HP Operations Agent - Performance Collection Component for HP-UX Dictionary of Operating System Performance Metrics Print Date 08/2012 HP Operations Agent for HP-UX Release 11.10 ************************************************************* Legal Notices ============= Warranty -------- The only warranties for HP products and services are set forth in the express warranty statements accompanying such products and services. Nothing herein should be construed as constituting an additional warranty. HP shall not be liable for technical or editorial errors or omissions contained herein. The information contained herein is subject to change without notice. Restricted Rights Legend ------------------------ Confidential computer software. Valid license from HP required for possession, use or copying. Consistent with FAR 12.211 and 12.212, Commercial Computer Software, Computer Software Documentation, and Technical Data for Commercial Items are licensed to the U.S. Government under vendor's standard commercial license. Copyright Notices ----------------- ©Copyright 2010-2012 Hewlett-Packard Development Company, L.P. All rights reserved. ***************************************************************** Introduction ============ This dictionary contains definitions of the HP-UX operating system performance metrics for the Performance Collection Component. This document is divided into the following sections: * "Metric Names by Data Class," which lists the metrics alphabetically by data class. Use these metric names for exporting data with the extract utility. You can also use these metric names in defining alarm conditions in your alarmdef file. * "Metric Definitions," which describes each metric in alphabetical order. Please note that the metric help has been put in a more generic format and references are made to the other platforms that also support each of the metrics. Metric Names by Data Class ========================== HP-UX Global Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR GBL_ACTIVE_CPU GBL_ACTIVE_PROC GBL_ALIVE_PROC GBL_COMPLETED_PROC GBL_CPU_CSWITCH_TIME GBL_CPU_CSWITCH_UTIL GBL_CPU_HISTOGRAM GBL_CPU_IDLE_TIME GBL_CPU_IDLE_UTIL GBL_CPU_INTERRUPT_TIME GBL_CPU_INTERRUPT_UTIL GBL_CPU_NICE_TIME GBL_CPU_NICE_UTIL GBL_CPU_NORMAL_TIME GBL_CPU_NORMAL_UTIL GBL_CPU_REALTIME_TIME GBL_CPU_REALTIME_UTIL GBL_CPU_SYSCALL_TIME GBL_CPU_SYSCALL_UTIL GBL_CPU_SYS_MODE_TIME GBL_CPU_SYS_MODE_UTIL GBL_CPU_TOTAL_TIME GBL_CPU_TOTAL_UTIL GBL_CPU_USER_MODE_TIME GBL_CPU_USER_MODE_UTIL GBL_DISK_FS_IO GBL_DISK_FS_IO_RATE GBL_DISK_FS_READ GBL_DISK_FS_READ_RATE GBL_DISK_FS_WRITE GBL_DISK_FS_WRITE_RATE GBL_DISK_HISTOGRAM GBL_DISK_LOGL_IO GBL_DISK_LOGL_IO_RATE GBL_DISK_LOGL_READ GBL_DISK_LOGL_READ_BYTE GBL_DISK_LOGL_READ_BYTE_RATE GBL_DISK_LOGL_READ_RATE GBL_DISK_LOGL_WRITE GBL_DISK_LOGL_WRITE_BYTE GBL_DISK_LOGL_WRITE_BYTE_RATE GBL_DISK_LOGL_WRITE_RATE GBL_DISK_PHYS_BYTE GBL_DISK_PHYS_BYTE_RATE GBL_DISK_PHYS_IO GBL_DISK_PHYS_IO_RATE GBL_DISK_PHYS_READ GBL_DISK_PHYS_READ_BYTE_RATE GBL_DISK_PHYS_READ_RATE GBL_DISK_PHYS_WRITE GBL_DISK_PHYS_WRITE_BYTE_RATE GBL_DISK_PHYS_WRITE_RATE GBL_DISK_RAW_IO GBL_DISK_RAW_IO_RATE GBL_DISK_RAW_READ GBL_DISK_RAW_READ_RATE GBL_DISK_RAW_WRITE GBL_DISK_RAW_WRITE_RATE GBL_DISK_SUBSYSTEM_QUEUE GBL_DISK_SYSTEM_IO GBL_DISK_SYSTEM_IO_RATE GBL_DISK_SYSTEM_READ GBL_DISK_SYSTEM_READ_RATE GBL_DISK_SYSTEM_WRITE GBL_DISK_SYSTEM_WRITE_RATE GBL_DISK_TIME_PEAK GBL_DISK_UTIL_PEAK GBL_DISK_VM_IO GBL_DISK_VM_IO_RATE GBL_DISK_VM_READ GBL_DISK_VM_READ_RATE GBL_DISK_VM_WRITE GBL_DISK_VM_WRITE_RATE GBL_FS_SPACE_UTIL_PEAK GBL_IPC_SUBSYSTEM_QUEUE GBL_LOST_MI_TRACE_BUFFERS GBL_MEM_ACTIVE_VIRT_UTIL GBL_MEM_CACHE_HIT_PCT GBL_MEM_FREE_UTIL GBL_MEM_PAGEOUT GBL_MEM_PAGEOUT_RATE GBL_MEM_PAGE_REQUEST GBL_MEM_PAGE_REQUEST_RATE GBL_MEM_QUEUE GBL_MEM_SWAP GBL_MEM_SWAPOUT_RATE GBL_MEM_SWAP_1_HR_RATE GBL_MEM_SYS_AND_CACHE_UTIL GBL_MEM_SYS_UTIL GBL_MEM_USER_UTIL GBL_MEM_UTIL GBL_NETWORK_SUBSYSTEM_QUEUE GBL_NET_COLLISION_1_MIN_RATE GBL_NET_COLLISION_PCT GBL_NET_ERROR_1_MIN_RATE GBL_NET_IN_ERROR_PCT GBL_NET_IN_PACKET GBL_NET_IN_PACKET_RATE GBL_NET_OUTQUEUE GBL_NET_OUT_ERROR_PCT GBL_NET_OUT_PACKET GBL_NET_OUT_PACKET_RATE GBL_NET_PACKET_RATE GBL_NFS_CALL GBL_NFS_CALL_RATE GBL_NUM_DISK GBL_NUM_NETWORK GBL_NUM_USER GBL_OTHER_QUEUE GBL_PRI_QUEUE GBL_PROC_RUN_TIME GBL_PROC_SAMPLE GBL_QUEUE_HISTOGRAM GBL_RUN_QUEUE GBL_SLEEP_QUEUE GBL_STARTED_PROC GBL_SWAP_SPACE_UTIL GBL_SYSCALL_RATE GBL_SYSTEM_UPTIME_HOURS GBL_TERM_FIRST_RESP_DIST_1 GBL_TERM_FIRST_RESP_DIST_10 GBL_TERM_FIRST_RESP_DIST_2 GBL_TERM_FIRST_RESP_DIST_3 GBL_TERM_FIRST_RESP_DIST_4 GBL_TERM_FIRST_RESP_DIST_5 GBL_TERM_FIRST_RESP_DIST_6 GBL_TERM_FIRST_RESP_DIST_7 GBL_TERM_FIRST_RESP_DIST_8 GBL_TERM_FIRST_RESP_DIST_9 GBL_TERM_IO_QUEUE GBL_TERM_RESP_DIST_1 GBL_TERM_RESP_DIST_10 GBL_TERM_RESP_DIST_2 GBL_TERM_RESP_DIST_3 GBL_TERM_RESP_DIST_4 GBL_TERM_RESP_DIST_5 GBL_TERM_RESP_DIST_6 GBL_TERM_RESP_DIST_7 GBL_TERM_RESP_DIST_8 GBL_TERM_RESP_DIST_9 GBL_TT_OVERFLOW_COUNT TBL_BUFFER_CACHE_USED TBL_FILE_LOCK_UTIL TBL_FILE_TABLE_UTIL TBL_INODE_CACHE_USED TBL_MSG_TABLE_UTIL TBL_PROC_TABLE_UTIL TBL_SEM_TABLE_UTIL TBL_SHMEM_TABLE_UTIL HP-UX Application Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR APP_ACTIVE_PROC APP_ALIVE_PROC APP_COMPLETED_PROC APP_CPU_NICE_TIME APP_CPU_NICE_UTIL APP_CPU_NORMAL_TIME APP_CPU_NORMAL_UTIL APP_CPU_REALTIME_TIME APP_CPU_REALTIME_UTIL APP_CPU_SYS_MODE_TIME APP_CPU_SYS_MODE_UTIL APP_CPU_TOTAL_TIME APP_CPU_TOTAL_UTIL APP_DISK_FS_IO APP_DISK_FS_IO_RATE APP_DISK_LOGL_IO APP_DISK_LOGL_IO_RATE APP_DISK_LOGL_READ APP_DISK_LOGL_READ_RATE APP_DISK_LOGL_WRITE APP_DISK_LOGL_WRITE_RATE APP_DISK_PHYS_IO APP_DISK_PHYS_IO_RATE APP_DISK_PHYS_READ APP_DISK_PHYS_READ_RATE APP_DISK_PHYS_WRITE APP_DISK_PHYS_WRITE_RATE APP_DISK_RAW_IO APP_DISK_RAW_IO_RATE APP_DISK_SUBSYSTEM_WAIT_PCT APP_DISK_SYSTEM_IO APP_DISK_SYSTEM_IO_RATE APP_DISK_VM_IO APP_DISK_VM_IO_RATE APP_IO_BYTE APP_IO_BYTE_RATE APP_IPC_SUBSYSTEM_WAIT_PCT APP_MEM_RES APP_MEM_VIRT APP_MEM_WAIT_PCT APP_NAME APP_NETWORK_SUBSYSTEM_WAIT_PCT APP_NUM APP_OTHER_IO_WAIT_PCT APP_PRI APP_PRI_STD_DEV APP_PRI_WAIT_PCT APP_PRM_CPUCAP_MODE APP_PRM_CPU_ENTITLEMENT APP_PRM_LOGGING_MODE APP_PRM_MEM_AVAIL APP_PRM_MEM_ENTITLEMENT APP_PRM_MEM_STATE APP_PRM_MEM_UPPERBOUND APP_PRM_MEM_UTIL APP_PRM_STATE APP_PROC_RUN_TIME APP_SAMPLE APP_SEM_WAIT_PCT APP_SLEEP_WAIT_PCT APP_TERM_IO_WAIT_PCT HP-UX Process Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR PROC_APP_ID PROC_CPU_CSWITCH_TIME PROC_CPU_CSWITCH_UTIL PROC_CPU_INTERRUPT_TIME PROC_CPU_INTERRUPT_UTIL PROC_CPU_NICE_TIME PROC_CPU_NICE_UTIL PROC_CPU_NORMAL_TIME PROC_CPU_NORMAL_UTIL PROC_CPU_REALTIME_TIME PROC_CPU_REALTIME_UTIL PROC_CPU_SYSCALL_TIME PROC_CPU_SYSCALL_UTIL PROC_CPU_SYS_MODE_TIME PROC_CPU_SYS_MODE_UTIL PROC_CPU_TOTAL_TIME PROC_CPU_TOTAL_TIME_CUM PROC_CPU_TOTAL_UTIL PROC_CPU_TOTAL_UTIL_CUM PROC_CPU_USER_MODE_TIME PROC_CPU_USER_MODE_UTIL PROC_DISK_FS_IO PROC_DISK_FS_IO_RATE PROC_DISK_FS_READ PROC_DISK_FS_READ_RATE PROC_DISK_FS_WRITE PROC_DISK_FS_WRITE_RATE PROC_DISK_LOGL_IO_CUM PROC_DISK_LOGL_IO_RATE_CUM PROC_DISK_LOGL_READ PROC_DISK_LOGL_READ_RATE PROC_DISK_LOGL_WRITE PROC_DISK_LOGL_WRITE_RATE PROC_DISK_PHYS_IO PROC_DISK_PHYS_IO_CUM PROC_DISK_PHYS_IO_RATE PROC_DISK_PHYS_IO_RATE_CUM PROC_DISK_SUBSYSTEM_WAIT_PCT PROC_DISK_SUBSYSTEM_WAIT_TIME PROC_DISK_SYSTEM_IO PROC_DISK_SYSTEM_IO_RATE PROC_DISK_VM_IO PROC_DISK_VM_IO_RATE PROC_GROUP_ID PROC_INTEREST PROC_INTERVAL_ALIVE PROC_IO_BYTE PROC_IO_BYTE_CUM PROC_IO_BYTE_RATE PROC_IO_BYTE_RATE_CUM PROC_IPC_SUBSYSTEM_WAIT_PCT PROC_IPC_SUBSYSTEM_WAIT_TIME PROC_LAN_WAIT_PCT PROC_LAN_WAIT_TIME PROC_MAJOR_FAULT PROC_MEM_RES PROC_MEM_VIRT PROC_MEM_WAIT_PCT PROC_MEM_WAIT_TIME PROC_MINOR_FAULT PROC_NFS_WAIT_PCT PROC_NFS_WAIT_TIME PROC_OTHER_IO_WAIT_PCT PROC_OTHER_IO_WAIT_TIME PROC_OTHER_WAIT_PCT PROC_OTHER_WAIT_TIME PROC_PARENT_PROC_ID PROC_PRI PROC_PRI_WAIT_PCT PROC_PRI_WAIT_TIME PROC_PRMID PROC_PROC_ID PROC_PROC_NAME PROC_RUN_TIME PROC_SEM_WAIT_PCT PROC_SEM_WAIT_TIME PROC_SLEEP_WAIT_PCT PROC_SLEEP_WAIT_TIME PROC_STOP_HISTOGRAM PROC_STOP_REASON PROC_SYS_WAIT_PCT PROC_SYS_WAIT_TIME PROC_TERM_IO_WAIT_PCT PROC_TERM_IO_WAIT_TIME PROC_THREAD_COUNT PROC_TTY PROC_USER_NAME HP-UX Transaction Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR TTBIN_TRANS_COUNT_1 TTBIN_TRANS_COUNT_10 TTBIN_TRANS_COUNT_2 TTBIN_TRANS_COUNT_3 TTBIN_TRANS_COUNT_4 TTBIN_TRANS_COUNT_5 TTBIN_TRANS_COUNT_6 TTBIN_TRANS_COUNT_7 TTBIN_TRANS_COUNT_8 TTBIN_TRANS_COUNT_9 TTBIN_UPPER_RANGE_1 TTBIN_UPPER_RANGE_10 TTBIN_UPPER_RANGE_2 TTBIN_UPPER_RANGE_3 TTBIN_UPPER_RANGE_4 TTBIN_UPPER_RANGE_5 TTBIN_UPPER_RANGE_6 TTBIN_UPPER_RANGE_7 TTBIN_UPPER_RANGE_8 TTBIN_UPPER_RANGE_9 TT_ABORT TT_ABORT_WALL_TIME_PER_TRAN TT_APP_NAME TT_APP_TRAN_NAME TT_CLIENT_ADDRESS TT_CLIENT_ADDRESS_FORMAT TT_CLIENT_TRAN_ID TT_COUNT TT_CPU_TOTAL_TIME_PER_TRAN TT_DISK_LOGL_IO_PER_TRAN TT_DISK_PHYS_IO_PER_TRAN TT_FAILED TT_INFO TT_NAME TT_NUM_BINS TT_SLO_COUNT TT_SLO_PERCENT TT_SLO_THRESHOLD TT_TERM_TRAN_1_HR_RATE TT_TRAN_1_MIN_RATE TT_TRAN_ID TT_UNAME TT_USER_MEASUREMENT_AVG TT_USER_MEASUREMENT_AVG_2 TT_USER_MEASUREMENT_AVG_3 TT_USER_MEASUREMENT_AVG_4 TT_USER_MEASUREMENT_AVG_5 TT_USER_MEASUREMENT_AVG_6 TT_USER_MEASUREMENT_COUNT TT_USER_MEASUREMENT_COUNT_2 TT_USER_MEASUREMENT_COUNT_3 TT_USER_MEASUREMENT_COUNT_4 TT_USER_MEASUREMENT_COUNT_5 TT_USER_MEASUREMENT_COUNT_6 TT_USER_MEASUREMENT_MAX TT_USER_MEASUREMENT_MAX_2 TT_USER_MEASUREMENT_MAX_3 TT_USER_MEASUREMENT_MAX_4 TT_USER_MEASUREMENT_MAX_5 TT_USER_MEASUREMENT_MAX_6 TT_USER_MEASUREMENT_MIN TT_USER_MEASUREMENT_MIN_2 TT_USER_MEASUREMENT_MIN_3 TT_USER_MEASUREMENT_MIN_4 TT_USER_MEASUREMENT_MIN_5 TT_USER_MEASUREMENT_MIN_6 TT_USER_MEASUREMENT_NAME TT_USER_MEASUREMENT_NAME_2 TT_USER_MEASUREMENT_NAME_3 TT_USER_MEASUREMENT_NAME_4 TT_USER_MEASUREMENT_NAME_5 TT_USER_MEASUREMENT_NAME_6 TT_WALL_TIME_PER_TRAN HP-UX Disk Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR BYDSK_AVG_SERVICE_TIME BYDSK_DEVNAME BYDSK_DIRNAME BYDSK_FS_READ BYDSK_FS_READ_RATE BYDSK_FS_WRITE BYDSK_FS_WRITE_RATE BYDSK_HISTOGRAM BYDSK_ID BYDSK_LOGL_READ BYDSK_LOGL_READ_RATE BYDSK_LOGL_WRITE BYDSK_LOGL_WRITE_RATE BYDSK_PHYS_BYTE BYDSK_PHYS_BYTE_RATE BYDSK_PHYS_IO BYDSK_PHYS_IO_RATE BYDSK_PHYS_READ BYDSK_PHYS_READ_BYTE BYDSK_PHYS_READ_BYTE_RATE BYDSK_PHYS_READ_RATE BYDSK_PHYS_WRITE BYDSK_PHYS_WRITE_BYTE BYDSK_PHYS_WRITE_BYTE_RATE BYDSK_PHYS_WRITE_RATE BYDSK_RAW_READ BYDSK_RAW_READ_RATE BYDSK_RAW_WRITE BYDSK_RAW_WRITE_RATE BYDSK_REQUEST_QUEUE BYDSK_SYSTEM_IO BYDSK_SYSTEM_IO_RATE BYDSK_UTIL BYDSK_VM_IO BYDSK_VM_IO_RATE HP-UX Logical Volume Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR FS_DIRNAME LV_DIRNAME LV_GROUP_NAME LV_LOGL_READ LV_LOGL_WRITE LV_READ_BYTE_RATE LV_READ_RATE LV_SPACE_UTIL LV_WRITE_BYTE_RATE LV_WRITE_RATE HP-UX Network Interface Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR BYNETIF_COLLISION BYNETIF_COLLISION_RATE BYNETIF_ERROR BYNETIF_ERROR_RATE BYNETIF_IN_BYTE_RATE BYNETIF_IN_BYTE_RATE_CUM BYNETIF_IN_PACKET BYNETIF_IN_PACKET_RATE BYNETIF_NAME BYNETIF_NET_MTU BYNETIF_NET_SPEED BYNETIF_OUT_BYTE_RATE BYNETIF_OUT_BYTE_RATE_CUM BYNETIF_OUT_PACKET BYNETIF_OUT_PACKET_RATE BYNETIF_QUEUE HP-UX CPU Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR BYCPU_CPU_SYS_MODE_UTIL BYCPU_CPU_TOTAL_UTIL BYCPU_CPU_USER_MODE_UTIL BYCPU_CSWITCH_RATE BYCPU_ID BYCPU_INTERRUPT_RATE BYCPU_STATE HP-UX Filesystem Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR FS_BLOCK_SIZE FS_DEVNAME FS_FRAG_SIZE FS_MAX_SIZE FS_SPACE_UTIL FS_TYPE HP-UX Configuration Metrics ---------------------------------- BLANK DATE DATE_SECONDS DAY INTERVAL RECORD_TYPE TIME YEAR GBL_COLLECTOR GBL_LOGFILE_VERSION GBL_LOGGING_TYPES GBL_MACHINE GBL_MACHINE_MODEL GBL_MEM_AVAIL GBL_MEM_PHYS GBL_NUM_CPU GBL_OSKERNELTYPE_INT GBL_OSNAME GBL_OSRELEASE GBL_OSVERSION GBL_SWAP_SPACE_AVAIL_KB GBL_SYSTEM_ID GBL_TERM_FIRST_RESP_RANGE_1 GBL_TERM_FIRST_RESP_RANGE_2 GBL_TERM_FIRST_RESP_RANGE_3 GBL_TERM_FIRST_RESP_RANGE_4 GBL_TERM_FIRST_RESP_RANGE_5 GBL_TERM_FIRST_RESP_RANGE_6 GBL_TERM_FIRST_RESP_RANGE_7 GBL_TERM_FIRST_RESP_RANGE_8 GBL_TERM_FIRST_RESP_RANGE_9 GBL_TERM_RESP_RANGE_1 GBL_TERM_RESP_RANGE_2 GBL_TERM_RESP_RANGE_3 GBL_TERM_RESP_RANGE_4 GBL_TERM_RESP_RANGE_5 GBL_TERM_RESP_RANGE_6 GBL_TERM_RESP_RANGE_7 GBL_TERM_RESP_RANGE_8 GBL_TERM_RESP_RANGE_9 GBL_THRESHOLD_CPU GBL_THRESHOLD_DISK GBL_THRESHOLD_FIRSTRESP GBL_THRESHOLD_NOKILLED GBL_THRESHOLD_NONEW GBL_THRESHOLD_PROCMEM GBL_THRESHOLD_RESPONSE GBL_THRESHOLD_SHORTLIVED GBL_THRESHOLD_TRANS GBL_THRESHOLD_WAIT_CPU GBL_THRESHOLD_WAIT_DISK GBL_THRESHOLD_WAIT_IMPEDE GBL_THRESHOLD_WAIT_MEMORY TBL_BUFFER_CACHE_AVAIL TBL_FILE_LOCK_AVAIL TBL_FILE_TABLE_AVAIL TBL_INODE_CACHE_AVAIL TBL_MSG_TABLE_AVAIL TBL_PROC_TABLE_AVAIL TBL_SEM_TABLE_AVAIL TBL_SHMEM_TABLE_AVAIL Metric Definitions ================== APP_ACTIVE_PROC ---------------------------------- An active process is one that exists and consumes some CPU time. APP_ACTIVE_PROC is the sum of the alive-process-time/interval-time ratios of every process belonging to an application that is active (uses any CPU time) during an interval. The following diagram of a four second interval showing two processes, A and B, for an application should be used to understand the above definition. Note the difference between active processes, which consume CPU time, and alive processes which merely exist on the system. ----------- Seconds ----------- 1 2 3 4 Proc ---- ---- ---- ---- ---- A live live live live B live/CPU live/CPU live dead Process A is alive for the entire four second interval, but consumes no CPU. A’s contribution to APP_ALIVE_PROC is 4*1/4. A contributes 0*1/4 to APP_ACTIVE_PROC. B’s contribution to APP_ALIVE_PROC is 3*1/4. B contributes 2*1/4 to APP_ACTIVE_PROC. Thus, for this interval, APP_ACTIVE_PROC equals 0.5 and APP_ALIVE_PROC equals 1.75. Because a process may be alive but not active, APP_ACTIVE_PROC will always be less than or equal to APP_ALIVE_PROC. This metric indicates the number of processes in an application group that are competing for the CPU. This metric is useful, along with other metrics, for comparing loads placed on the system by different groups of processes. On non HP-UX systems, this metric is derived from sampled process data. Since the data for a process is not available after the process has died on this operating system, a process whose life is shorter than the sampling interval may not be seen when the samples are taken. Thus this metric may be slightly less than the actual value. Increasing the sampling frequency captures a more accurate count, but the overhead of collection may also rise. APP_ALIVE_PROC ---------------------------------- An alive process is one that exists on the system. APP_ALIVE_PROC is the sum of the alive-process-time/interval-time ratios for every process belonging to a given application. The following diagram of a four second interval showing two processes, A and B, for an application should be used to understand the above definition. Note the difference between active processes, which consume CPU time, and alive processes which merely exist on the system. ----------- Seconds ----------- 1 2 3 4 Proc ---- ---- ---- ---- ---- A live live live live B live/CPU live/CPU live dead Process A is alive for the entire four second interval but consumes no CPU. A’s contribution to APP_ALIVE_PROC is 4*1/4. A contributes 0*1/4 to APP_ACTIVE_PROC. B’s contribution to APP_ALIVE_PROC is 3*1/4. B contributes 2*1/4 to APP_ACTIVE_PROC. Thus, for this interval, APP_ACTIVE_PROC equals 0.5 and APP_ALIVE_PROC equals 1.75. Because a process may be alive but not active, APP_ACTIVE_PROC will always be less than or equal to APP_ALIVE_PROC. On non HP-UX systems, this metric is derived from sampled process data. Since the data for a process is not available after the process has died on this operating system, a process whose life is shorter than the sampling interval may not be seen when the samples are taken. Thus this metric may be slightly less than the actual value. Increasing the sampling frequency captures a more accurate count, but the overhead of collection may also rise. APP_COMPLETED_PROC ---------------------------------- The number of processes in this group that completed during the interval. On non HP-UX systems, this metric is derived from sampled process data. Since the data for a process is not available after the process has died on this operating system, a process whose life is shorter than the sampling interval may not be seen when the samples are taken. Thus this metric may be slightly less than the actual value. Increasing the sampling frequency captures a more accurate count, but the overhead of collection may also rise. APP_CPU_NICE_TIME ---------------------------------- The time, in seconds, that processes in this group were using the CPU in user mode at a nice priority during the interval. On HP-UX, the NICE metrics include positive nice value CPU time only. Negative nice value CPU is broken out into NNICE (negative nice) metrics. Positive nice values range from 20 to 39. Negative nice values range from 0 to 19. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_NICE_UTIL ---------------------------------- The percentage of time that processes in this group were using the CPU in user mode at a nice priority during the interval. On HP-UX, the NICE metrics include positive nice value CPU time only. Negative nice value CPU is broken out into NNICE (negative nice) metrics. Positive nice values range from 20 to 39. Negative nice values range from 0 to 19. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_NORMAL_TIME ---------------------------------- The time, in seconds, that processes in this group were in user mode at a normal priority during the interval. Normal priority user mode CPU excludes CPU used at real-time and nice priorities. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_NORMAL_UTIL ---------------------------------- The percentage of time that processes in this group were in user mode running at normal priority during the interval. Normal priority user mode CPU excludes CPU used at real-time and nice priorities. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_REALTIME_TIME ---------------------------------- The time, in seconds, that the processes in this group were in user mode at a “realtime” priority during the interval. “Realtime” priority is 0-127. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_REALTIME_UTIL ---------------------------------- The percentage of time that processes in this group were in user mode at a “realtime” priority during the interval. “Realtime” priority is 0-127. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_SYS_MODE_TIME ---------------------------------- The time, in seconds, during the interval that the CPU was in system mode for processes in this group. A process operates in either system mode (also called kernel mode on Unix or privileged mode on Windows) or user mode. When a process requests services from the operating system with a system call, it switches into the machine’s privileged protection mode and runs in system mode. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_SYS_MODE_UTIL ---------------------------------- The percentage of time during the interval that the CPU was used in system mode for processes in this group. A process operates in either system mode (also called kernel mode on Unix or privileged mode on Windows) or user mode. When a process requests services from the operating system with a system call, it switches into the machine’s privileged protection mode and runs in system mode. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. High system CPU utilizations are normal for IO intensive groups. Abnormally high system CPU utilization can indicate that a hardware problem is causing a high interrupt rate. It can also indicate programs that are not making efficient system calls. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_TOTAL_TIME ---------------------------------- The total CPU time, in seconds, devoted to processes in this group during the interval. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_CPU_TOTAL_UTIL ---------------------------------- The percentage of the total CPU time devoted to processes in this group during the interval. This indicates the relative CPU load placed on the system by processes in this group. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. Large values for this metric may indicate that this group is causing a CPU bottleneck. This would be normal in a computation- bound workload, but might mean that processes are using excessive CPU time and perhaps looping. If the “other” application shows significant amounts of CPU, you may want to consider tuning your parm file so that process activity is accounted for in known applications. APP_CPU_TOTAL_UTIL = APP_CPU_SYS_MODE_UTIL + APP_CPU_USER_MODE_UTIL NOTE: On Windows, the sum of the APP_CPU_TOTAL_UTIL metrics may not equal GBL_CPU_TOTAL_UTIL. Microsoft states that “this is expected behavior” because the GBL_CPU_TOTAL_UTIL metric is taken from the NT performance library Processor objects while the APP_CPU_TOTAL_UTIL metrics are taken from the Process objects. Microsoft states that there can be CPU time accounted for in the Processor system objects that may not be seen in the Process objects. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. APP_DISK_FS_IO ---------------------------------- The number of file system disk IOs for processes in this group during the interval. These are physical reads generated by user file system access and do not include virtual memory reads, system reads (inode access), or reads relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical IOs in this category (they appear under virtual memory IOs). APP_DISK_FS_IO_RATE ---------------------------------- The number of file system disk IOs for processes in this group during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical IOs generated by user file system access and do not include virtual memory IOs, system IOs (inode updates), or IOs relating to raw disk access. An exception is user files accessed via the mmap(2) call, which will not show their physical IOs in this category. They appear under virtual memory IOs. APP_DISK_LOGL_IO ---------------------------------- The number of logical IOs for processes in this group during the interval. On many Unix systems, logical disk IOs are measured by counting the read and write system calls that are directed to disk devices. Also counted are read and write system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, writev, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. APP_DISK_LOGL_IO_RATE ---------------------------------- The number of logical IOs per second for processes in this group during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read and write system calls that are directed to disk devices. Also counted are read and write system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, writev, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. APP_DISK_LOGL_READ ---------------------------------- The number of logical reads for processes in this group during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. APP_DISK_LOGL_READ_RATE ---------------------------------- The number of logical reads per second for processes in this group during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. APP_DISK_LOGL_WRITE ---------------------------------- The number of logical writes for processes in this group during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. APP_DISK_LOGL_WRITE_RATE ---------------------------------- The number of logical writes per second for processes in this group during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. APP_DISK_PHYS_IO ---------------------------------- The number of physical IOs for processes in this group during the interval. On SUN systems, this metric is only available on Sun 5.X or later. APP_DISK_PHYS_IO_RATE ---------------------------------- The number of physical IOs per second for processes in this group during the interval. APP_DISK_PHYS_READ ---------------------------------- The number of physical reads for processes in this group during the interval. APP_DISK_PHYS_READ_RATE ---------------------------------- The number of physical reads per second for processes in this group during the interval. APP_DISK_PHYS_WRITE ---------------------------------- The number of physical writes for processes in this group during the interval. APP_DISK_PHYS_WRITE_RATE ---------------------------------- The number of physical writes per second for processes in this group during the interval. APP_DISK_RAW_IO ---------------------------------- The total number of raw IOs for processes in this group during the interval. Only accesses to local disk devices are counted. APP_DISK_RAW_IO_RATE ---------------------------------- The total number of raw IOs for processes in this group during the interval. Only accesses to local disk devices are counted. APP_DISK_SUBSYSTEM_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on the disk subsystem (waiting for their file system IOs to complete) during the interval. This is the sum of processes or kernel threads in the DISK, INODE, CACHE and CDFS wait states. It does not include processes or kernel threads doing raw IO to disk devices. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_DISK_SYSTEM_IO ---------------------------------- The number of physical IOs generated by the kernel for file system management (inode accesses or updates) for processes in this group during the interval APP_DISK_SYSTEM_IO_RATE ---------------------------------- The number of physical IOs per second generated by the kernel for file system management (inode accesses or updates) for processes in this group during the interval. APP_DISK_VM_IO ---------------------------------- The number of virtual memory IOs made on behalf of processes in this group during the interval. IOs to user file data are not included in this metric unless they were done via the mmap(2) system call. APP_DISK_VM_IO_RATE ---------------------------------- The number of virtual memory IOs per second made on behalf of processes in this group during the interval. IOs to user file data are not included in this metric unless they were done via the mmap(2) system call. APP_IO_BYTE ---------------------------------- The number of characters (in KB) transferred for processes in this group to all devices during the interval. This includes IO to disk, terminal, tape and printers. APP_IO_BYTE_RATE ---------------------------------- The number of characters (in KB) per second transferred for processes in this group to all devices during the interval. This includes IO to disk, terminal, tape and printers. APP_IPC_SUBSYSTEM_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on the InterProcess Communication (IPC) subsystems (waiting for their interprocess communication activity to complete) during the interval. This is the sum of processes or kernel threads in the IPC, MSG, SEM, PIPE, SOCKT (that is, sockets) and STRMS (that is, streams IO) wait states. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_MEM_RES ---------------------------------- On Unix systems, this is the sum of the size (in MB) of resident memory for processes in this group that were alive at the end of the interval. This consists of text, data, stack, and shared memory regions. On HP-UX, since PROC_MEM_RES typically takes shared region references into account, this approximates the total resident (physical) memory consumed by all processes in this group. On all other Unix systems, this is the sum of the resident memory region sizes for all processes in this group. When the resident memory size for processes includes shared regions, such as shared memory and library text and data, the shared regions are counted multiple times in this sum. For example, if the application contains four processes that are attached to a 500MB shared memory region that is all resident in physical memory, then 2000MB is contributed towards the sum in this metric. As such, this metric can overestimate the resident memory being used by processes in this group when they share memory regions. Refer to the help text for PROC_MEM_RES for additional information. On Windows, this is the sum of the size (in MB) of the working sets for processes in this group during the interval. The working set counts memory pages referenced recently by the threads making up this group. Note that the size of the working set is often larger than the amount of pagefile space consumed. APP_MEM_VIRT ---------------------------------- On Unix systems, this is the sum (in MB) of virtual memory for processes in this group that were alive at the end of the interval. This consists of text, data, stack, and shared memory regions. On HP-UX, since PROC_MEM_VIRT typically takes shared region references into account, this approximates the total virtual memory consumed by all processes in this group. On all other Unix systems, this is the sum of the virtual memory region sizes for all processes in this group. When the virtual memory size for processes includes shared regions, such as shared memory and library text and data, the shared regions are counted multiple times in this sum. For example, if the application contains four processes that are attached to a 500MB shared memory region, then 2000MB is reported in this metric. As such, this metric can overestimate the virtual memory being used by processes in this group when they share memory regions. On Windows, this is the sum (in MB) of paging file space used for all processes in this group during the interval. Groups of processes may have working set sizes (APP_MEM_RES) larger than the size of their pagefile space. APP_MEM_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on memory (waiting for virtual memory disk accesses to complete) during the interval. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_NAME ---------------------------------- The name of the application (up to 20 characters). This comes from the parm file where the applications are defined. The application called “other” captures all processes not aggregated into applications specifically defined in the parm file. In other words, if no applications are defined in the parm file, then all process data would be reflected in the “other” application. If the parm file switch to log PRM group data, instead of application data, is in effect (indicated by APP_PRM_LOGGING_MODE = 1 and the log statement in the parm file includes application=prm), then this name is the PRM groupname defined in the HP-UX Process Resource Manager configuration file. APP_NETWORK_SUBSYSTEM_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on the network subsystem (waiting for their network activity to complete) during the interval. This is the sum of processes or kernel threads in the LAN, NFS, and RPC wait states. This does not include processes or kernel threads blocked on SOCKT (that is, socket) waits, as some processes or kernel threads sit idle in SOCKT waits for long periods. This is calculated as the accumulated time that all processes or kernel threads in this group spent blocked on (LAN + NFS + RPC) divided by the interval time. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_NUM ---------------------------------- The sequentially assigned number of this application or, on Solaris, the project ID when application grouping by project is enabled. APP_OTHER_IO_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on “other IO” during the interval. “Other IO” includes all IO directed at a device (connected to the local computer) which is not a terminal or LAN. Examples of “other IO” devices are local printers, tapes, instruments, and disks. Time waiting for character (raw) IO to disks is included in this measurement. Time waiting for file system buffered IO to disks will typically been seen as IO or CACHE wait. Time waiting for IO to NFS disks is reported as NFS wait. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_PRI ---------------------------------- On Unix systems, this is the average priority of the processes in this group during the interval. On Windows, this is the average base priority of the processes in this group during the interval. APP_PRI_STD_DEV ---------------------------------- The standard deviation of priorities of the processes in this group during the interval. This metric is available on HP-UX 10.20. APP_PRI_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on PRI (waiting for their priority to become high enough to get the CPU) during the interval. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_PRM_CPUCAP_MODE ---------------------------------- The PRM CPU Cap Mode state on this system: 0 = PRM is not installed or not configured. 1 = CPU Cap Mode is not enabled (PRM CPU entitlements are in effect) 2 = CPU Cap Mode is enabled (The PRM CPU entitlements behave as caps or limits) APP_PRM_CPU_ENTITLEMENT ---------------------------------- The PRM CPU entitlement for this PRM Group ID entry as defined in the PRM configuration file. APP_PRM_LOGGING_MODE ---------------------------------- The PRM logging mode will be 1 when PRM group data is being logged by the MeasureWare agent in place of parm file defined application data. By default, parm file application definitions are used and the PRM logging mode will be 0. A parm file setting read during MeasureWare Agent (scopeux) startup is used to override the default to log PRM group data. PRM group data logging is set in the parm file by specifying: application=prm in the log statement in versions of HP-UX MeasureWare which support this capability. See the MeasureWare Agent Release Notes file (located in /opt/perf/ReleaseNotes/Mwa) for more information. APP_PRM_MEM_AVAIL ---------------------------------- PRM available memory is the amount of physical memory less the amount of memory reserved for the kernel and system processes running in the PRM_SYS group 0. PRM available memory is a dynamic value that changes with system usage. APP_PRM_MEM_ENTITLEMENT ---------------------------------- The PRM MEM entitlement for this PRM Group ID entry as defined in the PRM configuration file. APP_PRM_MEM_STATE ---------------------------------- The PRM MEM state on this system: 0 = PRM is not installed or no memory specification 1 = reset (PRM is installed in reset condition or no memory specification) 2 = configured/disabled (The PRM memory scheduler is configured, but the standard HP-UX scheduler is in effect) 3 = enabled (The PRM memory scheduler is configured and in effect) APP_PRM_MEM_UPPERBOUND ---------------------------------- The PRM MEM upperbound for this PRM Group ID entry as defined in the PRM configuration file. APP_PRM_MEM_UTIL ---------------------------------- The percent of PRM memory used by processes (process private space plus a process’ portion of shared memory) within the PRM groups during the interval. PRM available memory is the amount of physical memory less the amount of memory reserved for the kernel and system processes running in the PRM_SYS group 0. PRM available memory is a dynamic value that changes with system usage. APP_PRM_STATE ---------------------------------- The PRM CPU state on this system: 0 = PRM is not installed 1 = reset (PRM is configured with only the system group. The standard HP-UX CPU scheduler is in effect) 2 = configured/disabled (the PRM CPU scheduler is configured, but the standard HP-UX scheduler is in effect) 3 = enabled (the PRM CPU scheduler is configured and in effect) APP_PROC_RUN_TIME ---------------------------------- The average run time for processes in this group that completed during the interval. On non HP-UX systems, this metric is derived from sampled process data. Since the data for a process is not available after the process has died on this operating system, a process whose life is shorter than the sampling interval may not be seen when the samples are taken. Thus this metric may be slightly less than the actual value. Increasing the sampling frequency captures a more accurate count, but the overhead of collection may also rise. APP_SAMPLE ---------------------------------- The number of samples of process data that have been averaged or accumulated during this sample. APP_SEM_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on semaphores (waiting for their semaphore operations to complete) during the interval. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_SLEEP_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on SLEEP (waiting to awaken from sleep system calls) during the interval. A process or kernel thread enters the SLEEP state by putting itself to sleep using system calls such as sleep, wait, pause, sigpause, sigsuspend, poll and select. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). The Application QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues, within the context of a specific application. The Application WAIT PCT metrics, which are also based on block states, represent the percentage of processes or kernel threads that were alive on the system within the context of a specific application. These values will vary greatly depending on the application. No direct comparison is reasonable with the Global Queue metrics since they represent the average number of all processes or kernel threads that were alive on the system. As such, the Application WAIT PCT metrics cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. APP_TERM_IO_WAIT_PCT ---------------------------------- The percentage of time processes or kernel threads in this group were blocked on terminal IO (waiting for terminal IO to complete) during the interval. A percentage of time spent in a wait state is calculated as the accumulated time kernel threads belonging to processes in this group spent waiting in this state, divided by accumulated alive time of kernel threads belonging to processes in this group during the interval. For example, assume an application has 20 kernel threads. During the interval, ten kernel threads slept the entire time, while ten kernel threads waited on terminal input. As a result, the application wait percent values would be 50% for SLEEP and 50% for TERM (that is, terminal IO). This metric is available on HP-UX 10.20. BLANK ---------------------------------- An empty field used for spacing reports. For example, this field can be used to create a blank column in a spreadsheet that may be used to sum several items. BYCPU_CPU_SYS_MODE_UTIL ---------------------------------- The percentage of time that this CPU (or logical processor) was in system mode during the interval. A process operates in either system mode (also called kernel mode on Unix or privileged mode on Windows) or user mode. When a process requests services from the operating system with a system call, it switches into the machine’s privileged protection mode and runs in system mode. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. BYCPU_CPU_TOTAL_UTIL ---------------------------------- The percentage of time that this CPU (or logical processor) was not idle during the interval. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. BYCPU_CPU_USER_MODE_UTIL ---------------------------------- The percentage of time that this CPU (or logical processor) was in user mode during the interval. User CPU is the time spent in user mode at a normal priority, at real-time priority (on HP-UX, AIX, and Windows systems), and at a nice priority. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. BYCPU_CSWITCH_RATE ---------------------------------- The average number of context switches per second for this CPU during the interval. On HP-UX, this includes context switches that result in the execution of a different process and those caused by a process stopping, then resuming, with no other process running in the meantime. BYCPU_ID ---------------------------------- The ID number of this CPU. On some Unix systems, such as SUN, CPUs are not sequentially numbered. BYCPU_INTERRUPT_RATE ---------------------------------- The average number of device interrupts per second for this CPU during the interval. On HP-UX, a value of “na” is displayed on a system with multiple CPUs. BYCPU_STATE ---------------------------------- A text string indicating the current state of a processor. On HP-UX, this is either “Enabled”, “Disabled” or “Unknown”. On AIX, this is either “Idle/Offline” or “Online”. On all other systems, this is either “Offline”, “Online” or “Unknown”. BYDSK_AVG_SERVICE_TIME ---------------------------------- The average time, in milliseconds, that this disk device spent processing each disk request during the interval. For example, a value of 5.14 would indicate that disk requests during the last interval took on average slightly longer than five one-thousandths of a second to complete for this device. Some Linux kernels, typically 2.2 and older kernels, do not support the instrumentation needed to provide values for this metric. This metric will be “na” on the affected kernels. The “sar -d” command will also not be present on these systems. Distributions and OS releases that are known to be affected include: TurboLinux 7, SuSE 7.2, and Debian 3.0. This is a measure of the speed of the disk, because slower disk devices typically show a larger average service time. Average service time is also dependent on factors such as the distribution of I/O requests over the interval and their locality. It can also be influenced by disk driver and controller features such as I/O merging and command queueing. Note that this service time is measured from the perspective of the kernel, not the disk device itself. For example, if a disk device can find the requested data in its cache, the average service time could be quicker than the speed of the physical disk hardware. This metric can be used to help determine which disk devices are taking more time than usual to process requests. BYDSK_DEVNAME ---------------------------------- The name of this disk device. On HP-UX, the name identifying the specific disk spindle is the hardware path which specifies the address of the hardware components leading to the disk device. On SUN, these names are the same disk names displayed by “iostat”. On AIX, this is the path name string of this disk device. This is the fsname parameter in the mount(1M) command. If more than one file system is contained on a device (that is, the device is partitioned), this is indicated by an asterisk (“*”) at the end of the path name. On OSF1, this is the path name string of this disk device. This is the file-system parameter in the mount(1M) command. On Windows, this is the unit number of this disk device. BYDSK_DIRNAME ---------------------------------- The name of the file system directory mounted on this disk device. If more than one file system is mounted on this device, “Multiple FS” is seen. BYDSK_FS_READ ---------------------------------- The number of physical file system reads from this disk device during the interval. BYDSK_FS_READ_RATE ---------------------------------- The number of physical file system reads per second from this disk device during the interval. BYDSK_FS_WRITE ---------------------------------- The number of physical file system writes to this disk device during the interval. BYDSK_FS_WRITE_RATE ---------------------------------- The number of physical file system writes per second to this disk device during the interval. BYDSK_HISTOGRAM ---------------------------------- A bar chart of the disk IO. Shows breakout of the disk IO. Disk IO Rate = BYDSK_FS_READ_RATE + BYDSK_FS_WRITE_RATE + BYDSK_RAW_READ_RATE + BYDSK_RAW_WRITE_RATE + BYDSK_VM_IO_RATE + BYDSK_SYSTEM_IO_RATE ASCII and binary files contain a line of ASCII characters that make up one row of a printed histogram. This can be a quick way to get a graphical view of Disk IO on a character mode terminal display. BYDSK_ID ---------------------------------- The ID of the current disk device. BYDSK_LOGL_READ ---------------------------------- The number of logical reads for this disk device during the interval. On HP-UX, the logical IO rates by disk device cannot be obtained in a multi-disk LVM configuration because there is no reasonable means of tying logical IO transactions to physical spindles spanned on the logical volume. Therefore, if you have a multi- disk LVM configuration, you always see “na” for this metric. BYDSK_LOGL_READ_RATE ---------------------------------- The number of logical reads per second for this disk device during the interval. On HP-UX, the logical IO rates by disk device cannot be obtained in a multi-disk LVM configuration because there is no reasonable means of tying logical IO transactions to physical spindles spanned on the logical volume. Therefore, if you have a multi- disk LVM configuration, you always see “na” for this metric. BYDSK_LOGL_WRITE ---------------------------------- The number of logical writes for this disk device during the interval. On HP-UX, the logical IO rates by disk device cannot be obtained in a multi-disk LVM configuration because there is no reasonable means of tying logical IO transactions to physical spindles spanned on the logical volume. Therefore, if you have a multi- disk LVM configuration, you always see “na” for this metric. BYDSK_LOGL_WRITE_RATE ---------------------------------- The number of logical writes per second for this disk device during the interval. On HP-UX, the logical IO rates by disk device cannot be obtained in a multi-disk LVM configuration because there is no reasonable means of tying logical IO transactions to physical spindles spanned on the logical volume. Therefore, if you have a multi- disk LVM configuration, you always see “na” for this metric. BYDSK_PHYS_BYTE ---------------------------------- The number of KBs of physical IOs transferred to or from this disk device during the interval. On Unix systems, all types of physical disk IOs are counted, including file system, virtual memory, and raw IO. BYDSK_PHYS_BYTE_RATE ---------------------------------- The average KBs per second transferred to or from this disk device during the interval. On Unix systems, all types of physical disk IOs are counted, including file system, virtual memory, and raw IO. BYDSK_PHYS_IO ---------------------------------- The number of physical IOs for this disk device during the interval. On Unix systems, all types of physical disk IOs are counted, including file system, virtual memory, and raw reads. BYDSK_PHYS_IO_RATE ---------------------------------- The average number of physical IO requests per second for this disk device during the interval. On Unix systems, all types of physical disk IOs are counted, including file system IO, virtual memory and raw IO. BYDSK_PHYS_READ ---------------------------------- The number of physical reads for this disk device during the interval. On Unix systems, all types of physical disk reads are counted, including file system, virtual memory, and raw reads. On AIX, this is an estimated value based on the ratio of read bytes to total bytes transferred. The actual number of reads is not tracked by the kernel. This is calculated as BYDSK_PHYS_READ = BYDSK_PHYS_IO * (BYDSK_PHYS_READ_BYTE / BYDSK_PHYS_IO_BYTE) BYDSK_PHYS_READ_BYTE ---------------------------------- The KBs transferred from this disk device during the interval. On Unix systems, all types of physical disk reads are counted, including file system, virtual memory, and raw IO. BYDSK_PHYS_READ_BYTE_RATE ---------------------------------- The average KBs per second transferred from this disk device during the interval. On Unix systems, all types of physical disk reads are counted, including file system, virtual memory, and raw IO. BYDSK_PHYS_READ_RATE ---------------------------------- The average number of physical reads per second for this disk device during the interval. On Unix systems, all types of physical disk reads are counted, including file system, virtual memory, and raw reads. On AIX, this is an estimated value based on the ratio of read bytes to total bytes transferred. The actual number of reads is not tracked by the kernel. This is calculated as BYDSK_PHYS_READ_RATE = BYDSK_PHYS_IO_RATE * (BYDSK_PHYS_READ_BYTE / BYDSK_PHYS_IO_BYTE) BYDSK_PHYS_WRITE ---------------------------------- The number of physical writes for this disk device during the interval. On Unix systems, all types of physical disk writes are counted, including file system IO, virtual memory IO, and raw writes. On AIX, this is an estimated value based on the ratio of write bytes to total bytes transferred because the actual number of writes is not tracked by the kernel. This is calculated as BYDSK_PHYS_WRITE = BYDSK_PHYS_IO * (BYDSK_PHYS_WRITE_BYTE / BYDSK_PHYS_IO_BYTE) BYDSK_PHYS_WRITE_BYTE ---------------------------------- The KBs transferred to this disk device during the interval. On Unix systems, all types of physical disk writes are counted, including file system, virtual memory, and raw IO. BYDSK_PHYS_WRITE_BYTE_RATE ---------------------------------- The average KBs per second transferred to this disk device during the interval. On Unix systems, all types of physical disk writes are counted, including file system, virtual memory, and raw IO. BYDSK_PHYS_WRITE_RATE ---------------------------------- The average number of physical writes per second for this disk device during the interval. On Unix systems, all types of physical disk writes are counted, including file system IO, virtual memory IO, and raw writes. On AIX, this is an estimated value based on the ratio of write bytes to total bytes transferred. The actual number of writes is not tracked by the kernel. This is calculated as BYDSK_PHYS_WRITE_RATE = BYDSK_PHYS_IO_RATE * (BYDSK_PHYS_WRITE_BYTE / BYDSK_PHYS_IO_BYTE) BYDSK_RAW_READ ---------------------------------- The number of physical raw reads made from this disk device during the interval. BYDSK_RAW_READ_RATE ---------------------------------- The number of raw reads per second made from this disk device during the interval. BYDSK_RAW_WRITE ---------------------------------- The number of physical raw writes made to this disk device during the interval. BYDSK_RAW_WRITE_RATE ---------------------------------- The number of raw writes per second made to this disk device during the interval. BYDSK_REQUEST_QUEUE ---------------------------------- The average number of IO requests that were in the wait queue for this disk device during the interval. These requests are the physical requests (as opposed to logical IO requests). Some Linux kernels, typically 2.2 and older kernels, do not support the instrumentation needed to provide values for this metric. This metric will be “na” on the affected kernels. The “sar -d” command will also not be present on these systems. Distributions and OS releases that are known to be affected include: TurboLinux 7, SuSE 7.2, and Debian 3.0. BYDSK_SYSTEM_IO ---------------------------------- The number of physical system reads or writes to this disk device during the interval. BYDSK_SYSTEM_IO_RATE ---------------------------------- The number of physical system reads or writes per second to this disk device during the interval. BYDSK_UTIL ---------------------------------- On HP-UX, this is the percentage of the time during the interval that the disk device had IO in progress from the point of view of the Operating System. In other words, the utilization or percentage of time busy servicing requests for this device. On the non-HP-UX systems, this is the percentage of the time that this disk device was busy transferring data during the interval. Some Linux kernels, typically 2.2 and older kernels, do not support the instrumentation needed to provide values for this metric. This metric will be “na” on the affected kernels. The “sar -d” command will also not be present on these systems. Distributions and OS releases that are known to be affected include: TurboLinux 7, SuSE 7.2, and Debian 3.0. This is a measure of the ability of the IO path to meet the transfer demands being placed on it. Slower disk devices may show a higher utilization with lower IO rates than faster disk devices such as disk arrays. A value of greater than 50% utilization over time may indicate that this device or its IO path is a bottleneck, and the access pattern of the workload, database, or files may need reorganizing for better balance of disk IO load. BYDSK_VM_IO ---------------------------------- The number of virtual memory IOs to this disk device during the interval. BYDSK_VM_IO_RATE ---------------------------------- The number of virtual memory IOs per second to this disk device during the interval. BYNETIF_COLLISION ---------------------------------- The number of physical collisions that occurred on the network interface during the interval. A rising rate of collisions versus outbound packets is an indication that the network is becoming increasingly congested. This metric does not currently include deferred packets. This data is not collected for non-broadcasting devices, such as loopback (lo), and is always zero. For HP-UX, this will be the same as the sum of the “Single Collision Frames”, “Multiple Collision Frames”, “Late Collisions”, and “Excessive Collisions” values from the output of the “lanadmin” utility for the network interface. Remember that “lanadmin” reports cumulative counts. As of the HP-UX 11.0 release and beyond, “netstat -i” shows network activity on the logical level (IP) only. For most other Unix systems, this is the same as the sum of the “Coll” column from the “netstat -i” command (“collisions” from the “netstat -i -e” command on Linux) for a network device. See also netstat(1). If BYNETIF_NET_TYPE is “ESXVLan”, then this metric will be N/A. AIX does not support the collision count for the ethernet interface. The collision count is supported for the token ring (tr) and loopback (lo) interfaces. For more information, please refer to the netstat(1) man page. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. BYNETIF_COLLISION_RATE ---------------------------------- The number of physical collisions per second on the network interface during the interval. A rising rate of collisions versus outbound packets is an indication that the network is becoming increasingly congested. This metric does not currently include deferred packets. This data is not collected for non-broadcasting devices, such as loopback (lo), and is always zero. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric will be N/A. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. BYNETIF_ERROR ---------------------------------- The number of physical errors that occurred on the network interface during the interval. An increasing number of errors may indicate a hardware problem in the network. On Unix systems, this data is not available for loop-back (lo) devices and is always zero. For HP-UX, this will be the same as the sum of the “Inbound Errors” and “Outbound Errors” values from the output of the “lanadmin” utility for the network interface. Remember that “lanadmin” reports cumulative counts. As of the HP-UX 11.0 release and beyond, “netstat -i” shows network activity on the logical level (IP) only. For all other Unix systems, this is the same as the sum of “Ierrs” (RX-ERR on Linux) and “Oerrs” (TX-ERR on Linux) from the “netstat -i” command for a network device. See also netstat(1). If BYNETIF_NET_TYPE is “ESXVLan”, then this metric will be N/A. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. BYNETIF_ERROR_RATE ---------------------------------- The number of physical errors per second on the network interface during the interval. On Unix systems, this data is not available for loop-back (lo) devices and is always zero. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric will be N/A. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. BYNETIF_IN_BYTE_RATE ---------------------------------- The number of KBs per second received from the network via this interface during the interval. Only the bytes in packets that carry data are included in this rate. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_IN_BYTE_RATE_CUM ---------------------------------- The average number of KBs per second received from the network via this interface over the cumulative collection time. Only the bytes in packets that carry data are included in this rate. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_IN_PACKET ---------------------------------- The number of successful physical packets received through the network interface during the interval. Successful packets are those that have been processed without errors or collisions. For HP-UX, this will be the same as the sum of the “Inbound Unicast Packets” and “Inbound Non-Unicast Packets” values from the output of the “lanadmin” utility for the network interface. Remember that “lanadmin” reports cumulative counts. As of the HP- UX 11.0 release and beyond, “netstat -i” shows network activity on the logical level (IP) only. For all other Unix systems, this is the same as the sum of the “Ipkts” column (RX-OK on Linux) from the “netstat -i” command for a network device. See also netstat(1). If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_IN_PACKET_RATE ---------------------------------- The number of successful physical packets per second received through the network interface during the interval. Successful packets are those that have been processed without errors or collisions. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_NAME ---------------------------------- The name of the network interface. For HP-UX 11.0 and beyond, these are the same names that appear in the “Description” field of the “lanadmin” command output. On all other Unix systems, these are the same names that appear in the “Name” column of the “netstat -i” command. Some examples of device names are: lo - loop-back driver ln - Standard Ethernet driver en - Standard Ethernet driver le - Lance Ethernet driver ie - Intel Ethernet driver tr - Token-Ring driver et - Ether Twist driver bf - fiber optic driver All of the device names will have the unit number appended to the name. For example, a loop-back device in unit 0 will be “lo0”. On vMA for Lan cards which are of type ESXVLan, this metric contains the vmnic as first half and the second half is the ESX host name. BYNETIF_NET_MTU ---------------------------------- The size of the maximum transfer unit (MTU) for this interface. BYNETIF_NET_SPEED ---------------------------------- The speed of this interface. This is the bandwidth in Mega bits/sec. BYNETIF_OUT_BYTE_RATE ---------------------------------- The number of KBs per second sent to the network via this interface during the interval. Only the bytes in packets that carry data are included in this rate. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_OUT_BYTE_RATE_CUM ---------------------------------- The average number of KBs per second sent to the network via this interface over the cumulative collection time. Only the bytes in packets that carry data are included in this rate. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_OUT_PACKET ---------------------------------- The number of successful physical packets sent through the network interface during the interval. Successful packets are those that have been processed without errors or collisions. For HP-UX, this will be the same as the sum of the “Outbound Unicast Packets” and “Outbound Non-Unicast Packets” values from the output of the “lanadmin” utility for the network interface. Remember that “lanadmin” reports cumulative counts. As of the HP- UX 11.0 release and beyond, “netstat -i” shows network activity on the logical level (IP) only. For all other Unix systems, this is the same as the sum of the “Opkts” column (TX-OK on Linux) from the “netstat -i” command for a network device. See also netstat(1). If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_OUT_PACKET_RATE ---------------------------------- The number of successful physical packets per second sent through the network interface during the interval. Successful packets are those that have been processed without errors or collisions. If BYNETIF_NET_TYPE is “ESXVLan”, then this metric shows the values for the Lan card in the host. Physical statistics are packets recorded by the network drivers. These numbers most likely will not be the same as the logical statistics. The values returned for the loopback interface will show “na” for the physical statistics since there is no network driver activity. Logical statistics are packets seen only by the Interface Protocol (IP) layer of the networking subsystem. Not all packets seen by IP will go out and come in through a network driver. An example is the loopback interface (127.0.0.1). Pings or other network generating commands (ftp, rlogin, and so forth) to 127.0.0.1 will not change physical driver statistics. Pings to IP addresses on remote systems will change physical driver statistics. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. BYNETIF_QUEUE ---------------------------------- The length of the outbound queue at the time of the last sample. This metric will be the same as the “Outbound Queue Length” values from the output of “lanadmin” utility. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On HP-UX, this metric is only available for LAN interfaces. For WAN (Wide-Area Network) interfaces such as ATM and X.25, with interface names such as el, cip/ixe, and netisdn, this metric returns “na”. DATE ---------------------------------- The date the information in this record was captured, based on local time. The date is an ASCII field in mm/dd/yyyy format unless localized. If localized, the separators may be different and the subfield may be in a different sequence. In ASCII files this field will always contain 10 characters. Each subfield (mm, dd, yyyy) will contain a leading zero if the value is less than 10. This metric is extracted from GBL_STATTIME, which is obtained using the time() system call at the time of data collection. This field responds to language localization. For example, in Italy the field would appear as dd/mm/yyyy and in Japan it would be yyyy/mm/dd. In binary files this field is in MPE CALENDAR format in the least significant 16 bits of the field. The most significant 16 bits should all be zero. Dividing the field by 512 will isolate the year (that is, 94). This field MOD 512 will isolate the day of the year. DATE_SECONDS ---------------------------------- The time that the data in this record was captured, expressed in seconds since January 1, 1970, based on local time. This is related to the standard time-stamp returned by the unix system call time(), but has had the local time zone correction applied. DAY ---------------------------------- The julian day of the year that the data in this record was captured. This metric is extracted from GBL_STATTIME. FS_BLOCK_SIZE ---------------------------------- The maximum block size of this file system, in bytes. A value of “na” may be displayed if the file system is not mounted. If the product is restarted, these unmounted file systems are not displayed until remounted. FS_DEVNAME ---------------------------------- On Unix systems, this is the path name string of the current device. On Windows, this is the disk drive string of the current device. On HP-UX, this is the “fsname” parameter in the mount(1M) command. For NFS devices, this includes the name of the node exporting the file system. It is possible that a process may mount a device using the mount(2) system call. This call does not update the “/etc/mnttab” and its name is blank. This situation is rare, and should be corrected by syncer(1M). Note that once a device is mounted, its entry is displayed, even after the device is unmounted, until the midaemon process terminates. On SUN, this is the path name string of the current device, or “tmpfs” for memory based file systems. See tmpfs(7). FS_DIRNAME ---------------------------------- On Unix systems, this is the path name of the mount point of the file system. On Windows, this is the drive letter associated with the selected disk partition. On HP-UX, this is the path name of the mount point of the file system if the logical volume has a mounted file system. This is the directory parameter of the mount(1M) command for most entries. Exceptions are: * For lvm swap areas, this field contains “lvm swap device”. * For logical volumes with no mounted file systems, this field contains “Raw Logical Volume” (relevant only to Perf Agent). On HP-UX, the file names are in the same order as shown in the “/usr/sbin/mount -p” command. File systems are not displayed until they exhibit IO activity once the midaemon has been started. Also, once a device is displayed, it continues to be displayed (even after the device is unmounted) until the midaemon process terminates. On SUN, only “UFS”, “HSFS” and “TMPFS” file systems are listed. See mount(1M) and mnttab(4). “TMPFS” file systems are memory based filesystems and are listed here for convenience. See tmpfs(7). On AIX, see mount(1M) and filesystems(4). On OSF1, see mount(2). FS_FRAG_SIZE ---------------------------------- The fundamental file system block size, in bytes. A value of “na” may be displayed if the file system is not mounted. If the product is restarted, these unmounted file systems are not displayed until remounted. FS_MAX_SIZE ---------------------------------- Maximum number that this file system could obtain if full, in MB. Note that this is the user space capacity - it is the file system space accessible to non root users. On most Unix systems, the df command shows the total file system capacity which includes the extra file system space accessible to root users only. The equivalent fields to look at are “used” and “avail”. For the target file system, to calculate the maximum size in MB, use FS Max Size = (used + avail)/1024 A value of “na” may be displayed if the file system is not mounted. If the product is restarted, these unmounted file systems are not displayed until remounted. On HP-UX, this metric is updated at 4 minute intervals to minimize collection overhead. FS_SPACE_UTIL ---------------------------------- Percentage of the file system space in use during the interval. Note that this is the user space capacity - it is the file system space accessible to non root users. On most Unix systems, the df command shows the total file system capacity which includes the extra file system space accessible to root users only. A value of “na” may be displayed if the file system is not mounted. If the product is restarted, these unmounted file systems are not displayed until remounted. On HP-UX, this metric is updated at 4 minute intervals to minimize collection overhead. FS_TYPE ---------------------------------- A string indicating the file system type. On Unix systems, some of the possible types are: hfs - user file system ufs - user file system ext2 - user file system cdfs - CD-ROM file system vxfs - Veritas (vxfs) file system nfs - network file system nfs3 - network file system Version 3 On Windows, some of the possible types are: NTFS - New Technology File System FAT - 16-bit File Allocation Table FAT32 - 32-bit File Allocation Table FAT uses a 16-bit file allocation table entry (216 clusters). FAT32 uses a 32-bit file allocation table entry. However, Windows 2000 reserves the first 4 bits of a FAT32 file allocation table entry, which means FAT32 has a theoretical maximum of 228 clusters. NTFS is native file system of Windows NT and beyond. GBL_ACTIVE_CPU ---------------------------------- The number of CPUs online on the system. For HP-UX and certain versions of Linux, the sar(1M) command allows you to check the status of the system CPUs. For SUN and DEC, the commands psrinfo(1M) and psradm(1M) allow you to check or change the status of the system CPUs. For AIX, the pstat(1) command allows you to check the status of the system CPUs. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment if RSET is not configured for the System WPAR. If RSET is configured for the System WPAR, this metric value will report the number of CPUs in the RSET. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_ACTIVE_PROC ---------------------------------- An active process is one that exists and consumes some CPU time. GBL_ACTIVE_PROC is the sum of the alive-process-time/interval-time ratios of every process that is active (uses any CPU time) during an interval. The following diagram of a four second interval during which two processes exist on the system should be used to understand the above definition. Note the difference between active processes, which consume CPU time, and alive processes which merely exist on the system. ----------- Seconds ----------- 1 2 3 4 Proc ---- ---- ---- ---- ---- A live live live live B live/CPU live/CPU live dead Process A is alive for the entire four second interval but consumes no CPU. A’s contribution to GBL_ALIVE_PROC is 4*1/4. A contributes 0*1/4 to GBL_ACTIVE_PROC. B’s contribution to GBL_ALIVE_PROC is 3*1/4. B contributes 2*1/4 to GBL_ACTIVE_PROC. Thus, for this interval, GBL_ACTIVE_PROC equals 0.5 and GBL_ALIVE_PROC equals 1.75. Because a process may be alive but not active, GBL_ACTIVE_PROC will always be less than or equal to GBL_ALIVE_PROC. This metric is a good overall indicator of the workload of the system. An unusually large number of active processes could indicate a CPU bottleneck. To determine if the CPU is a bottleneck, compare this metric with GBL_CPU_TOTAL_UTIL and GBL_RUN_QUEUE. If GBL_CPU_TOTAL_UTIL is near 100 percent and GBL_RUN_QUEUE is greater than one, there is a bottleneck. On non HP-UX systems, this metric is derived from sampled process data. Since the data for a process is not available after the process has died on this operating system, a process whose life is shorter than the sampling interval may not be seen when the samples are taken. Thus this metric may be slightly less than the actual value. Increasing the sampling frequency captures a more accurate count, but the overhead of collection may also rise. GBL_ALIVE_PROC ---------------------------------- An alive process is one that exists on the system. GBL_ALIVE_PROC is the sum of the alive-process-time/interval-time ratios for every process. The following diagram of a four second interval during which two processes exist on the system should be used to understand the above definition. Note the difference between active processes, which consume CPU time, and alive processes which merely exist on the system. ----------- Seconds ----------- 1 2 3 4 Proc ---- ---- ---- ---- ---- A live live live live B live/CPU live/CPU live dead Process A is alive for the entire four second interval but consumes no CPU. A’s contribution to GBL_ALIVE_PROC is 4*1/4. A contributes 0*1/4 to GBL_ACTIVE_PROC. B’s contribution to GBL_ALIVE_PROC is 3*1/4. B contributes 2*1/4 to GBL_ACTIVE_PROC. Thus, for this interval, GBL_ACTIVE_PROC equals 0.5 and GBL_ALIVE_PROC equals 1.75. Because a process may be alive but not active, GBL_ACTIVE_PROC will always be less than or equal to GBL_ALIVE_PROC. On non HP-UX systems, this metric is derived from sampled process data. Since the data for a process is not available after the process has died on this operating system, a process whose life is shorter than the sampling interval may not be seen when the samples are taken. Thus this metric may be slightly less than the actual value. Increasing the sampling frequency captures a more accurate count, but the overhead of collection may also rise. GBL_COLLECTOR ---------------------------------- ASCII field containing collector name and version. The collector name will appear as either “SCOPE/xx V.UU.FF.LF” or “Coda RV.UU.FF.LF”. xx identifies the platform; V = version, UU = update level, FF = fix level, and LF = lab fix id. For example, SCOPE/UX C.04.00.00; or Coda A.07.10.04. GBL_COMPLETED_PROC ---------------------------------- The number of processes that terminated during the interval. On non HP-UX systems, this metric is derived from sampled process data. Since the data for a process is not available after the process has died on this operating system, a process whose life is shorter than the sampling interval may not be seen when the samples are taken. Thus this metric may be slightly less than the actual value. Increasing the sampling frequency captures a more accurate count, but the overhead of collection may also rise. GBL_CPU_CSWITCH_TIME ---------------------------------- The time, in seconds, that the CPU spent context switching during the interval. On HP-UX, this includes context switches that result in the execution of a different process and those caused by a process stopping, then resuming, with no other process running in the meantime. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_CSWITCH_UTIL ---------------------------------- The percentage of time that the CPU spent context switching during the interval. On HP-UX, this includes context switches that result in the execution of a different process and those caused by a process stopping, then resuming, with no other process running in the meantime. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_HISTOGRAM ---------------------------------- Histogram of CPU utilization components. Shows breakout: GBL_CPU_TOTAL_UTIL = GBL_CPU_SYSCALL_UTIL + GBL_CPU_NICE_UTIL + GBL_CPU_NORMAL_UTIL + GBL_CPU_REALTIME_UTIL + GBL_CPU_CSWITCH_UTIL + GBL_CPU_INTERRUPT_UTIL ASCII and BINARY files contain a line of ASCII characters that make up one row of a printed histogram. This can be a quick way to get a graphical view of CPU usage on a character-mode terminal display. GBL_CPU_IDLE_TIME ---------------------------------- The time, in seconds, that the CPU was idle during the interval. This is the total idle time, including waiting for I/O. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. On AIX System WPARs, this metric value is calculated against physical cpu time. On Solaris non-global zones, this metric is N/A. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_IDLE_UTIL ---------------------------------- The percentage of time that the CPU was idle during the interval. This is the total idle time, including waiting for I/O. On Unix systems, this is the same as the sum of the “%idle” and “%wio” fields reported by the “sar -u” command. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. On Solaris non-global zones, this metric is N/A. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_INTERRUPT_TIME ---------------------------------- The time, in seconds, that the CPU spent processing interrupts during the interval. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On Hyper-V host, this metric is NA. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_INTERRUPT_UTIL ---------------------------------- The percentage of time that the CPU spent processing interrupts during the interval. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On Hyper-V host, this metric is NA. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_NICE_TIME ---------------------------------- The time, in seconds, that the CPU was in user mode at a nice priority during the interval. On HP-UX, the NICE metrics include positive nice value CPU time only. Negative nice value CPU is broken out into NNICE (negative nice) metrics. Positive nice values range from 20 to 39. Negative nice values range from 0 to 19. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_NICE_UTIL ---------------------------------- The percentage of time that the CPU was in user mode at a nice priority during the interval. On HP-UX, the NICE metrics include positive nice value CPU time only. Negative nice value CPU is broken out into NNICE (negative nice) metrics. Positive nice values range from 20 to 39. Negative nice values range from 0 to 19. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_NORMAL_TIME ---------------------------------- The time, in seconds, that the CPU was in user mode at normal priority during the interval. Normal priority user mode CPU excludes CPU used at real-time and nice priorities. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_NORMAL_UTIL ---------------------------------- The percentage of time that the CPU was in user mode at normal priority during the interval. Normal priority user mode CPU excludes CPU used at real-time and nice priorities. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_REALTIME_TIME ---------------------------------- The time, in seconds, that the CPU was in user mode at a realtime priority during the interval. Running at a realtime priority means that the process or kernel thread was run using the rtprio command or the rtprio system call to alter its priority. Realtime priorities range from zero to 127 and are absolute priorities, meaning the realtime process with the lowest priority runs as long as it wants to. Since this can have a huge impact on the system, the realtime CPU is tracked separately to make visible the effect of using realtime priorities. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_REALTIME_UTIL ---------------------------------- The percentage of time that the CPU was in user mode at a realtime priority during the interval. Running at a realtime priority means that the process or kernel thread was run using the rtprio command or the rtprio system call to alter its priority. Realtime priorities range from zero to 127 and are absolute priorities, meaning the realtime process with the lowest priority runs as long as it wants to. Since this can have a huge impact on the system, the realtime CPU is tracked separately to make visible the effect of using realtime priorities. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_SYSCALL_TIME ---------------------------------- The time, in seconds, that the CPU was in system mode (excluding interrupt, context switch, trap, or vfault CPU) during the interval. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_SYSCALL_UTIL ---------------------------------- The percentage of time that the CPU was in system mode (excluding interrupt, context switch, trap, or vfault CPU) during the interval. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_SYS_MODE_TIME ---------------------------------- The time, in seconds, that the CPU was in system mode during the interval. A process operates in either system mode (also called kernel mode on Unix or privileged mode on Windows) or user mode. When a process requests services from the operating system with a system call, it switches into the machine’s privileged protection mode and runs in system mode. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. On AIX System WPARs, this metric value is calculated against physical cpu time. On Hyper-V host, this metric indicates the time spent in Hypervisor code. GBL_CPU_SYS_MODE_UTIL ---------------------------------- Percentage of time the CPU was in system mode during the interval. A process operates in either system mode (also called kernel mode on Unix or privileged mode on Windows) or user mode. When a process requests services from the operating system with a system call, it switches into the machine’s privileged protection mode and runs in system mode. This metric is a subset of the GBL_CPU_TOTAL_UTIL percentage. This is NOT a measure of the amount of time used by system daemon processes, since most system daemons spend part of their time in user mode and part in system calls, like any other process. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. High system mode CPU percentages are normal for IO intensive applications. Abnormally high system mode CPU percentages can indicate that a hardware problem is causing a high interrupt rate. It can also indicate programs that are not calling system calls efficiently. On a logical system, this metric indicates the percentage of time the logical processor was in kernel mode during this interval. On Hyper-V host, this metric indicates the percentage of time spent in Hypervisor code. GBL_CPU_TOTAL_TIME ---------------------------------- The total time, in seconds, that the CPU was not idle in the interval. This is calculated as GBL_CPU_TOTAL_TIME = GBL_CPU_USER_MODE_TIME + GBL_CPU_SYS_MODE_TIME On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. On AIX System WPARs, this metric value is calculated against physical cpu time. GBL_CPU_TOTAL_UTIL ---------------------------------- Percentage of time the CPU was not idle during the interval. This is calculated as GBL_CPU_TOTAL_UTIL = GBL_CPU_USER_MODE_UTIL + GBL_CPU_SYS_MODE_UTIL On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. GBL_CPU_TOTAL_UTIL + GBL_CPU_IDLE_UTIL = 100% This metric varies widely on most systems, depending on the workload. A consistently high CPU utilization can indicate a CPU bottleneck, especially when other indicators such as GBL_RUN_QUEUE and GBL_ACTIVE_PROC are also high. High CPU utilization can also occur on systems that are bottlenecked on memory, because the CPU spends more time paging and swapping. NOTE: On Windows, this metric may not equal the sum of the APP_CPU_TOTAL_UTIL metrics. Microsoft states that “this is expected behavior” because this GBL_CPU_TOTAL_UTIL metric is taken from the performance library Processor objects while the APP_CPU_TOTAL_UTIL metrics are taken from the Process objects. Microsoft states that there can be CPU time accounted for in the Processor system objects that may not be seen in the Process objects. On a logical system, this metric indicates the logical utilization with respect to number of processors available for the logical system (GBL_NUM_CPU). On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. GBL_CPU_USER_MODE_TIME ---------------------------------- The time, in seconds, that the CPU was in user mode during the interval. User CPU is the time spent in user mode at a normal priority, at real-time priority (on HP-UX, AIX, and Windows systems), and at a nice priority. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. On AIX System WPARs, this metric value is calculated against physical cpu time. On Hyper-V host, this metric indicates the time spent in guest code. GBL_CPU_USER_MODE_UTIL ---------------------------------- The percentage of time the CPU was in user mode during the interval. User CPU is the time spent in user mode at a normal priority, at real-time priority (on HP-UX, AIX, and Windows systems), and at a nice priority. This metric is a subset of the GBL_CPU_TOTAL_UTIL percentage. On a system with multiple CPUs, this metric is normalized. That is, the CPU used over all processors is divided by the number of processors online. This represents the usage of the total processing capacity available. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. High user mode CPU percentages are normal for computation- intensive applications. Low values of user CPU utilization compared to relatively high values for GBL_CPU_SYS_MODE_UTIL can indicate an application or hardware problem. On a logical system, this metric indicates the percentage of time the logical processor was in user mode during this interval. On Hyper-V host, this metric indicates the percentage of time spent in guest code. GBL_DISK_FS_IO ---------------------------------- The total of physical file system disk reads and writes during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical IOs generated by user file system access and do not include virtual memory IOs, system IOs (inode updates), or IOs relating to raw disk access. An exception is user files accessed via the mmap(2) call, which will not show their physical IOs in this category. They appear under virtual memory IOs. GBL_DISK_FS_IO_RATE ---------------------------------- The total of file system disk physical reads and writes per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical IOs generated by user file system access and do not include virtual memory IOs, system IOs (inode updates), or IOs relating to raw disk access. An exception is user files accessed via the mmap(2) call, which will not show their physical IOs in this category. They appear under virtual memory IOs. GBL_DISK_FS_READ ---------------------------------- The number of file system disk reads during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical reads generated by user file system access and do not include virtual memory reads, system reads (inode access), or reads relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical reads in this category. They appear under virtual memory reads. GBL_DISK_FS_READ_RATE ---------------------------------- The number of file system disk reads per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical reads generated by user file system access and do not include virtual memory reads, system reads (inode access), or reads relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical reads in this category. They appear under virtual memory reads. GBL_DISK_FS_WRITE ---------------------------------- The number of file system disk writes during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical writes generated by user file system access and do not include virtual memory writes, system writes (inode updates), or writes relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical writes in this category. They appear under virtual memory writes. GBL_DISK_FS_WRITE_RATE ---------------------------------- The number of file system disk writes per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical writes generated by user file system access and do not include virtual memory writes, system writes (inode updates), or writes relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical writes in this category. They appear under virtual memory writes. GBL_DISK_HISTOGRAM ---------------------------------- Histogram of physical Disk IO rate components. On HP-UX, this shows a breakout of: GBL_DISK_PHYS_IO_RATE = GBL_DISK_VM_IO_RATE + GBL_DISK_SYSTEM_IO_RATE + GBL_DISK_FS_IO_RATE + GBL_DISK_RAW_IO_RATE On SUN systems, this shows a breakout of: GBL_DISK_PHYS_IO_RATE = GBL_DISK_BLOCK_READ_RATE + GBL_DISK_BLOCK_WRITE_RATE + GBL_DISK_RAW_READ_RATE + GBL_DISK_RAW_WRITE_RATE + GBL_DISK_VM_IO_RATE On the remaining Unix systems, this shows a breakout of: GBL_DISK_PHYS_IO_RATE = GBL_DISK_BLOCK_IO_RATE + GBL_DISK_VM_IO_RATE + GBL_DISK_RAW_IO_RATE On Windows, this shows a breakout of: GBL_DISK_PHYS_IO_RATE = GBL_DISK_PHYS_READ_RATE + GBL_DISK_PHYS_WRITE_RATE ASCII and BINARY files contain a line of ASCII characters that make up one row of a printed histogram. This can be a quick way to get a graphical view of Disk usage on a character-mode terminal display. GBL_DISK_LOGL_IO ---------------------------------- The number of logical IOs made during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read and write system calls that are directed to disk devices. Also counted are read and write system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, writev, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. GBL_DISK_LOGL_IO_RATE ---------------------------------- The number of logical IOs per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read and write system calls that are directed to disk devices. Also counted are read and write system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, writev, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. GBL_DISK_LOGL_READ ---------------------------------- On most systems, this is the number of logical reads made during the interval. On SUN, this is the number of logical block reads made during the interval. On Windows, this includes both buffered (cached) read requests and unbuffered reads. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. GBL_DISK_LOGL_READ_BYTE ---------------------------------- The number of KBs transferred through logical reads during the last interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. GBL_DISK_LOGL_READ_BYTE_RATE ---------------------------------- The number of KBs transferred per second via logical reads during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. GBL_DISK_LOGL_READ_RATE ---------------------------------- On most systems, this is The average number of logical reads per second made during the interval. On SUN, this is the average number of logical block reads per second made during the interval. On Windows, this includes both buffered (cached) read requests and unbuffered reads. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_LOGL_WRITE ---------------------------------- On most systems, this is the number of logical writes made during the interval. On SUN, this is the number of logical block writes during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. GBL_DISK_LOGL_WRITE_BYTE ---------------------------------- The number of KBs transferred via logical writes during the last interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. GBL_DISK_LOGL_WRITE_BYTE_RATE ---------------------------------- The number of KBs per second transferred via logical writes during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. GBL_DISK_LOGL_WRITE_RATE ---------------------------------- On most systems, this is the average number of logical writes per second made during the interval. On SUN, this is the average number of logical block writes per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_PHYS_BYTE ---------------------------------- The number of KBs transferred to and from disks during the interval. The bytes for all types of physical IOs are counted. Only local disks are counted in this measurement. NFS devices are excluded. It is not directly related to the number of IOs, since IO requests can be of differing lengths. On Unix systems, this includes file system IO, virtual memory IO, and raw IO. On Windows, all types of physical IOs are counted. On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_BYTE_RATE ---------------------------------- The average number of KBs per second at which data was transferred to and from disks during the interval. The bytes for all types physical IOs are counted. Only local disks are counted in this measurement. NFS devices are excluded. This is a measure of the physical data transfer rate. It is not directly related to the number of IOs, since IO requests can be of differing lengths. This is an indicator of how much data is being transferred to and from disk devices. Large spikes in this metric can indicate a disk bottleneck. On Unix systems, all types of physical disk IOs are counted, including file system, virtual memory, and raw reads. On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_IO ---------------------------------- The number of physical IOs during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Unix systems, all types of physical disk IOs are counted, including file system IO, virtual memory IO and raw IO. On HP-UX, this is calculated as GBL_DISK_PHYS_IO = GBL_DISK_FS_IO + GBL_DISK_VM_IO + GBL_DISK_SYSTEM_IO + GBL_DISK_RAW_IO On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_IO_RATE ---------------------------------- The number of physical IOs per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Unix systems, all types of physical disk IOs are counted, including file system IO, virtual memory IO and raw IO. On HP-UX, this is calculated as GBL_DISK_PHYS_IO_RATE = GBL_DISK_FS_IO_RATE + GBL_DISK_VM_IO_RATE + GBL_DISK_SYSTEM_IO_RATE + GBL_DISK_RAW_IO_RATE On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_READ ---------------------------------- The number of physical reads during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Unix systems, all types of physical disk reads are counted, including file system, virtual memory, and raw reads. On HP-UX, there are many reasons why there is not a direct correlation between the number of logical IOs and physical IOs. For example, small sequential logical reads may be satisfied from the buffer cache, resulting in fewer physical IOs than logical IOs. Conversely, large logical IOs or small random IOs may result in more physical than logical IOs. Logical volume mappings, logical disk mirroring, and disk striping also tend to remove any correlation. On HP-UX, this is calculated as GBL_DISK_PHYS_READ = GBL_DISK_FS_READ + GBL_DISK_VM_READ + GBL_DISK_SYSTEM_READ + GBL_DISK_RAW_READ On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_READ_BYTE_RATE ---------------------------------- The average number of KBs transferred from the disk per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_READ_RATE ---------------------------------- The number of physical reads per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Unix systems, all types of physical disk reads are counted, including file system, virtual memory, and raw reads. On HP-UX, this is calculated as GBL_DISK_PHYS_READ_RATE = GBL_DISK_FS_READ_RATE + GBL_DISK_VM_READ_RATE + GBL_DISK_SYSTEM_READ_RATE + GBL_DISK_RAW_READ_RATE On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_WRITE ---------------------------------- The number of physical writes during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Unix systems, all types of physical disk writes are counted, including file system IO, virtual memory IO, and raw writes. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On HP-UX, there are many reasons why there is not a direct correlation between logical IOs and physical IOs. For example, small logical writes may end up entirely in the buffer cache, and later generate fewer physical IOs when written to disk due to the larger IO size. Or conversely, small logical writes may require physical prefetching of the corresponding disk blocks before the data is merged and posted to disk. Logical volume mappings, logical disk mirroring, and disk striping also tend to remove any correlation. On HP-UX, this is calculated as GBL_DISK_PHYS_WRITE = GBL_DISK_FS_WRITE + GBL_DISK_VM_WRITE + GBL_DISK_SYSTEM_WRITE + GBL_DISK_RAW_WRITE On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_WRITE_BYTE_RATE ---------------------------------- The average number of KBs transferred to the disk per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Unix systems, all types of physical disk writes are counted, including file system IO, virtual memory IO, and raw writes. On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_PHYS_WRITE_RATE ---------------------------------- The number of physical writes per second during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Unix systems, all types of physical disk writes are counted, including file system IO, virtual memory IO, and raw writes. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On HP-UX, this is calculated as GBL_DISK_PHYS_WRITE_RATE = GBL_DISK_FS_WRITE_RATE + GBL_DISK_VM_WRITE_RATE + GBL_DISK_SYSTEM_WRITE_RATE + GBL_DISK_RAW_WRITE_RATE On SUN, if a CD drive is powered off, or no CD is inserted in the CD drive at boottime, the operating system does not provide performance data for that device. This can be determined by checking the “by-disk” data when provided in a product. If the CD drive has an entry in the list of active disks on a system, then data for that device is being collected. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_RAW_IO ---------------------------------- The total number of raw reads and writes during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On Sun, tape drive accesses are included in raw IOs, but not in physical IOs. To determine if raw IO is tape access versus disk access, compare the global physical disk accesses to the total raw, block, and vm IOs. If the totals are the same, the raw IO activity is to a disk, floppy, or CD drive. Check physical IO data for each individual disk device to isolate a device. If the totals are different, there is raw IO activity to a non-disk device like a tape drive. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_RAW_IO_RATE ---------------------------------- The total number of raw reads and writes per second during the interval. Only accesses to local disk devices are counted. On Sun, tape drive accesses are included in raw IOs, but not in physical IOs. To determine if raw IO is tape access versus disk access, compare the global physical disk accesses to the total raw, block, and vm IOs. If the totals are the same, the raw IO activity is to a disk, floppy, or CD drive. Check physical IO data for each individual disk device to isolate a device. If the totals are different, there is raw IO activity to a non-disk device like a tape drive. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_RAW_READ ---------------------------------- The number of raw reads during the interval. Only accesses to local disk devices are counted. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_RAW_READ_RATE ---------------------------------- The number of raw reads per second during the interval. Only accesses to local disk devices are counted. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_RAW_WRITE ---------------------------------- The number of raw writes during the interval. Only accesses to local disk devices are counted. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_RAW_WRITE_RATE ---------------------------------- The number of raw writes per second during the interval. Only accesses to local disk devices are counted. On Sun, tape drive accesses are included in raw IOs, but not in physical IOs. To determine if raw IO is tape access versus disk access, compare the global physical disk accesses to the total raw, block, and vm IOs. If the totals are the same, the raw IO activity is to a disk, floppy, or CD drive. Check physical IO data for each individual disk device to isolate a device. If the totals are different, there is raw IO activity to a non-disk device like a tape drive. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. GBL_DISK_SUBSYSTEM_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on the disk subsystem (in a “queue” waiting for their file system disk IO to complete) during the interval. This is the sum of processes or kernel threads in the DISK, INODE, CACHE and CDFS wait states. Processes or kernel threads doing raw IO to a disk are not included in this measurement. As this number rises, it is an indication of a disk bottleneck. This is calculated as the accumulated time that all processes or kernel threads spent blocked on (DISK + INODE + CACHE + CDFS) divided by the interval time. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_DISK_SYSTEM_IO ---------------------------------- The number of physical disk IOs generated by the kernel for file system management (inode accesses or updates) during the interval. Only local disks are counted in this measurement. NFS devices are excluded. GBL_DISK_SYSTEM_IO_RATE ---------------------------------- The number of physical disk IOs per second generated by the kernel for file system management (inode accesses or updates) during the interval. Only local disks are counted in this measurement. NFS devices are excluded. GBL_DISK_SYSTEM_READ ---------------------------------- Number of physical disk reads generated by the kernel for file system management (inode accesses) during the interval. Only local disks are counted in this measurement. NFS devices are excluded. GBL_DISK_SYSTEM_READ_RATE ---------------------------------- Number of physical disk reads per second generated by the kernel for file system management (inode accesses) during the interval. Only local disks are counted in this measurement. NFS devices are excluded. GBL_DISK_SYSTEM_WRITE ---------------------------------- Number of physical disk writes generated by the kernel for file system management (inode updates) during the interval. Only local disks are counted in this measurement. NFS devices are excluded. GBL_DISK_SYSTEM_WRITE_RATE ---------------------------------- Number of physical disk writes per second generated by the kernel for file system management (inode updates) during the interval. Only local disks are counted in this measurement. NFS devices are excluded. GBL_DISK_TIME_PEAK ---------------------------------- The time, in seconds, during the interval that the busiest disk was performing IO transfers. This is for the busiest disk only, not all disk devices. This counter is based on an end-to-end measurement for each IO transfer updated at queue entry and exit points. Only local disks are counted in this measurement. NFS devices are excluded. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_UTIL_PEAK ---------------------------------- The utilization of the busiest disk during the interval. On HP-UX, this is the percentage of time during the interval that the busiest disk device had IO in progress from the point of view of the Operating System. On all other systems, this is the percentage of time during the interval that the busiest disk was performing IO transfers. It is not an average utilization over all the disk devices. Only local disks are counted in this measurement. NFS devices are excluded. Some Linux kernels, typically 2.2 and older kernels, do not support the instrumentation needed to provide values for this metric. This metric will be “na” on the affected kernels. The “sar -d” command will also not be present on these systems. Distributions and OS releases that are known to be affected include: TurboLinux 7, SuSE 7.2, and Debian 3.0. A peak disk utilization of more than 50 percent often indicates a disk IO subsystem bottleneck situation. A bottleneck may not be in the physical disk drive itself, but elsewhere in the IO path. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_VM_IO ---------------------------------- The total number of virtual memory IOs made during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On HP-UX, the IOs to user file data are not included in this metric unless they were done via the mmap(2) system call. On SUN, when a file is accessed, it is memory mapped by the operating system. Accesses generate virtual memory IOs. Reading a file generates block IOs as the file’s inode information is cached. File writes are a combination of posting to memory mapped allocations (VM IOs) and posting updated inode information to disk (block IOs). On SUN, this metric is calculated by subtracting raw and block IOs from physical IOs. Tape drive accesses are included in the raw IOs, but not in the physical IOs. Therefore, when tape drive accesses are occurring on a system, all virtual memory and raw IO is counted as raw IO. For example, you may see heavy raw IO occurring during system backup. Raw IOs for disks are counted in the physical IOs. To determine if the raw IO is tape access versus disk access, compare the global physical disk accesses to the total of raw, block, and VM IOs. If the totals are the same, the raw IO activity is to a disk, floppy, or CD drive. Check physical IO data for each individual disk device to isolate a device. If the totals are different, there is raw IO activity to a non-disk device like a tape drive. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_VM_IO_RATE ---------------------------------- The number of virtual memory IOs per second made during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On HP-UX, the IOs to user file data are not included in this metric unless they were done via the mmap(2) system call. On SUN, when a file is accessed, it is memory mapped by the operating system. Accesses generate virtual memory IOs. Reading a file generates block IOs as the file’s inode information is cached. File writes are a combination of posting to memory mapped allocations (VM IOs) and posting updated inode information to disk (block IOs). On SUN, this metric is calculated by subtracting raw and block IOs from physical IOs. Tape drive accesses are included in the raw IOs, but not in the physical IOs. Therefore, when tape drive accesses are occurring on a system, all virtual memory and raw IO is counted as raw IO. For example, you may see heavy raw IO occurring during system backup. Raw IOs for disks are counted in the physical IOs. To determine if the raw IO is tape access versus disk access, compare the global physical disk accesses to the total of raw, block, and VM IOs. If the totals are the same, the raw IO activity is to a disk, floppy, or CD drive. Check physical IO data for each individual disk device to isolate a device. If the totals are different, there is raw IO activity to a non-disk device like a tape drive. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_DISK_VM_READ ---------------------------------- The number of virtual memory reads made during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On HP-UX, the reads to user file data are not included in this metric unless they were accessed via the mmap(2) system call. On AIX System WPARs, this metric is NA. GBL_DISK_VM_READ_RATE ---------------------------------- The number of virtual memory reads per second made during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On HP-UX, the reads to user file data are not included in this metric unless they were accessed via the mmap(2) system call. On AIX System WPARs, this metric is NA. GBL_DISK_VM_WRITE ---------------------------------- The number of virtual memory writes made during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On HP-UX, the writes to user file data are not included in this metric unless they were done via the mmap(2) system call. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On AIX System WPARs, this metric is NA. GBL_DISK_VM_WRITE_RATE ---------------------------------- The number of virtual memory writes per second made during the interval. Only local disks are counted in this measurement. NFS devices are excluded. On HP-UX, the writes to user file data are not included in this metric unless they were done via the mmap(2) system call. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On AIX System WPARs, this metric is NA. GBL_FS_SPACE_UTIL_PEAK ---------------------------------- The percentage of occupied disk space to total disk space for the fullest file system found during the interval. Only locally mounted file systems are counted in this metric. This metric can be used as an indicator that at least one file system on the system is running out of disk space. On Unix systems, CDROM and PC file systems are also excluded. This metric can exceed 100 percent. This is because a portion of the file system space is reserved as a buffer and can only be used by root. If the root user has made the file system grow beyond the reserved buffer, the utilization will be greater than 100 percent. This is a dangerous situation since if the root user totally fills the file system, the system may crash. On Windows, CDROM file systems are also excluded. On Solaris non-global zones, this metric shows data from the global zone. GBL_IPC_SUBSYSTEM_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on the InterProcess Communication (IPC) subsystems (waiting for their interprocess communication activity to complete) during the interval. This is the sum of processes or kernel threads in the IPC, MSG, SEM, PIPE, SOCKT (that is, sockets) and STRMS (that is, streams IO) wait states. This is calculated as the accumulated time that all processes or kernel threads spent blocked on (IPC + MSG + SEM + PIPE + SOCKT + STRMS) divided by the interval time. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_LOGFILE_VERSION ---------------------------------- Three byte ASCII field containing the log file version number. The log file version is assigned by scopeux and is incremented when changes to the log file causes the layout to be different from previous versions. The current version is “ D”. Every effort is made to protect the information investment maintained in historical log files by providing forward compatibility and/or conversion utilities when log files change. GBL_LOGGING_TYPES ---------------------------------- A 13-byte field indicating the types of data logged by the collector. This is controlled by the LOG statement in the parm file. Each position will contain either a space or the characters as shown below. Note that positions two (all applications) and four (all processes) were implemented for HP internal use only and are not normally used outside of HP. An @ in position two indicates that all applications are logged each five minute interval even if they had no activity during the interval. An @ in position four indicates that all processes, not just the interesting ones, are logged each one minute interval. This can result in very large log files.An @ in position 6 indicates all devices( File System Device,Disk,CPU,LAN,Logical Volume) are logged. Position Char Meaning 1 G Global data 2 @ All applications 3 A Applications 4 @ All processes 5 P Interesting processes 6 @ All Devices 7 F File System Device 8 D Disk 9 C CPU 10 L LAN 11 V Logical Volume 12 T Transaction data 13 space Not used By default, global, interesting process, LAN data is logged, in which case this field would be “ G P L”. GBL_LOST_MI_TRACE_BUFFERS ---------------------------------- The number of trace buffers lost by the measurement processing daemon. On HP-UX systems, if this value is > 0, the measurement subsystem is not keeping up with the system events that generate traces. For other Unix systems, if this value is > 0, the measurement subsystem is not keeping up with the ARM API calls that generate traces. Note: The value reported for this metric will roll over to 0 once it crosses INTMAX. GBL_MACHINE ---------------------------------- An ASCII string representing the Processor Architecture. And machine hardware model is represented by GBL_MACHINE_MODEL metric. GBL_MACHINE_MODEL ---------------------------------- The CPU model. This is similar to the information returned by the GBL_MACHINE metric and the uname command(except for Solaris 10 x86/x86_64). However, this metric returns more information on some processors. On HP-UX, this is the same information returned by the model command. GBL_MEM_ACTIVE_VIRT_UTIL ---------------------------------- The percentage of total virtual memory active at the end of the interval. Active virtual memory is the virtual memory associated with processes that are currently on the run queue or processes that have executed recently. This is the sum of the virtual memory sizes of the data and stack regions for these processes. On HP-UX, this is the sum of the virtual memory of all processes which have had a thread run in the last 20 seconds. GBL_MEM_AVAIL ---------------------------------- The amount of physical available memory in the system (in MBs unless otherwise specified). On Windows, memory resident operating system code and data is not included as available memory. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_CACHE_HIT_PCT ---------------------------------- On HP-UX, the percentage of buffer cache reads resolved from the buffer cache (rather than going to disk) during the interval. Buffer cache reads can occur as a result of a logical read (for example, file read system call), a read generated by a client, a read-ahead on behalf of a logical read or a system procedure. On HP-UX, this metric is obtained by measuring the number of buffered read calls that were satisfied by the data that was in the file system buffer cache. Reads to filesystem file buffers that are not in the buffer cache result in disk IO. Reads to raw IO and virtual memory IO (including memory mapped files), do not go through the filesystem buffer cache, and so are not relevant to this metric. On HP-UX, a low cache hit rate may indicate low efficiency of the buffer cache, either because applications have poor data locality or because the buffer cache is too small. Overly large buffer cache sizes can lead to a memory bottleneck. The buffer cache should be sized small enough so that pageouts do not occur even when the system is busy. However, in the case of VxFS, all memory-mapped IOs show up as page ins/page outs and are not a result of memory pressure. On AIX, the percentage of disk reads that were satisfied in the file system buffer cache (rather than going to disk) during the interval. On AIX, the traditional file system buffer cache is not normally used, since files are implicitly memory mapped and the access is through the virtual memory system rather than the buffer cache. However, if a file is read as a block device (e.g /dev/hdisk1), the file system buffer cache is used, making this metric meaningful in that situation. If no IO through the buffer cache occurs during the interval, this metric is 0. On the remaining Unix systems, this is the percentage of logical reads satisfied in memory (rather than going to disk) during the interval. This includes inode, indirect block and cylinder group related disk reads, plus file reads from files memory mapped by the virtual memory IO system. On Windows, this is the percentage of buffered reads satisfied in the buffer cache (rather than going to disk) during the interval. This metric is obtained by measuring the number of buffered read calls that were satisfied by the data that was in the system buffer cache. Reads that are not in the buffer cache result in disk IO. Unbuffered IO and virtual memory IO (including memory mapped files), are not counted in this metric. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_MEM_FREE_UTIL ---------------------------------- The percentage of physical memory that was free at the end of the interval. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_PAGEOUT ---------------------------------- The total number of page outs to the disk during the interval. On HP-UX, Solaris, Linux and AIX, this reflects paging activity between memory and paging space. It does not include activity between memory and file systems. On Windows, this includes paging activity for both file systems and paging space. On HP-UX, this is the same as the “page outs” value from the “vmstat -s” command. On HP-UX 11iv3 and above this includes filecache page outs also. On AIX, this is the same as the “paging space page outs” value. Remember that “vmstat -s” reports cumulative counts. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_PAGEOUT_RATE ---------------------------------- The total number of page outs to the disk per second during the interval. On HP-UX, Solaris, Linux and AIX, this reflects paging activity between memory and paging space. It does not include activity between memory and file systems. On Windows, this includes paging activity for both file systems and paging space. On HP-UX and AIX, this is the same as the “po” value from the vmstat command. On Solaris, this is the same as the sum of the “epo” and “apo” values from the “vmstat -p” command, divided by the page size in KB. On Windows, this counter also includes paging traffic on behalf of the system cache to access file data for applications and so may be high when there is no memory pressure. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_PAGE_REQUEST ---------------------------------- The number of page requests to or from the disk during the interval. On HP-UX, Solaris, and AIX, this includes pages paged to or from the paging space and not to the file system. On Windows, this includes pages paged to or from both paging space and the file system. On HP-UX, this is the same as the sun of the “page ins” and “page outs” values from the “vmstat -s” command. On AIX, this is the same as the sum of the “paging space page ins” and “paging space page outs” values. Remember that “vmstat -s” reports cumulative counts. On Windows, this counter also includes paging traffic on behalf of the system cache to access file data for applications and so may be high when there is no memory pressure. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_PAGE_REQUEST_RATE ---------------------------------- The number of page requests to or from the disk per second during the interval. On HP-UX, Solaris, and AIX, this includes pages paged to or from the paging space and not to or from the file system. On Windows, this includes pages paged to or from both paging space and the file system. On HP-UX and AIX, this is the same as the sum of the “pi” and “po” values from the vmstat command. On Solaris, this is the same as the sum of the “epi”, “epo”, “api”, and “apo” values from the “vmstat -p” command, divided by the page size in KB. Higher than normal rates can indicate either a memory or a disk bottleneck. Compare GBL_DISK_UTIL_PEAK and GBL_MEM_UTIL to determine which resource is more constrained. High rates may also indicate memory thrashing caused by a particular application or set of applications. Look for processes with high major fault rates to identify the culprits. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_PHYS ---------------------------------- The amount of physical memory in the system (in MBs unless otherwise specified). On HP-UX, banks with bad memory are not counted. Note that on some machines, the Processor Dependent Code (PDC) code uses the upper 1MB of memory and thus reports less than the actual physical memory of the system. Thus, on a system with 256MB of physical memory, this metric and dmesg(1M) might only report 267,386,880 bytes (255MB). This is all the physical memory that software on the machine can access. On Windows, this is the total memory available, which may be slightly less than the total amount of physical memory present in the system. This value is also reported in the Control Panel’s About Windows NT help topic. On Linux, this is the amount of memory given by dmesg(1M). If the value is not available in kernel ring buffer, then the sum of system memory and available memory will be reported as physical memory. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on memory (waiting for virtual memory disk accesses to complete) during the interval. This typically happens when processes or kernel threads are allocating a large amount of memory. It can also happen when processes or kernel threads access memory that has been paged out to disk (swap) because of overall memory pressure on the system. Note that large programs can block on VM disk access when they are initializing, bringing their text and data pages into memory. When this metric rises, it can be an indication of a memory bottleneck, especially if overall system memory utilization (GBL_MEM_UTIL) is near 100% and there is also swapout or page out activity. This is calculated as the accumulated time that all processes or kernel threads spent blocked on memory divided by the interval time. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_MEM_SWAP ---------------------------------- The total number of swap ins and swap outs (or deactivations and reactivations on HP-UX) during the interval. On Linux and AIX, swap metrics are equal to the corresponding page metrics. On HP-UX, process swapping was replaced by a combination of paging and deactivation. Process deactivation occurs when the system is thrashing or when the amount of free memory falls below a critical level. The swapper then marks certain processes for deactivation and removes them from the run queue. Pages within the associated memory regions are reused or paged out by the memory management vhand process in favor of pages belonging to processes that are not deactivated. Unlike traditional process swapping, deactivated memory pages may or may not be written out to the swap area, because a process could be reactivated before the paging occurs. To summarize, a process swap-out on HP-UX is a process deactivation. A swap-in is a reactivation of a deactivated process. Swap metrics that report swap-out bytes now represent bytes paged out to swap areas from deactivated regions. Because these pages are pushed out over time based on memory demands, these counts are much smaller than HP-UX 9.x counts where the entire process was written to the swap area when it was swapped- out. Likewise, swap-in bytes now represent bytes paged in as a result of reactivating a deactivated process and reading in any pages that were actually paged out to the swap area while the process was deactivated. GBL_MEM_SWAPOUT_RATE ---------------------------------- The number of swap outs (or deactivations on HP-UX) per second during the interval. On Linux and AIX, swap metrics are equal to the corresponding page metrics. On HP-UX, process swapping was replaced by a combination of paging and deactivation. Process deactivation occurs when the system is thrashing or when the amount of free memory falls below a critical level. The swapper then marks certain processes for deactivation and removes them from the run queue. Pages within the associated memory regions are reused or paged out by the memory management vhand process in favor of pages belonging to processes that are not deactivated. Unlike traditional process swapping, deactivated memory pages may or may not be written out to the swap area, because a process could be reactivated before the paging occurs. To summarize, a process swap-out on HP-UX is a process deactivation. A swap-in is a reactivation of a deactivated process. Swap metrics that report swap-out bytes now represent bytes paged out to swap areas from deactivated regions. Because these pages are pushed out over time based on memory demands, these counts are much smaller than HP-UX 9.x counts where the entire process was written to the swap area when it was swapped- out. Likewise, swap-in bytes now represent bytes paged in as a result of reactivating a deactivated process and reading in any pages that were actually paged out to the swap area while the process was deactivated. On Solaris non-global zones with Uncapped Memory scenario, this metric value is same as seen in global zone. GBL_MEM_SWAP_1_HR_RATE ---------------------------------- The number of deactivations/reactivations per hour during the interval. On HP-UX, process swapping was replaced by a combination of paging and deactivation. Process deactivation occurs when the system is thrashing or when the amount of free memory falls below a critical level. The swapper then marks certain processes for deactivation and removes them from the run queue. Pages within the associated memory regions are reused or paged out by the memory management vhand process in favor of pages belonging to processes that are not deactivated. Unlike traditional process swapping, deactivated memory pages may or may not be written out to the swap area, because a process could be reactivated before the paging occurs. To summarize, a process swap-out on HP-UX is a process deactivation. A swap-in is a reactivation of a deactivated process. Swap metrics that report swap-out bytes now represent bytes paged out to swap areas from deactivated regions. Because these pages are pushed out over time based on memory demands, these counts are much smaller than HP-UX 9.x counts where the entire process was written to the swap area when it was swapped- out. Likewise, swap-in bytes now represent bytes paged in as a result of reactivating a deactivated process and reading in any pages that were actually paged out to the swap area while the process was deactivated. GBL_MEM_SYS_AND_CACHE_UTIL ---------------------------------- The percentage of physical memory used by the system (kernel) and the buffer cache at the end of the interval. On HP-UX 11iv3, this includes file cache also. On HP-UX 11.0, this metric does not include some kinds of dynamically allocated kernel memory. This has always been reported in the GBL_MEM_USER* metrics. On HP-UX 11.11 and beyond, this metric includes some kinds of dynamically allocated kernel memory. On Solaris non-global zones, this metric is N/A. GBL_MEM_SYS_UTIL ---------------------------------- The percentage of physical memory used by the system during the interval. System memory does not include the buffer cache. On HP-UX and Linux this does not include filecache also. On HP-UX 11.0, this metric does not include some kinds of dynamically allocated kernel memory. This has always been reported in the GBL_MEM_USER* metrics. On HP-UX 11.11 and beyond, this metric includes some kinds of dynamically allocated kernel memory. On Solaris non-global zones, this metric shows value as 0. GBL_MEM_USER_UTIL ---------------------------------- The percent of physical memory allocated to user code and data at the end of the interval. This metric shows the percent of memory owned by user memory regions such as user code, heap, stack and other data areas including shared memory. This does not include memory for buffer cache. On HP-UX and Linux this does not include filecache also. On HP-UX 11.0, this metric includes some kinds of dynamically allocated kernel memory. On HP-UX 11.11 and beyond, this metric does not include some kinds of dynamically allocated kernel memory. This is now reported in the GBL_MEM_SYS* metrics. Large fluctuations in this metric can be caused by programs which allocate large amounts of memory and then either release the memory or terminate. A slow continual increase in this metric may indicate a program with a memory leak. GBL_MEM_UTIL ---------------------------------- The percentage of physical memory in use during the interval. This includes system memory (occupied by the kernel), buffer cache and user memory. On HP-UX 11iv3 and above, this includes file cache. This excludes file cache when cachemem parameter in the parm file is set to free. On HP-UX, this calculation is done using the byte values for physical memory and used memory, and is therefore more accurate than comparing the reported kilobyte values for physical memory and used memory. On Linux, the value of this metric includes file cache when the cachemem parameter in the parm file is set to user. On SUN, high values for this metric may not indicate a true memory shortage. This metric can be influenced by the VMM (Virtual Memory Management) system. This excludes ZFS ARC cache when cachemem parameter in the parm file is set to free. On AIX, this excludes file cache when cachemem parameter in the parm file is set to free. Locality Domain metrics are available on HP-UX 11iv2 and above. GBL_MEM_FREE and LDOM_MEM_FREE, as well as the memory utilization metrics derived from them, may not always fully match. GBL_MEM_FREE represents free memory in the kernel’s reservation layer while LDOM_MEM_FREE shows actual free pages. If memory has been reserved but not actually consumed from the Locality Domains, the two values won’t match. Because GBL_MEM_FREE includes pre- reserved memory, the GBL_MEM_* metrics are a better indicator of actual memory consumption in most situations. GBL_NETWORK_SUBSYSTEM_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on the network subsystem (waiting for their network activity to complete) during the interval. This is the sum of processes or kernel threads in the LAN, NFS, and RPC wait states. This does not include processes or kernel threads blocked on SOCKT (that is, sockets) waits, as some processes or kernel threads sit idle in SOCKT waits for long periods. This is calculated as the accumulated time that all processes or kernel threads spent blocked on (LAN + NFS + RPC) divided by the interval time. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_NET_COLLISION_1_MIN_RATE ---------------------------------- The number of collisions per minute on all network interfaces during the interval. This metric does not include deferred packets. This does not include data for loopback interface. Collisions occur on any busy network, but abnormal collision rates could indicate a hardware or software problem. AIX does not support the collision count for the ethernet interface. The collision count is supported for the token ring (tr) and loopback (lo) interfaces. For more information, please refer to the netstat(1) man page. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_COLLISION_PCT ---------------------------------- The percentage of collisions to total outbound packet attempts during the interval. Outbound packet attempts include both successful packets and collisions. This does not include data for loopback interface. A rising rate of collisions versus outbound packets is an indication that the network is becoming increasingly congested. This metric does not currently include deferred packets. AIX does not support the collision count for the ethernet interface. The collision count is supported for the token ring (tr) and loopback (lo) interfaces. For more information, please refer to the netstat(1) man page. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_ERROR_1_MIN_RATE ---------------------------------- The number of errors per minute on all network interfaces during the interval. This rate should normally be zero or very small. A large error rate can indicate a hardware or software problem. This does not include data for loopback interface. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. GBL_NET_IN_ERROR_PCT ---------------------------------- The percentage of inbound network errors to total inbound packet attempts during the interval. Inbound packet attempts include both packets successfully received and those that encountered errors. This does not include data for loopback interface. A large number of errors may indicate a hardware problem on the network. The percentage of inbound errors to total packets attempted should remain low. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_IN_PACKET ---------------------------------- The number of successful packets received through all network interfaces during the interval. Successful packets are those that have been processed without errors or collisions. This does not include data for loopback interface. For HP-UX, this will be the same as the sum of the “Inbound Unicast Packets” and “Inbound Non-Unicast Packets” values from the output of the “lanadmin” utility for the network interface. Remember that “lanadmin” reports cumulative counts. As of the HP- UX 11.0 release and beyond, “netstat -i” shows network activity on the logical level (IP) only. For all other Unix systems, this is the same as the sum of the “Ipkts” column (RX-OK on Linux) from the “netstat -i” command for a network device. See also netstat(1). This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On Windows system, the packet size for NBT connections is defined as 1 Kbyte. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_IN_PACKET_RATE ---------------------------------- The number of successful packets per second received through all network interfaces during the interval. Successful packets are those that have been processed without errors or collisions. This does not include data for loopback interface. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On Windows system, the packet size for NBT connections is defined as 1 Kbyte. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_OUTQUEUE ---------------------------------- The sum of the outbound queue lengths for all network interfaces (BYNETIF_QUEUE). This metric is derived from the same source as the Outbound Queue Length shown in the lanadmin(1M) program. This does not include data for loopback interface. For most interfaces, the outbound queue is usually zero. When the value is non-zero over a period of time, the network may be experiencing a bottleneck. Determine which network interface has a non-zero queue and compare its traffic levels to normal. Also see if processes are blocking on network wait states. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. GBL_NET_OUT_ERROR_PCT ---------------------------------- The percentage of outbound network errors to total outbound packet attempts during the interval. Outbound packet attempts include both packets successfully sent and those that encountered errors. This does not include data for loopback interface. The percentage of outbound errors to total packets attempted to be transmitted should remain low. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_OUT_PACKET ---------------------------------- The number of successful packets sent through all network interfaces during the last interval. Successful packets are those that have been processed without errors or collisions. This does not include data for loopback interface. For HP-UX, this will be the same as the sum of the “Outbound Unicast Packets” and “Outbound Non-Unicast Packets” values from the output of the “lanadmin” utility for the network interface. Remember that “lanadmin” reports cumulative counts. As of the HP- UX 11.0 release and beyond, “netstat -i” shows network activity on the logical level (IP) only. For all other Unix systems, this is the same as the sum of the “Opkts” column (TX-OK on Linux) from the “netstat -i” command for a network device. See also netstat(1). This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On Windows system, the packet size for NBT connections is defined as 1 Kbyte. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_OUT_PACKET_RATE ---------------------------------- The number of successful packets per second sent through the network interfaces during the interval. Successful packets are those that have been processed without errors or collisions. This does not include data for loopback interface. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On Windows system, the packet size for NBT connections is defined as 1 Kbyte. On Solaris non-global zones, this metric shows data from the global zone. GBL_NET_PACKET_RATE ---------------------------------- The number of successful packets per second (both inbound and outbound) for all network interfaces during the interval. Successful packets are those that have been processed without errors or collisions. This does not include data for loopback interface. This metric is updated at the sampling interval, regardless of the number of IP addresses on the system. On Windows system, the packet size for NBT connections is defined as 1 Kbyte. On Solaris non-global zones, this metric shows data from the global zone. GBL_NFS_CALL ---------------------------------- The number of NFS calls the local system has made as either a NFS client or server during the interval. This includes both successful and unsuccessful calls. Unsuccessful calls are those that cannot be completed due to resource limitations or LAN packet errors. NFS calls include create, remove, rename, link, symlink, mkdir, rmdir, statfs, getattr, setattr, lookup, read, readdir, readlink, write, writecache, null and root operations. On AIX System WPARs, this metric is NA. GBL_NFS_CALL_RATE ---------------------------------- The number of NFS calls per second the system made as either a NFS client or NFS server during the interval. Each computer can operate as both a NFS server, and as an NFS client. This metric includes both successful and unsuccessful calls. Unsuccessful calls are those that cannot be completed due to resource limitations or LAN packet errors. NFS calls include create, remove, rename, link, symlink, mkdir, rmdir, statfs, getattr, setattr, lookup, read, readdir, readlink, write, writecache, null and root operations. On AIX System WPARs, this metric is NA. GBL_NUM_CPU ---------------------------------- The number of physical CPUs on the system. This includes all CPUs, either online or offline. For HP-UX and certain versions of Linux, the sar(1M) command allows you to check the status of the system CPUs. For SUN and DEC, the commands psrinfo(1M) and psradm(1M) allow you to check or change the status of the system CPUs. For AIX, this metric indicates the maximum number of CPUs the system ever had. On a logical system, this metric indicates the number of virtual CPUs configured. When hardware threads are enabled, this metric indicates the number of logical processors. On Solaris non-global zones with Uncapped CPUs, this metric shows data from the global zone. On AIX System WPARs, this metric value is identical to the value on AIX Global Environment. The Linux kernel currently doesn’t provide any metadata information for disabled CPUs. This means that there is no way to find out types, speeds, as well as hardware IDs or any other information that is used to determine the number of cores, the number of threads, the HyperThreading state, etc... If the agent (or Glance) is started while some of the CPUs are disabled, some of these metrics will be “na”, some will be based on what is visible at startup time. All information will be updated if/when additional CPUs are enabled and information about them becomes available. The configuration counts will remain at the highest discovered level (i.e. if CPUs are then disabled, the maximum number of CPUs/cores/etc... will remain at the highest observed level). It is recommended that the agent be started with all CPUs enabled. GBL_NUM_DISK ---------------------------------- The number of disks on the system. Only local disk devices are counted in this metric. On HP-UX, this is a count of the number of disks on the system that have ever had activity over the cumulative collection time. On Solaris non-global zones, this metric shows value as 0. On AIX System WPARs, this metric shows value as 0. GBL_NUM_NETWORK ---------------------------------- The number of network interfaces on the system. This includes the loopback interface. On certain platforms, this also include FDDI, Hyperfabric, ATM, Serial Software interfaces such as SLIP or PPP, and Wide Area Network interfaces (WAN) such as ISDN or X.25. The “netstat -i” command also displays the list of network interfaces on the system. GBL_NUM_USER ---------------------------------- The number of users logged in at the time of the interval sample. This is the same as the command “who | wc -l”. For Unix systems, the information for this metric comes from the utmp file which is updated by the login command. For more information, read the man page for utmp. Some applications may create users on the system without using login and updating the utmp file. These users are not reflected in this count. This metric can be a general indicator of system usage. In a networked environment, however, users may maintain inactive logins on several systems. On Windows, the information for this metric comes from the Server Sessions counter in the Performance Libraries Server object. It is a count of the number of users using this machine as a file server. GBL_OSKERNELTYPE_INT ---------------------------------- This indicates the word size of the current kernel on the system. Some hardware can load the 64-bit kernel or the 32-bit kernel. GBL_OSNAME ---------------------------------- A string representing the name of the operating system. On Unix systems, this is the same as the output from the “uname -s” command. GBL_OSRELEASE ---------------------------------- The current release of the operating system. On most Unix systems, this is same as the output from the “uname - r” command. On AIX, this is the actual patch level of the operating system. This is similar to what is returned by the command “lslpp -l bos.rte” as the most recent level of the COMMITTED Base OS Runtime. For example, “5.2.0”. GBL_OSVERSION ---------------------------------- A string representing the version of the operating system. This is the same as the output from the “uname -v” command. This string is limited to 20 characters, and as a result, the complete version name might be truncated. On Windows, this is a string representing the service pack installed on the operating system. GBL_OTHER_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on other (unknown) activities during the interval. This includes processes or kernel threads that were started and subsequently suspended before the midaemon was started and have not been resumed, or the block state is unknown. This is calculated as the accumulated time that all processes or kernel threads spent blocked on OTHER divided by the interval time. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_PRI_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on PRI (waiting for their priority to become high enough to get the CPU) during the interval. To determine if the CPU is a bottleneck, compare this metric with GBL_CPU_TOTAL_UTIL. If GBL_CPU_TOTAL_UTIL is near 100 percent and GBL_PRI_QUEUE is greater than three, there is a high probability of a CPU bottleneck. This is calculated as the accumulated time that all processes or kernel threads spent blocked on PRI divided by the interval time. HP-UX RUN/PRI/CPU Queue differences for multi-cpu systems: For example, let’s assume we’re using a system with eight processors. We start eight CPU intensive threads that consume almost all of the CPU resources. The approximate values shown for the CPU related queue metrics would be: GBL_RUN_QUEUE = 1.0 GBL_PRI_QUEUE = 0.1 GBL_CPU_QUEUE = 1.0 Assume we start an additional eight CPU intensive threads. The approximate values now shown are: GBL_RUN_QUEUE = 2.0 GBL_PRI_QUEUE = 8.0 GBL_CPU_QUEUE = 16.0 At this point, we have sixteen CPU intensive threads running on the eight processors. Keeping the definitions of the three queue metrics in mind, the run queue is 2 (that is, 16 / 8); the pri queue is 8 (only half of the threads can be active at any given time); and the cpu queue is 16 (half of the threads waiting in the cpu queue that are ready to run, plus one for each active thread). This illustrates that the run queue is the average of number of threads waiting in the runqueue for all processors; the pri queue is the number of threads that are blocked on “PRI” (priority); and the cpu queue is the number of threads in the cpu queue that are ready to run, including the threads using the CPU. Note that if the value for GBL_PRI_QUEUE greatly exceeds the value for GBL_RUN_QUEUE, this may be a side-effect of the measurement interface having lost trace data. In this case, check the value of the GBL_LOST_MI_TRACE_BUFFERS metric. If there has been buffer loss, you can correct the value of GBL_PRI_QUEUE by restarting the midaemon and the performance tools. You can use the /opt/perf/bin/midaemon -T command to force immediate shutdown of the measurement interface. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_PROC_RUN_TIME ---------------------------------- The average run time, in seconds, for processes that terminated during the interval. GBL_PROC_SAMPLE ---------------------------------- The number of process data samples that have been averaged into global metrics (such as GBL_ACTIVE_PROC) that are based on process samples. GBL_QUEUE_HISTOGRAM ---------------------------------- A bar chart of the processes waiting in queues. Shows breakout of the components of processes waiting in queues. The sum of processes waiting = GBL_RUN_QUEUE + GBL_DISK_SUBSYSTEM_QUEUE + GBL_MEM_QUEUE + GBL_NETWORK_SUBSYSTEM_QUEUE ASCII and binary files contain a line of ASCII characters that make up one row of a printed histogram. This can be a quick way to get a graphical view of processes waiting on a character mode terminal display. GBL_RUN_QUEUE ---------------------------------- On UNIX systems except Linux, this is the average number of threads waiting in the runqueue over the interval. The average is computed against the number of times the run queue is occupied instead of time. The average is updated by the kernel at a fine grain interval, only when the run queue is occupied. It is not averaged against the interval and can therefore be misleading for long intervals when the run queue is empty most or part of the time. This value matches runq-sz reported by the “sar -q” command. The GBL_LOADAVG* metrics are better indicators of run queue pressure. On Linux and Windows, this is instantaneous value obtained at the time of logging. On Linux, it shows the number of threads waiting in the runqueue. On Windows, it shows the Processor Queue Length. On Unix systems, GBL_RUN_QUEUE will typically be a small number. Larger than normal values for this metric indicate CPU contention among threads. This CPU bottleneck is also normally indicated by 100 percent GBL_CPU_TOTAL_UTIL. It may be OK to have GBL_CPU_TOTAL_UTIL be 100 percent if no other threads are waiting for the CPU. However, if GBL_CPU_TOTAL_UTIL is 100 percent and GBL_RUN_QUEUE is greater than the number of processors, it indicates a CPU bottleneck. On Windows, the Processor Queue reflects a count of process threads which are ready to execute. A thread is ready to execute (in the Ready state) when the only resource it is waiting on is the processor. The Windows operating system itself has many system threads which intermittently use small amounts of processor time. Several low priority threads intermittently wake up and execute for very short intervals. Depending on when the collection process samples this queue, there may be none or several of these low-priority threads trying to execute. Therefore, even on an otherwise quiescent system, the Processor Queue Length can be high. High values for this metric during intervals where the overall CPU utilization (gbl_cpu_total_util) is low do not indicate a performance bottleneck. Relatively high values for this metric during intervals where the overall CPU utilization is near 100% can indicate a CPU performance bottleneck. HP-UX RUN/PRI/CPU Queue differences for multi-cpu systems: For example, let’s assume we’re using a system with eight processors. We start eight CPU intensive threads that consume almost all of the CPU resources. The approximate values shown for the CPU related queue metrics would be: GBL_RUN_QUEUE = 1.0 GBL_PRI_QUEUE = 0.1 GBL_CPU_QUEUE = 1.0 Assume we start an additional eight CPU intensive threads. The approximate values now shown are: GBL_RUN_QUEUE = 2.0 GBL_PRI_QUEUE = 8.0 GBL_CPU_QUEUE = 16.0 At this point, we have sixteen CPU intensive threads running on the eight processors. Keeping the definitions of the three queue metrics in mind, the run queue is 2 (that is, 16 / 8); the pri queue is 8 (only half of the threads can be active at any given time); and the cpu queue is 16 (half of the threads waiting in the cpu queue that are ready to run, plus one for each active thread). This illustrates that the run queue is the average of number of threads waiting in the runqueue for all processors; the pri queue is the number of threads that are blocked on “PRI” (priority); and the cpu queue is the number of threads in the cpu queue that are ready to run, including the threads using the CPU. On Solaris non-global zones, this metric shows data from the global zone. GBL_SLEEP_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on SLEEP (waiting to awaken from sleep system calls) during the interval. A process or kernel thread enters the SLEEP state by putting itself to sleep using system calls such as sleep, wait, pause, sigpause, sigsuspend, poll and select. This is calculated as the accumulated time that all processes or kernel threads spent blocked on SLEEP divided by the interval time. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_STARTED_PROC ---------------------------------- The number of processes that started during the interval. GBL_SWAP_SPACE_AVAIL_KB ---------------------------------- The total amount of potential swap space, in KB. On HP-UX, this is the sum of the device swap areas enabled by the swapon command, the allocated size of any file system swap areas, and the allocated size of pseudo swap in memory if enabled. Note that this is potential swap space. Since swap is allocated in fixed (SWCHUNK) sizes, not all of this space may actually be usable. For example, on a 61MB disk using 2 MB swap size allocations, 1 MB remains unusable and is considered wasted space. On HP-UX, this is the same as (AVAIL: total) as reported by the “swapinfo -t” command. On SUN, this is the total amount of swap space available from the physical backing store devices (disks) plus the amount currently available from main memory. This is the same as (used + available)/1024, reported by the “swap -s” command. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_SWAP_SPACE_UTIL ---------------------------------- The percent of available swap space that was being used by running processes in the interval. On Windows, this is the percentage of virtual memory, which is available to user processes, that is in use at the end of the interval. It is not an average over the entire interval. It reflects the ratio of committed memory to the current commit limit. The limit may be increased by the operating system if the paging file is extended. This is the same as (Committed Bytes / Commit Limit) * 100 when comparing the results to Performance Monitor. On HP-UX, swap space must be reserved (but not allocated) before virtual memory can be created. If all of available swap is reserved, then no new processes or virtual memory can be created. Swap space locations are actually assigned (used) when a page is actually written to disk or locked in memory (pseudo swap in memory). This is the same as (PCT USED: total) as reported by the “swapinfo -mt” command. On Unix systems, this metric is a measure of capacity rather than performance. As this metric nears 100 percent, processes are not able to allocate any more memory and new processes may not be able to run. Very low swap utilization values may indicate that too much area has been allocated to swap, and better use of disk space could be made by reallocating some swap partitions to be user filesystems. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. On Solaris non-global zones, this metric is N/A. On AIX System WPARs, this metric is NA. GBL_SYSCALL_RATE ---------------------------------- The average number of system calls per second during the interval. High system call rates are normal on busy systems, especially with IO intensive applications. Abnormally high system call rates may indicate problems such as a “hung” terminal that is stuck in a loop generating read system calls. On HP-UX, system call rates affect the overhead of the midaemon. Due to the system call instrumentation on HP-UX, the fork and vfork system calls are double counted. In the case of fork and vfork, one process starts the system call, but two processes exit. HP-UX lightweight system calls, such as umask, do not show up in the Glance System Calls display, but will get added to the global system call rates. If a process is being traced (debugged) using standard debugging tools (such as adb or xdb), all system calls used by that process will show up in the System Calls display while being traced. On HP-UX, compare this metric to GBL_DISK_LOGL_IO_RATE to see if high system callrates correspond to high disk IO. GBL_CPU_SYSCALL_UTIL shows the CPU utilization due to processing system calls. GBL_SYSTEM_ID ---------------------------------- The network node hostname of the system. This is the same as the output from the “uname -n” command. On Windows, the name obtained from GetComputerName. GBL_SYSTEM_UPTIME_HOURS ---------------------------------- The time, in hours, since the last system reboot. GBL_TERM_FIRST_RESP_DIST_1 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_10 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_2 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_3 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_4 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_5 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_6 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_7 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_8 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_DIST_9 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with first response times falling into the 10 ranges collected. GBL_TERM_FIRST_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the first response times for the interval. GBL_TERM_FIRST_RESP_RANGE_1 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_2 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_3 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_4 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_5 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_6 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_7 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_8 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_FIRST_RESP_RANGE_9 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 first response time bins in seconds. This is for reporting of histograms showing the distribution of first response times. GBL_TERM_IO_QUEUE ---------------------------------- The average number of processes or kernel threads blocked on terminal IO (waiting for their terminal IO to complete) during the interval. This is calculated as the accumulated time that all processes or kernel threads spent blocked on TERM (that is, terminal IO) divided by the interval time. The Global QUEUE metrics, which are based on block states, represent the average number of process or kernel thread counts, not actual queues. The Global WAIT PCT metrics, which are also based on block states, represent the percentage of all processes or kernel threads that were alive on the system. No direct comparison is reasonable with the Application WAIT PCT metrics since they represent percentages within the context of a specific application and cannot be summed or compared with global values easily. In addition, the sum of each Application WAIT PCT for all applications will not equal 100% since these values will vary greatly depending on the number of processes or kernel threads in each application. For example, the GBL_DISK_SUBSYSTEM_QUEUE values can be low, while the APP_DISK_SUBSYSTEM_WAIT_PCT values can be high. In this case, there are many processes on the system, but there are only a very small number of processes in the specific application that is being examined and there is a high percentage of those few processes that are blocked on the disk I/O subsystem. GBL_TERM_RESP_DIST_1 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_10 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_2 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_3 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_4 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_5 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_6 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_7 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_8 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_DIST_9 ---------------------------------- An array of 10 numbers showing the distribution of transactions done during the interval with response times falling into the 10 ranges collected. GBL_TERM_RESP_RANGE_(1-10) returns the limits for the 10 ranges. This presents a histogram distribution of the response times for the interval. GBL_TERM_RESP_RANGE_1 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_2 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_3 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_4 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_5 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_6 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_7 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_8 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_TERM_RESP_RANGE_9 ---------------------------------- An array of 10 numbers showing the upper limit of ranges for the 10 response time bins in seconds. This is for reporting of histograms showing the distribution of response times. GBL_THRESHOLD_CPU ---------------------------------- The percent of CPU that a process must use to become interesting during an interval. The default for this threshold is “5.0”, which means a process must have a value of at least 5.0% for PROC_CPU_TOTAL_UTIL to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_THRESHOLD_DISK ---------------------------------- On HP-UX, this is the rate (IOs/sec) of physical disk IOs that a process must generate to become interesting during an interval. On Linux, this is the KB rate of physical disk IOs that the system must generate to become interesting during an interval. On the other Unix systems, this is the rate of either block disk IOs or major faults that a process must generate to become interesting during an interval. The default values and corresponding metric for this threshold are noted below. In order to exceed this threshold, the metric noted must match or exceed the value shown. HP-UX 5.0 for PROC_DISK_PHYS_IO_RATE for the given process SUN 5.0 for PROC_DISK_BLOCK_IO_RATE for the given process AIX 5.0 for PROC_DISK_BLOCK_IO_RATE for the given process OSF1 2.0 for PROC_IO_BYTE_RATE for the given process Linux 15.0 for GBL_DISK_PHYS_BYTE_RATE All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_THRESHOLD_FIRSTRESP ---------------------------------- The number of seconds that the average transaction time to first response must exceed for a process to become interesting during an interval. The default for this threshold is “ 5.0” which means a process must have a PROC_TERM_FIRST_RESP value of 5.0 seconds or greater to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_THRESHOLD_NOKILLED ---------------------------------- This is a flag specifying that terminating processes are not interesting. The flag is set by the THRESHOLD NOKILLED statement in the parm file. If this flag is set, then the process will be logged only if it exceeds at least one of the thresholds. The default (blank) is for the flag to be turned off, which means a terminating process will be logged in the interval it exits even if it did not exceed any thresholds during that interval. This is so that the death of a process is recorded even if it does not exceed any of the thresholds. On HP-UX, an exception to this is short-lived processes that are alive for less than one second. By default, short-lived processes are not considered interesting. However, there is a flag (THRESHOLD_SHORTLIVED) to turn on the logging of short-lived processes. GBL_THRESHOLD_NONEW ---------------------------------- This is a flag specifying that newly created processes are not interesting. The flag is set by the THRESHOLD NONEW statement in the parm file. If this flag is set, then the process will be logged only if it exceeds at least one of the thresholds. The default (blank) is for the flag to be turned off, which means a new process will be logged in the interval it was created even if it did not exceed any thresholds during that interval. This is so that the existence of a process is recorded even if it does not exceed any of the thresholds. On HP-UX, an exception to this is short-lived processes that are alive for less than one second. By default, short-lived processes are not considered interesting. However, there is a flag (THRESHOLD_SHORTLIVED) to turn on the logging of short-lived processes. GBL_THRESHOLD_PROCMEM ---------------------------------- The process memory threshold specified in the parm file. GBL_THRESHOLD_RESPONSE ---------------------------------- The number of seconds that the average transaction time to prompt must exceed for a process to become interesting during an interval. The default for this threshold is “ 30.0” which means a process must have a PROC_TERM_RESP value of 30.0 seconds or greater to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_THRESHOLD_SHORTLIVED ---------------------------------- This is a flag to specify that short-lived processes are to be considered interesting. A short-lived process is one that runs for less than one second. By default, a short-lived process is not logged unless it exceeded any threshold during its short life. This is because on UNIX systems there are many of these short- lived processes that do not consume much of the system resources. Many are simply processes created to fork another process. Logging them can become quite expensive. The THRESHOLD SHORTLIVED statement in the parm file can be used to turn logging of short- lived processes on. GBL_THRESHOLD_TRANS ---------------------------------- The number of terminal transactions that a process must complete to become interesting during an interval. The default for this threshold is “ 100” which means a process must have a value of at least 100 for PROC_TERM_TRAN to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. This metric is available on HP-UX 10.20. GBL_THRESHOLD_WAIT_CPU ---------------------------------- The percent of time that a process must spend waiting on the CPU to become interesting during an interval. The default for this threshold is “ 100.0” which means a process must have a value of at least 100.0% for PROC_PRI_WAIT_PCT to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_THRESHOLD_WAIT_DISK ---------------------------------- The percent of time that a process must spend waiting on the disk subsystem to become interesting during an interval. The disk subsystem includes disk transfers, memory cache, inode access, or CD-ROM disk transfers. The default for this threshold is “ 100.0” which means a process must have a value of at least 100.0% for PROC_DISK_SUBSYSTEM_WAIT_PCT to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_THRESHOLD_WAIT_IMPEDE ---------------------------------- The percent of time that a process must spend waiting on a semaphore to become interesting during an interval. The default for this threshold is “ 100.0” which means a process must have a value of at least 100.0% for PROC_SEM_WAIT_PCT to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_THRESHOLD_WAIT_MEMORY ---------------------------------- The percent of time that a process must spend waiting on access to main memory to become interesting during an interval. The default for this threshold is “ 100.0” which means a process must have a value of at least 100.0% for PROC_MEM_WAIT_PCT to exceed this threshold. All threshold values are supplied by the parm file. A process must exceed at least one threshold value in any given interval before it will be considered interesting and be logged. GBL_TT_OVERFLOW_COUNT ---------------------------------- The number of new transactions that could not be measured because the Measurement Processing Daemon’s (midaemon) Measurement Performance Database is full. If this happens, the default Measurement Performance Database size is not large enough to hold all of the registered transactions on this system. This can be remedied by stopping and restarting the midaemon process using the -smdvss option to specify a larger Measurement Performance Database size. The current Measurement Performance Database size can be checked using the midaemon -sizes option. INTERVAL ---------------------------------- The number of seconds in the measurement interval. For the process data class, this is the number of seconds the process was alive during the interval. LV_DIRNAME ---------------------------------- The path name of this logical volume or volume/disk group. On HP-UX 11i and beyond, data is available from VERITAS Volume Manager (VxVM). LVM (Logical Volume Manager) uses the terminology “volume group” to describe a set of related volumes. VERITAS Volume Manager uses the terminology “disk group” to describe a collection of VM disks. For additional information on VERITAS Volume Manager, see vxintro(1M). For LVM logical volumes, this is the name used as a parameter to the lvdisplay(1M) command. For volume groups, this is the name used as a parameter to the vgdisplay(1M) command. The entry referred to as the “/dev/vgXX/group” entry shows the internal resources used by the LVM software to manage the logical volumes. LV_GROUP_NAME ---------------------------------- On HP-UX, this is the name of this volume/disk group associated with a logical volume. On SUN and AIX, this is the name of this volume group associated with a logical volume. On SUN, this metric is applicable only for the Veritas LVM. On HP-UX 11i and beyond, data is available from VERITAS Volume Manager (VxVM). LVM (Logical Volume Manager) uses the terminology “volume group” to describe a set of related volumes. VERITAS Volume Manager uses the terminology “disk group” to describe a collection of VM disks. For additional information on VERITAS Volume Manager, see vxintro(1M). LV_LOGL_READ ---------------------------------- The number of logical reads for the current logical volume during the interval. Note: This metric is not available on HP-UX. LV_LOGL_WRITE ---------------------------------- The number of logical writes to the current logical volume during the interval. Note: This metric is not available on HP-UX. LV_READ_BYTE_RATE ---------------------------------- The number of physical KBs per second read from this logical volume during the interval. Note that bytes read from the buffer cache are not included in this calculation. LV_READ_RATE ---------------------------------- The number of physical reads per second for this logical volume during the interval. This may not correspond to the physical read rate from a particular disk drive since a logical volume may be composed of many disk drives or it may be a subset of a disk drive. An individual physical read from one logical volume may span multiple individual disk drives. Since this is a physical read rate, there may not be any correspondence to the logical read rate since many small reads are satisfied in the buffer cache, and large logical read requests must be broken up into physical read requests. LV_SPACE_UTIL ---------------------------------- Percentage of the logical volume file system space in use during the interval. A value of “na” is displayed for volume groups and logical volumes which have no mounted filesystem. LV_WRITE_BYTE_RATE ---------------------------------- The number of KBs per second written to this logical volume during the interval. LV_WRITE_RATE ---------------------------------- The number of physical writes per second to this logical volume during the interval. This may not correspond to the physical write rate to a particular disk drive since a logical volume may be composed of many disk drives or it may be a subset of a disk drive. Since this is a physical write rate, there may not be any correspondence to the logical write rate since many small writes are combined in the buffer cache, and many large logical writes must be broken up. PROC_APP_ID ---------------------------------- The ID number of the application to which the process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) belonged during the interval. Application “other” always has an ID of 1. There can be up to 999 user-defined applications, which are defined in the parm file. PROC_CPU_CSWITCH_TIME ---------------------------------- The time, in seconds, that the process or kernel thread spent in context switching during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_CSWITCH_UTIL ---------------------------------- The percentage of time spent in context switching the current process or kernel thread during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_INTERRUPT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread spent processing interrupts during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_INTERRUPT_UTIL ---------------------------------- The percentage of time that this process or kernel thread was in interrupt mode during the last interval. Interrupt mode means that interrupts were being handled while the process or kernel thread was loaded and running on the CPU. The interrupts may have been generated by any process, not just the running process, but they were handled while the process or kernel thread was running and may have had an impact on the performance of this process or kernel thread. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_NICE_TIME ---------------------------------- The time, in seconds, that this niced process or kernel thread was using the CPU in user mode during the interval. On HP-UX, the NICE metrics include positive nice value CPU time only. Negative nice value CPU is broken out into NNICE (negative nice) metrics. Positive nice values range from 20 to 39. Negative nice values range from 0 to 19. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_NICE_UTIL ---------------------------------- The percentage of time that this niced process or kernel thread was in user mode during the interval. On HP-UX, the NICE metrics include positive nice value CPU time only. Negative nice value CPU is broken out into NNICE (negative nice) metrics. Positive nice values range from 20 to 39. Negative nice values range from 0 to 19. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi- processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_NORMAL_TIME ---------------------------------- The time, in seconds, that the selected process or kernel thread was in user mode at normal priority during the interval. Normal priority user mode CPU excludes CPU used at real-time and nice priorities. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_NORMAL_UTIL ---------------------------------- The percentage of time that this process or kernel thread was in user mode at a normal priority during the interval. “At a normal priority” means the neither rtprio or nice had been used to alter the priority of the process or kernel thread during the interval. Normal priority user mode CPU excludes CPU used at real-time and nice priorities. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_REALTIME_TIME ---------------------------------- The time, in seconds, that the selected process or kernel thread was in user mode at a realtime priority during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_REALTIME_UTIL ---------------------------------- The percentage of time that this process or kernel thread was at a realtime priority during the interval. The realtime CPU is separated out to allow users to see the effect of using the realtime facilities to alter priority. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_SYSCALL_TIME ---------------------------------- The time, in seconds, that this process or kernel thread spent executing system calls in system mode, excluding interrupt or context processing, during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_SYSCALL_UTIL ---------------------------------- The percentage of the total CPU time this process or kernel thread spent in system mode (excluding interrupt, context switch, trap, or vfault CPU) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_SYS_MODE_TIME ---------------------------------- The CPU time in system mode in the context of the process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) during the interval. A process operates in either system mode (also called kernel mode on Unix or privileged mode on Windows) or user mode. When a process requests services from the operating system with a system call, it switches into the machine’s privileged protection mode and runs in system mode. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_SYS_MODE_UTIL ---------------------------------- The percentage of time that the CPU was in system mode in the context of the process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) during the interval. A process operates in either system mode (also called kernel mode on Unix or privileged mode on Windows) or user mode. When a process requests services from the operating system with a system call, it switches into the machine’s privileged protection mode and runs in system mode. Unlike the global and application CPU metrics, process CPU is not averaged over the number of processors on systems with multiple CPUs. Single-threaded processes can use only one CPU at a time and never exceed 100% CPU utilization. High system mode CPU utilizations are normal for IO intensive programs. Abnormally high system CPU utilization can indicate that a hardware problem is causing a high interrupt rate. It can also indicate programs that are not using system calls efficiently. A classic “hung shell” shows up with very high system mode CPU because it gets stuck in a loop doing terminal reads (a system call) to a device that never responds. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_TOTAL_TIME ---------------------------------- The total CPU time, in seconds, consumed by a process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) during the interval. Unlike the global and application CPU metrics, process CPU is not averaged over the number of processors on systems with multiple CPUs. Single-threaded processes can use only one CPU at a time and never exceed 100% CPU utilization. On HP-UX, the total CPU time is the sum of the CPU time components for a process or kernel thread, including system, user, context switch, interrupts processing, realtime, and nice utilization values. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_TOTAL_TIME_CUM ---------------------------------- The total CPU time consumed by a process (or kernel thread, if HP- UX/Linux Kernel 2.6 and above) over the cumulative collection time. CPU time is in seconds unless otherwise specified. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. This is calculated as PROC_CPU_TOTAL_TIME_CUM = PROC_CPU_SYS_MODE_TIME_CUM + PROC_CPU_USER_MODE_TIME_CUM On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_TOTAL_UTIL ---------------------------------- The total CPU time consumed by a process (or kernel thread, if HP- UX/Linux Kernel 2.6 and above) as a percentage of the total CPU time available during the interval. Unlike the global and application CPU metrics, process CPU is not averaged over the number of processors on systems with multiple CPUs. Single-threaded processes can use only one CPU at a time and never exceed 100% CPU utilization. On HP-UX, the total CPU utilization is the sum of the CPU utilization components for a process or kernel thread, including system, user, context switch, interrupts processing, realtime, and nice utilization values. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_TOTAL_UTIL_CUM ---------------------------------- The total CPU time consumed by a process (or kernel thread, if HP- UX/Linux Kernel 2.6 and above) as a percentage of the total CPU time available over the cumulative collection time. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. Unlike the global and application CPU metrics, process CPU is not averaged over the number of processors on systems with multiple CPUs. Single-threaded processes can use only one CPU at a time and never exceed 100% CPU utilization. On HP-UX, the total CPU utilization is the sum of the CPU utilization components for a process or kernel thread, including system, user, context switch, interrupts processing, realtime, and nice utilization values. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_USER_MODE_TIME ---------------------------------- The time, in seconds, the process (or kernel threads, if HP- UX/Linux Kernel 2.6 and above) was using the CPU in user mode during the interval. User CPU is the time spent in user mode at a normal priority, at real-time priority (on HP-UX, AIX, and Windows systems), and at a nice priority. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_CPU_USER_MODE_UTIL ---------------------------------- The percentage of time the process (or kernel thread, if HP- UX/Linux Kernel 2.6 and above) was using the CPU in user mode during the interval. User CPU is the time spent in user mode at a normal priority, at real-time priority (on HP-UX, AIX, and Windows systems), and at a nice priority. Unlike the global and application CPU metrics, process CPU is not averaged over the number of processors on systems with multiple CPUs. Single-threaded processes can use only one CPU at a time and never exceed 100% CPU utilization. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On multi-processor HP-UX systems, processes which have component kernel threads executing simultaneously on different processors could have resource utilization sums over 100%. The maximum percentage is 100% times the number of CPUs online. On platforms other than HPUX, If the ignore_mt flag is set(true) in parm file, this metric will report values normalized against the number of active cores in the system. If the ignore_mt flag is not set(false) in parm file, this metric will report values normalized against the number of threads in the system. This flag will be a no-op if Multithreading is turned off. On HPUX, CPU utilization normalization is controlled by the “- ignore_mt” option of the midaemon(1m). To change normalization from core-based to logical-cpu-based, or vice-versa, all performance components (scopeux, glance, perfd) must be shut down and the midaemon restarted in the desired mode. To start the midaemon with “-ignore_mt” by default, this option should be added in the /etc/rc.config.d/ovpa control file. Refer to the documentation regarding ovpa startup. Note that, on HPUX, unlike other platforms, specifying core-based normalization affects CPU, application, process and thread metrics. PROC_DISK_FS_IO ---------------------------------- The number of file system disk IOs for a process during the interval. These are physical reads generated by user file system access and do not include virtual memory reads, system reads (inode access, or reads relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical IOs in this category (they appear under virtual memory IOs). PROC_DISK_FS_IO_RATE ---------------------------------- The number of file system disk IOs per second for a process during the interval. These are physical reads generated by user file system access and do not include virtual memory reads, system reads (inode access), or reads relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical IOs in this category (they appear under virtual memory IOs). PROC_DISK_FS_READ ---------------------------------- Number of file system physical disk reads made by a process or kernel thread during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical reads generated by user file system access and do not include virtual memory reads, system reads (inode access), or reads relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical reads in this category. They appear under virtual memory reads. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_FS_READ_RATE ---------------------------------- The number of file system physical disk reads made by a process or kernel thread during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical reads generated by user file system access and do not include virtual memory reads, system reads (inode access), or reads relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical reads in this category. They appear under virtual memory reads. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_FS_WRITE ---------------------------------- Number of file system physical disk writes made by a process or kernel thread during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical writes generated by user file system access and do not include virtual memory writes, system writes (inode updates), or writes relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical writes in this category. They appear under virtual memory writes. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_FS_WRITE_RATE ---------------------------------- The number of file system physical disk writes made by a process or kernel thread during the interval. Only local disks are counted in this measurement. NFS devices are excluded. These are physical writes generated by user file system access and do not include virtual memory writes, system writes (inode updates), or writes relating to raw disk access. An exception is user files accessed via the mmap(2) call, which does not show their physical writes in this category. They appear under virtual memory writes. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_LOGL_IO_CUM ---------------------------------- The number of logical IOs made by (or for) a process or kernel thread over the cumulative collection time. NFS mounted disks are not included in this list. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. On many Unix systems, logical disk IOs are measured by counting the read and write system calls that are directed to disk devices. Also counted are read and write system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, writev, send, sento, sendmsg, and ipcsend. “Disk” refers to a physical drive (that is, “spindle”), not a partition on a drive (unless the partition occupies the entire physical disk). On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_LOGL_IO_RATE_CUM ---------------------------------- The average number of logical IOs per second made by (or for) a process or kernel thread over the cumulative collection time. Only local disks are counted in this measurement. NFS devices are excluded. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. On many Unix systems, logical disk IOs are measured by counting the read and write system calls that are directed to disk devices. Also counted are read and write system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, writev, send, sento, sendmsg, and ipcsend. On many Unix systems, there are several reasons why logical IOs may not correspond with physical IOs. Logical IOs may not always result in a physical disk access, since the data may already reside in memory -- either in the buffer cache, or in virtual memory if the IO is to a memory mapped file. Several logical IOs may all map to the same physical page or block. In these two cases, logical IOs are greater than physical IOs. The reverse can also happen. A single logical write can cause a physical read to fetch the block to be updated from disk, and then cause a physical write to put it back on disk. A single logical IO can require more than one physical page or block, and these can be found on different disks. Mirrored disks further distort the relationship between logical and physical IO, since physical writes are doubled. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_LOGL_READ ---------------------------------- The number of disk logical reads made by a process or kernel thread during the interval. Calls destined for NFS mounted files are not counted. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_LOGL_READ_RATE ---------------------------------- The number of logical reads per second made by (or for) a process or kernel thread during the interval. Calls destined for NFS mounted files are not counted. On many Unix systems, logical disk IOs are measured by counting the read system calls that are directed to disk devices. Also counted are read system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. “Disk” refers to a physical drive (that is, “spindle”), not a partition on a drive (unless the partition occupies the entire physical disk). On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_LOGL_WRITE ---------------------------------- Number of disk logical writes made by a process or kernel thread during the interval. Calls destined for NFS mounted files are not counted. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_LOGL_WRITE_RATE ---------------------------------- The number of logical writes per second made by (or for) a process or kernel thread during the interval. NFS mounted disks are not included in this list. “Disk” refers to a physical drive (that is, “spindle”), not a partition on a drive (unless the partition occupies the entire physical disk). On many Unix systems, logical disk IOs are measured by counting the write system calls that are directed to disk devices. Also counted are write system calls made indirectly through other system calls, including writev, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, send, sento, sendmsg, and ipcsend. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_PHYS_IO ---------------------------------- The number of physical IOs made by (or for) a process during the interval. PROC_DISK_PHYS_IO_CUM ---------------------------------- This is the total number of physical disk transfers done by a process since it was created. It is calculated as: PROC_DISK_PHYS_IO_CUM = PROC_DISK_PHYS_IO_RATE_CUM * PROC_RUN_TIME PROC_DISK_PHYS_IO_RATE ---------------------------------- The average number of physical disk IOs per second made by the process or kernel thread during the interval. For processes which run for less than the measurement interval, this metric is normalized over the measurement interval. For example, a process ran for 1 second and did 50 IOs during its life. If the measurement interval is 5 seconds, it is reported as having done 10 IOs per second. If the measurement interval is 60 seconds, it is reported as having done 50/60 or 0.83 IOs per second. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. Linux release versions vary with regards to the amount of process-level IO statistics that are available. Some kernels instrument only disk IO, while some provide statistics for all devices together (including tty and other devices with disk IO). When it is available from your specific release of Linux, the PROC_DISK_PHYS* metrics will report pages of disk IO specifically. The PROC_IO* metrics will report the sum of all types of IO including disk IO, in Kilobytes or KB rates. These metrics will have “na” values on kernels that do not support the instrumentation. For multi-threaded processes, some Linux kernels only report IO statistics for the main thread. In that case, patches are available that will allow the process instrumentation to report the sum of all thread’s IOs, and will also enable per-thread reporting. PROC_DISK_PHYS_IO_RATE_CUM ---------------------------------- The number of physical disk IOs per second made by the selected process or kernel thread over the cumulative collection time. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. Linux release versions vary with regards to the amount of process-level IO statistics that are available. Some kernels instrument only disk IO, while some provide statistics for all devices together (including tty and other devices with disk IO). When it is available from your specific release of Linux, the PROC_DISK_PHYS* metrics will report pages of disk IO specifically. The PROC_IO* metrics will report the sum of all types of IO including disk IO, in Kilobytes or KB rates. These metrics will have “na” values on kernels that do not support the instrumentation. For multi-threaded processes, some Linux kernels only report IO statistics for the main thread. In that case, patches are available that will allow the process instrumentation to report the sum of all thread’s IOs, and will also enable per-thread reporting. PROC_DISK_SUBSYSTEM_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on the disk subsystem (waiting for its file system IOs to complete) during the interval. This includes time spent waiting in the DISK, INODE, CACHE, and CDFS wait states. It does not include processes doing raw IO to disk devices. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_DISK_SUBSYSTEM_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on the disk subsystem (waiting for its file system IOs to complete) during the interval. This includes time spent waiting in the DISK, INODE, CACHE, and CDFS wait states. It does not include processes doing raw IO to disk devices. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_DISK_SYSTEM_IO ---------------------------------- Number of file system management physical disk IOs made for a process or kernel thread during the interval. File system management IOs are the physical accesses required to obtain or update internal information about the file system structure (inode access). Accesses or updates to user data are not included in this metric. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_SYSTEM_IO_RATE ---------------------------------- The number of file system management physical disk IOs per second made for a process or kernel thread during the interval. File system management IOs are the physical accesses required to obtain or update internal information about the file system structure (inode access). Accesses or updates to user data are not included in this metric. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_VM_IO ---------------------------------- The number of virtual memory IOs made for a process or kernel thread during the interval. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_DISK_VM_IO_RATE ---------------------------------- The number of virtual memory IOs per second made for a process or kernel thread during the interval. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. PROC_GROUP_ID ---------------------------------- On most systems, this is the real group ID number of the process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above). On AIX, this is the effective group ID number of the process. On HP-UX, this is the effective group ID number of the process if not in setgid mode. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. PROC_INTEREST ---------------------------------- A string containing the reason(s) why the process or thread is of interest, based on the thresholds specified in the parm file. An ‘A’ indicates that the process or thread exceeds the process CPU threshold, computed using the actual time the process or thread was alive during the interval. A ‘C’ indicates that the process or thread exceeds the process CPU threshold, computed using the collection interval. Currently, the same CPU threshold is used for both CPU interest reasons. A ‘D’ indicates that the process or thread exceeds the process disk IO threshold. An ‘I’ indicates that the process or thread exceeds the IO threshold. An ‘M’ indicates that the process exceeds the process memory threshold. This interest reason is only meaningful for processes and therefore not shown for threads. New processes or threads are identified with an ‘N’, terminated processes or threads are identified with a ‘K’. Note that the parm file ‘nonew’, ‘nokill’ and ‘shortlived’ settings are logging only options and therefore ignored in Glance components. 4 D Disk IOs exceeded threshold 5 M Memory used exceeded threshold 6 F First- response time exceeded threshold 7 T Transaction rate exceeded threshold 8 c Wait for CPU percentage exceeded threshold 9 d Wait for disk percentage exceeded threshold 10 m Wait for memory percentage exceeded threshold 11 i Wait for impede percentage exceeded threshold 12 blank Special purpose field PROC_INTERVAL_ALIVE ---------------------------------- The number of seconds that the process (or kernel thread, if HP- UX/Linux Kernel 2.6 and above) was alive during the interval. This may be less than the time of the interval if the process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) was new or died during the interval. PROC_IO_BYTE ---------------------------------- On HP-UX, this is the total number of physical IO KBs (unless otherwise specified) that was used by this process or kernel thread, either directly or indirectly, during the interval. On all other systems, this is the total number of physical IO KBs (unless otherwise specified) that was used by this process during the interval. IOs include disk, terminal, tape and network IO. On HP-UX, indirect IOs include paging and deactivation/reactivation activity done by the kernel on behalf of the process or kernel thread. Direct IOs include disk, terminal, tape, and network IO, but exclude all NFS traffic. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On SUN, counts in the MB ranges in general can be attributed to disk accesses and counts in the KB ranges can be attributed to terminal IO. This is useful when looking for processes with heavy disk IO activity. This may vary depending on the sample interval length. Linux release versions vary with regards to the amount of process-level IO statistics that are available. Some kernels instrument only disk IO, while some provide statistics for all devices together (including tty and other devices with disk IO). When it is available from your specific release of Linux, the PROC_DISK_PHYS* metrics will report pages of disk IO specifically. The PROC_IO* metrics will report the sum of all types of IO including disk IO, in Kilobytes or KB rates. These metrics will have “na” values on kernels that do not support the instrumentation. For multi-threaded processes, some Linux kernels only report IO statistics for the main thread. In that case, patches are available that will allow the process instrumentation to report the sum of all thread’s IOs, and will also enable per-thread reporting. PROC_IO_BYTE_CUM ---------------------------------- On HP-UX, this is the total number of physical IO KBs (unless otherwise specified) that was used by this process or kernel thread, either directly or indirectly, over the cumulative collection time. On all other systems, this is the total number of physical IO KBs (unless otherwise specified) that was used by this process over the cumulative collection time. IOs include disk, terminal, tape and network IO. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. On HP-UX, indirect IOs include paging and deactivation/reactivation activity done by the kernel on behalf of the process or kernel thread. Direct IOs include disk, terminal, tape, and network IO, but exclude all NFS traffic. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. Linux release versions vary with regards to the amount of process-level IO statistics that are available. Some kernels instrument only disk IO, while some provide statistics for all devices together (including tty and other devices with disk IO). When it is available from your specific release of Linux, the PROC_DISK_PHYS* metrics will report pages of disk IO specifically. The PROC_IO* metrics will report the sum of all types of IO including disk IO, in Kilobytes or KB rates. These metrics will have “na” values on kernels that do not support the instrumentation. For multi-threaded processes, some Linux kernels only report IO statistics for the main thread. In that case, patches are available that will allow the process instrumentation to report the sum of all thread’s IOs, and will also enable per-thread reporting. PROC_IO_BYTE_RATE ---------------------------------- On HP-UX, this is the number of physical IO KBs per second that was used by this process or kernel thread, either directly or indirectly, during the interval. On all other systems, this is the number of physical IO KBs per second that was used by this process during the interval. IOs include disk, terminal, tape and network IO. On HP-UX, indirect IOs include paging and deactivation/reactivation activity done by the kernel on behalf of the process or kernel thread. Direct IOs include disk, terminal, tape, and network IO, but exclude all NFS traffic. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On SUN, counts in the MB ranges in general can be attributed to disk accesses and counts in the KB ranges can be attributed to terminal IO. This is useful when looking for processes with heavy disk IO activity. This may vary depending on the sample interval length. Certain types of disk IOs are not counted by AIX at the process level, so they are excluded from this metric. Linux release versions vary with regards to the amount of process-level IO statistics that are available. Some kernels instrument only disk IO, while some provide statistics for all devices together (including tty and other devices with disk IO). When it is available from your specific release of Linux, the PROC_DISK_PHYS* metrics will report pages of disk IO specifically. The PROC_IO* metrics will report the sum of all types of IO including disk IO, in Kilobytes or KB rates. These metrics will have “na” values on kernels that do not support the instrumentation. For multi-threaded processes, some Linux kernels only report IO statistics for the main thread. In that case, patches are available that will allow the process instrumentation to report the sum of all thread’s IOs, and will also enable per-thread reporting. PROC_IO_BYTE_RATE_CUM ---------------------------------- On HP-UX, this is the average number of physical IO KBs per second that was used by this process or kernel thread, either directly or indirectly, over the cumulative collection time. On all other systems, this is the average number of physical IO KBs per second that was used by this process over the cumulative collection time. IOs include disk, terminal, tape and network IO. The cumulative collection time is defined from the point in time when either: a) the process (or thread) was first started, or b) the performance tool was first started, or c) the cumulative counters were reset (relevant only to Glance, if available for the given platform), whichever occurred last. On HP-UX, all cumulative collection times and intervals start when the midaemon starts. On other Unix systems, non-process collection time starts from the start of the performance tool, process collection time starts from the start time of the process or measurement start time, which ever is older. Regardless of the process start time, application cumulative intervals start from the time the performance tool is started. On systems where the performance components are 32-bit or where the 64-bit model is LLP64 (Windows), all INTERVAL_CUM metrics will start reporting “o/f” (overflow) after the performance agent (or the midaemon on HPUX) has been up for 466 days and the cumulative metrics will fail to report accurate data after 497 days. On Linux, Solaris and AIX, if measurement is started after the system has been up for more than 466 days, cumulative process CPU data won’t include times accumulated prior to the performance tool’s start and a message will be logged to indicate this. On HP-UX, indirect IOs include paging and deactivation/reactivation activity done by the kernel on behalf of the process or kernel thread. Direct IOs include disk, terminal, tape, and network IO, but exclude all NFS traffic. On a threaded operating system, such as HP-UX 11.0 and beyond, process usage of a resource is calculated by summing the usage of that resource by its kernel threads. If this metric is reported for a kernel thread, the value is the resource usage by that single kernel thread. If this metric is reported for a process, the value is the sum of the resource usage by all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. On SUN, counts in the MB ranges in general can be attributed to disk accesses and counts in the KB ranges can be attributed to terminal IO. This is useful when looking for processes with heavy disk IO activity. This may vary depending on the sample interval length. Linux release versions vary with regards to the amount of process-level IO statistics that are available. Some kernels instrument only disk IO, while some provide statistics for all devices together (including tty and other devices with disk IO). When it is available from your specific release of Linux, the PROC_DISK_PHYS* metrics will report pages of disk IO specifically. The PROC_IO* metrics will report the sum of all types of IO including disk IO, in Kilobytes or KB rates. These metrics will have “na” values on kernels that do not support the instrumentation. For multi-threaded processes, some Linux kernels only report IO statistics for the main thread. In that case, patches are available that will allow the process instrumentation to report the sum of all thread’s IOs, and will also enable per-thread reporting. PROC_IPC_SUBSYSTEM_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on the InterProcess Communication (IPC) subsystems (waiting for its interprocess communication activity to complete) during the interval. This is the sum of processes or kernel threads in the IPC, MSG, SEM, PIPE, SOCKT (that is, sockets) and STRMS (that is, streams IO) wait states. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_IPC_SUBSYSTEM_WAIT_TIME ---------------------------------- The time, in seconds, the process or kernel thread was blocked on the InterProcess Communication (IPC) subsystems (waiting for its interprocess communication activity to complete) during the interval. This is the sum of processes or kernel threads in the IPC, MSG, SEM, PIPE, SOCKT (that is, sockets) and STRMS (that is, streams IO) wait states. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_LAN_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on LAN (waiting for IO over the LAN to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_LAN_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on LAN (waiting for IO over the LAN to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_MAJOR_FAULT ---------------------------------- Number of major page faults for this process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) during the interval. On HP-UX, major page faults and minor page faults are a subset of vfaults (virtual faults). Stack and heap accesses can cause vfaults, but do not result in a disk page having to be loaded into memory. PROC_MEM_RES ---------------------------------- The size (in KB) of resident memory allocated for the process(or kernel thread, if HP-UX/Linux Kernel 2.6 and above). On HP-UX, the calculation of this metric differs depending on whether this process has used any CPU time since the midaemon process was started. This metric is less accurate and does not include shared memory regions in its calculation when the process has been idle since the midaemon was started. On HP-UX, for processes that use CPU time subsequent to midaemon startup, the resident memory is calculated as RSS = sum of private region pages + (sum of shared region pages / number of references) The number of references is a count of the number of attachments to the memory region. Attachments, for shared regions, may come from several processes sharing the same memory, a single process with multiple attachments, or combinations of these. This value is only updated when a process uses CPU. Thus, under memory pressure, this value may be higher than the actual amount of resident memory for processes which are idle because their memory pages may no longer be resident or the reference count for shared segments may have changed. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. A value of “na” is displayed when this information is unobtainable. This information may not be obtainable for some system (kernel) processes. It may also not be available for processes. On AIX, this is the same as the RSS value shown by “ps v”. On Windows, this is the number of KBs in the working set of this process. The working set includes the memory pages touched recently by the threads of the process. If free memory in the system is above a threshold, then pages are left in the working set even if they are not in use. When free memory falls below a threshold, pages are trimmed from the working set, but not necessarily paged out to disk from memory. If those pages are subsequently referenced, they will be page faulted back into the working set. Therefore, the working set is a general indicator of the memory resident set size of this process, but it will vary depending on the overall status of memory on the system. Note that the size of the working set is often larger than the amount of pagefile space consumed (PROC_MEM_VIRT). PROC_MEM_VIRT ---------------------------------- The size (in KB) of virtual memory allocated for the process(or kernel thread, if HP-UX/Linux Kernel 2.6 and above). On HP-UX, this consists of the sum of the virtual set size of all private memory regions used by this process, plus this process’ share of memory regions which are shared by multiple processes. For processes that use CPU time, the value is divided by the reference count for those regions which are shared. On HP-UX, this metric is less accurate and does not reflect the reference count for shared regions for processes that were started prior to the midaemon process and have not used any CPU time since the midaemon was started. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. On all other Unix systems, this consists of private text, private data, private stack and shared memory. The reference count for shared memory is not taken into account, so the value of this metric represents the total virtual size of all regions regardless of the number of processes sharing access. Note also that lazy swap algorithms, sparse address space malloc calls, and memory-mapped file access can result in large VSS values. On systems that provide Glance memory regions detail reports, the drilldown detail per memory region is useful to understand the nature of memory allocations for the process. A value of “na” is displayed when this information is unobtainable. This information may not be obtainable for some system (kernel) processes. It may also not be available for processes. On Windows, this is the number of KBs the process has used in the paging file(s). Paging files are used to store pages of memory used by the process, such as local data, that are not contained in other files. Examples of memory pages which are contained in other files include pages storing a program’s .EXE and .DLL files. These would not be kept in pagefile space. Thus, often programs will have a memory working set size (PROC_MEM_RES) larger than the size of its pagefile space. On Linux this value is rounded to PAGESIZE. PROC_MEM_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on memory (waiting for memory resources to become available) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_MEM_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on VM (waiting for virtual memory resources to become available) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_MINOR_FAULT ---------------------------------- Number of minor page faults for this process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) during the interval. On HP-UX, major page faults and minor page faults are a subset of vfaults (virtual faults). Stack and heap accesses can cause vfaults, but do not result in a disk page having to be loaded into memory. PROC_NFS_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on NFS (waiting for network file system IO to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_NFS_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on NFS (waiting for its network file system IO to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_OTHER_IO_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on “other IO” during the interval. “Other IO” includes all IO directed at a device (connected to the local computer) which is not a terminal or LAN. Examples of “other IO” devices are local printers, tapes, instruments, and disks. Time waiting for character (raw) IO to disks is included in this measurement. Time waiting for file system buffered IO to disks will typically been seen as IO or CACHE wait. Time waiting for IO to NFS disks is reported as NFS wait. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_OTHER_IO_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on other IO during the interval. “Other IO” includes all IO directed at a device (connected to the local computer) which is not a terminal or LAN. Examples of “other IO” devices are local printers, tapes, instruments, and disks. Time waiting for character (raw) IO to disks is included in this measurement. Time waiting for file system buffered IO to disks will typically been seen as IO or CACHE wait. Time waiting for IO to NFS disks is reported as NFS wait. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_OTHER_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on other (unknown) activities during the interval. This includes processes or kernel threads that were started and subsequently suspended before the midaemon was started and have not been resumed, or the block state is unknown. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_OTHER_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on other (unknown) activities during the interval. This includes processes or kernel threads that were started and subsequently suspended before the midaemon was started and have not been resumed, or the block state is unknown. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_PARENT_PROC_ID ---------------------------------- The parent process’ PID number. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. PROC_PRI ---------------------------------- On Unix systems, this is the dispatch priority of a process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) at the end of the interval. The lower the value, the more likely the process is to be dispatched. On Windows, this is the current base priority of this process. On HP-UX, whenever the priority is changed for the selected process or kernel thread, the new value will not be reflected until the process or kernel thread is reactivated if it is currently idle (for example, SLEEPing). On HP-UX, the lower the value, the more the process or kernel thread is likely to be dispatched. Values between zero and 127 are considered to be “real-time” priorities, which the kernel does not adjust. Values above 127 are normal priorities and are modified by the kernel for load balancing. Some special priorities are used in the HP-UX kernel and subsystems for different activities. These values are described in /usr/include/sys/param.h. Priorities less than PZERO 153 are not signalable. Note that on HP-UX, many network-related programs such as inetd, biod, and rlogind run at priority 154 which is PPIPE. Just because they run at this priority does not mean they are using pipes. By examining the open files, you can determine if a process or kernel thread is using pipes. For HP-UX 10.0 and later releases, priorities between -32 and -1 can be seen for processes or kernel threads using the Posix Real- time Schedulers. When specifying a Posix priority, the value entered must be in the range from 0 through 31, which the system then remaps to a negative number in the range of -1 through -32. Refer to the rtsched man pages for more information. On a threaded operating system, such as HP-UX 11.0 and beyond, this metric represents a kernel thread characteristic. If this metric is reported for a process, the value for its last executing kernel thread is given. For example, if a process has multiple kernel threads and kernel thread one is the last to execute during the interval, the metric value for kernel thread one is assigned to the process. On AIX, values for priority range from 0 to 127. Processes running at priorities less than PZERO (40) are not signalable. On Windows, the higher the value the more likely the process or thread is to be dispatched. Values for priority range from 0 to 31. Values of 16 and above are considered to be “realtime” priorities. Threads within a process can raise and lower their own base priorities relative to the process’s base priority. PROC_PRI_WAIT_PCT ---------------------------------- The percentage of time during the interval the process or kernel thread was blocked on priority (waiting for its priority to become high enough to get the CPU). On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_PRI_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on PRI (waiting for its priority to become high enough to get the CPU) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_PRMID ---------------------------------- The PRM Group ID this process is assigned to. The PRM group configuration is kept in the PRM configuration file. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. PROC_PROC_ID ---------------------------------- The process ID number (or PID) of this process(or associated process for kernel threads, if HPUX/LInux Kernel 2.6 and above) that is used by the kernel to uniquely identify the process. Process numbers are reused, so they only identify a process for its lifetime. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. PROC_PROC_NAME ---------------------------------- The process(or kernel thread, if HP-UX/Linux Kernel 2.6 and above) program name. It is limited to 16 characters. On Unix systems, this is derived from the 1st parameter to the exec(2) system call. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. On Windows, the “System Idle Process” is not reported by Perf Agent since Idle is a process that runs to occupy the processors when they are not executing other threads. Idle has one thread per processor. PROC_RUN_TIME ---------------------------------- The elapsed time since a process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) started, in seconds. This metric is less than the interval time if the process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) was not alive during the entire first or last interval. On a threaded operating system such as HP-UX 11.0 and beyond, this metric is available for a process or kernel thread. PROC_SEM_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on semaphores (waiting on a semaphore operation to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_SEM_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on semaphores (waiting on a semaphore operation to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_SLEEP_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on SLEEP (waiting to awaken from sleep system calls) during the interval. A process or kernel thread enters the SLEEP state by putting itself to sleep using system calls such as sleep, wait, pause, sigpause, sigsuspend, poll and select. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_SLEEP_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on SLEEP (waiting to awaken from sleep system calls) during the interval. A process or kernel thread enters the SLEEP state by putting itself to sleep using system calls such as sleep, wait, pause, sigpause, sigsuspend, poll and select. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_STOP_HISTOGRAM ---------------------------------- A bar chart of the process stop reasons. Shows breakout of the percent of time processes were in different wait states. The percent of time = PROC_CPU_TOTAL_UTIL_CUM + PROC_PRI_WAIT_PCT + PROC_DISK_SUBSYSTEM_WAIT_PCT + PROC_MEM_WAIT_PCT + PROC_LAN_WAIT_PCT ASCII and binary files contain a line of ASCII characters that make up one row of a printed histogram. This can be a quick way to get a graphical view of the wait states on a character mode terminal display. PROC_STOP_REASON ---------------------------------- A text string describing what caused the process (or kernel thread, if HP-UX/Linux Kernel 2.6 and above) to stop executing. For example, if the process is waiting for a CPU while higher priority processes are executing, then its block reason is PRI. A complete list of block reasons follows: String Reason for Process Block ------------------------------------ CACHE Waiting at the buffer cache level trying to lock down a buffer cache structure, or waiting for an IO operation to or from a buffer cache to complete. File system access will block on IO more often than CACHE on HP-UX 11.x. CDFS Waiting for CD-ROM file system node structure allocation or locks while accessing a CD-ROM device through the file system. died Process terminated during the interval. DISK Waiting for an IO operation to complete at the logical device manager or disk driver level. Waits from raw disk IO and diagnostic requests can be seen here. Buffered IO requests can also block on DISK, but will more often be seen waiting on “IO”. CDFS access will block on “CDFS”. Virtual memory activity will block on “VM”. GRAPH Waiting for a graphics card or framebuf semaphore operation. INODE Waiting while accessing an inode structure. This includes inode gets and waiting due to inode locks. IO Waiting for IO to local disks, printers, tapes, or instruments to complete (above the driver, but below the buffer cache). Both file system and raw disk access can block in this state. CDFS access will block on “CDFS”. Virtual memory activity will block on “VM”. IPC Waiting for a process or kernel thread event (that is, waiting for a child to receive a signal). This includes both inter and intra process or kernel thread operations, such as IPC locks, kernel thread mutexes, and database IPC operations. System V message queue operations will block on “MESG”, while semaphore operations will block on “SEM”. JOBCL Waiting for tracing resume, debug resume, or job control start. A background process incurs this block when attempting to write to a terminal set with “stty tostop”. On HP-UX 11i, scheduler activation threads (user threads) will show this block. LAN Waiting for a network IO completion. This includes waiting on the LAN hardware and low level LAN device driver. It does not include waiting on the higher level network software such as the streams based transport or NFS, which has its own stop state. MESG Waiting for a System V message queue operation such as msgrcv or msgsnd. new Process was created (via the fork/vfork system calls) during the interval. NFS Waiting for a Networked File System request to complete. This includes both NFS V2 and V3 requests. This does not include stops where kernel threads or deamons are waiting for a NFS event or request (such as biod or nfsd). These will block on SLEEP to show they are waiting for some activity. NONE Zombie process - waiting to die. OTHER The process was started before the midaemon was started and has not been resumed, or the block state is unknown. PIPE Waiting for operations involving pipes. This includes opening, closing, reading, and writing using pipes. Named pipes will block on PIPE. PRI Waiting because a higher priority process is running, or waiting for a spinlock or alpha semaphore. RPC Waiting for remote procedure call operations to complete. This includes both NFS and DCE RPC requests. SEM Waiting for a System V semaphore operation (such as semop, semget, or semctl) or waiting for a memory mapped file semaphore operation (such as msem_init or msem_lock). SLEEP Waiting because the process put itself to sleep using system calls such as sleep, wait, pause, sigpause, poll, sigsuspend and select. This is the standard stop reason for idle system daemons. SOCKT Waiting for an operation to complete while accessing a device through a socket. This is used primarily in networking code and includes all protocols using sockets (X25, UDP, TCP, and so on). STRMS Waiting for an operation to complete while accessing a “streams” device. This is the normal stop reason for kernel threads and daemons waiting for a streams event. This includes the network transport and pseudo terminal IO requests. For example, waiting for a read on a streams device or waiting for an internal streams synchronization. SYSTM Waiting for access to a system resource or lock. These resources include data structures from the LVM, VFS, UFS, JFS, and Disk Quota subsystems. “SYSTM” is the “catch-all” wait state for blocks on system resources that are not common enough or long enough to warrant their own stop state. TERM Waiting for a non-streams terminal transfer (tty or pty). VM Waiting for a virtual memory operation to complete, or waiting for free memory, or blocked while creating/ accessing a virtual memory structure. For a process or kernel thread currently running, the last reason it was stopped before obtaining the CPU is shown. On HP-UX 11.0 and beyond, mikslp.text (located in /opt/perf/lib) contains the blocking functions and their corresponding block states for use by midaemon. On a threaded operating system, such as HP-UX 11.0 and beyond, this metric represents a kernel thread characteristic. If this metric is reported for a process, the value for its last executing kernel thread is given. For example, if a process has multiple kernel threads and kernel thread one is the last to execute during the interval, the metric value for kernel thread one is assigned to the process. PROC_SYS_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on system resources during the interval. These resources include data structures from the LVM, VFS, UFS, JFS, and Disk Quota subsystems. “SYSTM” is the “catch-all” wait state for blocks on system resources that are not common enough or long enough to warrant their own stop state. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_SYS_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on SYSTM (that is, system resources) during the interval. These resources include data structures from the LVM, VFS, UFS, JFS, and Disk Quota subsystems. “SYSTM” is the “catch-all” wait state for blocks on system resources that are not common enough or long enough to warrant their own stop state. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_TERM_IO_WAIT_PCT ---------------------------------- The percentage of time the process or kernel thread was blocked on terminal IO (waiting for its terminal IO to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. A percentage of time spent in a wait state is calculated as the time a kernel thread (or all kernel threads of a process) spent waiting in this state, divided by the alive time of the kernel thread (or all kernel threads of the process) during the interval. If this metric is reported for a kernel thread, the percentage value is for that single kernel thread. If this metric is reported for a process, the percentage value is calculated with the sum of the wait and alive times of all of its kernel threads. For example, if a process has 2 kernel threads, one sleeping for the entire interval and one waiting on terminal input for the interval, the process wait percent values will be 50% on Sleep and 50% on Terminal. The kernel thread wait values will be 100% on Sleep for the first kernel thread and 100% on Terminal for the second kernel thread. For another example, consider the same process as above, with 2 kernel threads, one of which was created half-way through the interval, and which then slept for the remainder of the interval. The other kernel thread was waiting for terminal input for half the interval, then used the CPU actively for the remainder of the interval. The process wait percent values will be 33% on Sleep and 33% on Terminal (each one third of the total alive time). The kernel thread wait values will be 100% on Sleep for the first kernel thread and 50% on Terminal for the second kernel thread. PROC_TERM_IO_WAIT_TIME ---------------------------------- The time, in seconds, that the process or kernel thread was blocked on terminal IO (waiting for its terminal IO to complete) during the interval. On a threaded operating system, such as HP-UX 11.0 and beyond, process wait time is calculated by summing the wait times of its kernel threads. If this metric is reported for a kernel thread, the value is the wait time of that single kernel thread. If this metric is reported for a process, the value is the sum of the wait times of all of its kernel threads. Alive kernel threads and kernel threads that have died during the interval are included in the summation. For multi-threaded processes, the wait times can exceed the length of the measurement interval. PROC_THREAD_COUNT ---------------------------------- The total number of kernel threads for the current process. On Linux systems with Kernel 2.5 and below, every thread has its own process ID so this metric will always be 1. On Solaris systems, this metric reflects the total number of Light Weight Processes (LWPs) associated with the process. PROC_TTY ---------------------------------- The controlling terminal for a process(or kernel threads, if HP- UX/Linux Kernel 2.6 and above). This field is blank if there is no controlling terminal. On HP-UX, Linux, and AIX, this is the same as the “TTY” field of the ps command. On all other Unix systems, the controlling terminal name is found by searching the directories provided in the /etc/ttysrch file. See man page ttysrch(4) for details. The matching criteria field (“M”, “F” or “I” values) of the ttysrch file is ignored. If a terminal is not found in one of the ttysrch file directories, the following directories are searched in the order here: “/dev”, “/dev/pts”, “/dev/term” and “dev/xt”. When a match is found in one of the “/dev” subdirectories, “/dev/” is not displayed as part of the terminal name. If no match is found in the directory searches, the major and minor numbers of the controlling terminal are displayed. In most cases, this value is the same as the “TTY” field of the ps command. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. PROC_USER_NAME ---------------------------------- On Unix systems, this is real user name of a process or the login account (from /etc/passwd) of a process (or kernel thread, if HP- UX/Linux Kernel 2.6 and above). If more than one account is listed in /etc/passwd with the same user ID (uid) field, the first one is used. If an account cannot be found that matches the uid field, then the uid number is returned. This would occur if the account was removed after a process was started. On Windows, this is the process owner account name, without the domain name this account resides in. On HP-UX, this metric is specific to a process. If this metric is reported for a kernel thread, the value for its associated process is given. RECORD_TYPE ---------------------------------- ASCII string that identifies the record. Possibilities include: GLOB for global 5 minute detail GSUM for global hourly summary APPL for application 5 minute detail ASUM for application hourly summary CONF for configuration TRAN for transaction tracker detail TSUM for transaction tracker summary Except for Windows Desktop, this also includes: PROC for process 1 minute detail DISK for disk device 5 minute detail DSUM for disk device summary On HP-UX, this also includes: VOLS for logical volume disk detail VSUM for logical volume disk summary TBL_BUFFER_CACHE_AVAIL ---------------------------------- The size (in KBs unless otherwise specified) of the file system buffer cache on the system. On HP-UX 11i v2 and below, these buffers are used for all file system IO operations, as well as all other block IO operations in the system (exec, mount, inode reading, and some device drivers). If dynamic buffer cache is enabled, the system allocates a percentage of available memory not less than dbc_min_pct nor more than dbc_max_pct, depending on the system needs at any given time. On systems with a static buffer cache, this value will remain equal to bufpages, or not less than dbc_min_pct nor more than dbc_max_pct. On HP-UX 11i v3 and above the limits of the file system buffer cache which is still being used for file system metadata are automatically set to certain percentages of filecache_min and filecache_max. On SUN, this value is obtained by multiplying the system page size times the number of buffer headers (nbuf). For example, on a SPARCstation 10 the buffer size is usually (200 (page size buffers) * 4096 (bytes/page) = 800 KB). NOTE: (For SUN systems with VERITAS File System installed) Veritas implemented their Direct I/O feature in their file system to provide mechanism for bypassing the Unix system buffer cache while retaining the on disk structure of a file system. The way in which Direct I/O works involves the way the system buffer cache is handled by the Unix OS. Once the VERITAS file system returns with the requested block, instead of copying the content to a system buffer page, it copies the block into the application’s buffer space. That’s why if you have installed vxfs on your system, the TBL_BUFFER_CACHE_AVAIL can exceed the TBL_BUFFER_CACHE_HWM metric. On SUN, the buffer cache is a memory pool used by the system to cache inode, indirect block and cylinder group related disk accesses. This is different from the traditional concept of a buffer cache that also holds file system data. On Solaris 5.X, as file data is cached, accesses to it show up as virtual memory IOs. File data caching occurs through memory mapping managed by the virtual memory system, not through the buffer cache. The “nbuf” value is dynamic, but it is very hard to create a situation where the memory cache metrics change, since most systems have more than adequate space for inode, indirect block, and cylinder group data caching. This cache is more heavily utilized on NFS file servers. On AIX, this cache is used for all block IO. On AIX System WPARs, this metric is NA. TBL_BUFFER_CACHE_USED ---------------------------------- The size (in KBs unless otherwise specified) of the sum of the currently used buffers. On HP-UX 11i v2 and below, this is normally greater than the amount requested due to internal fragmentation of the buffer cache. Since this is a cache, it is normal for it to be filled. The buffer cache is used to stage all block IOs to disk. On a dynamic buffer cache configuration, this metric is always equal to TBL_BUFFER_CACHE_AVAIL. With dynamic buffer cache, the system allocates a percentage of available memory not less than dbc_min_pct nor more than dbc_max_pct, depending on the system needs at any given time. On systems with a static buffer cache, this value will remain equal to bufpages, or not less than dbc_min_pct nor more than dbc_max_pct. With a static buffer cache, this metric shows the amount of memory within the configured size that is actually used. On HP-UX 11i v3 and above this metric value represents the usage of the file system buffer cache which is still being used for file system metadata. On AIX, this is normally greater than the amount requested due to internal fragmentation of the buffer cache. Since this is a cache, it is normal for it to be filled. The buffer cache is used to stage all block IOs to disk. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. TBL_FILE_LOCK_AVAIL ---------------------------------- The configured number of file or record locks that can be allocated on the system. Files and/or records are locked by calls to lockf(2). On Linux kernel versions 2.4 and above, available file orrecord locks is a dynamic value which can grow upto max unsigned long. TBL_FILE_LOCK_UTIL ---------------------------------- The percentage of configured file or record locks currently in use. On Linux 2.4 and above kernel versions, this may not give correct picture as file or record locks available may change dynamically and can grow upto max unsigned long. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. TBL_FILE_TABLE_AVAIL ---------------------------------- The number of entries in the file table. On HP-UX and AIX, this is the configured maximum number of the file table entries used by the kernel to manage open file descriptors. On HP-UX, this is the sum of the “nfile” and “file_pad” values used in kernel generation. On SUN, this is the number of entries in the file cache. This is a size. All entries are not always in use. The cache size is dynamic. Entries in this cache are used to manage open file descriptors. They are reused as files are closed and new ones are opened. The size of the cache will go up or down in chunks as more or less space is required in the cache. On AIX, the file table entries are dynamically allocated by the kernel if there is no entry available. These entries are allocated in chunks. TBL_FILE_TABLE_UTIL ---------------------------------- The percentage of file table entries currently used by file descriptors. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. TBL_INODE_CACHE_AVAIL ---------------------------------- On HP-UX, this is the configured total number of entries for the incore inode tables on the system. For HP-UX releases prior to 11.2x, this value reflects only the HFS inode table. For subsequent HP-UX releases, this value is the sum of inode tables for both HFS and VxFS file systems (ninode plus vxfs_ninode). On HP-UX, file system directory activity is done through inodes that are stored on disk. The kernel keeps a memory cache of active and recently accessed inodes to reduce disk IOs. When a file is opened through a pathname, the kernel converts the pathname to an inode number and attempts to obtain the inode information from the cache based on the filesystem type. If the inode entry is not in the cache, the inode is read from disk into the inode cache. On HP-UX, the number of used entries in the inode caches are usually at or near the capacity. This does not necessarily indicate that the configured sizes are too small because the tables may contain recently used inodes and inodes referenced by entries in the directory name lookup cache. When a new inode cache entry is required and a free entry does not exist, inactive entries referenced by the directory name cache are used. If after freeing inode entries only referenced by the directory name cache does not create enough free space, the message “inode: table is full” message may appear on the console. If this occurs, increase the size of the kernel parameter, ninode. Low directory name cache hit ratios may also indicate an underconfigured inode cache. On HP-UX, the default formula for the ninode size is: ninode = ((nproc+16+maxusers)+32+ (2*npty)+(4*num_clients)) On all other Unix systems, this is the number of entries in the inode cache. This is a size. All entries are not always in use. The cache size is dynamic. Entries in this cache are reused as files are closed and new ones are opened. The size of the cache will go up or down in chunks as more or less space is required in the cache. Inodes are used to store information about files within the file system. Every file has at least two inodes associated with it (one for the directory and one for the file itself). The information stored in an inode includes the owners, timestamps, size, and an array of indices used to translate logical block numbers to physical sector numbers. There is a separate inode maintained for every view of a file, so if two processes have the same file open, they both use the same directory inode, but separate inodes for the file. TBL_INODE_CACHE_USED ---------------------------------- The number of inode cache entries currently in use. On HP-UX, this is the number of “non-free” inodes currently used. Since the inode table contains recently closed inodes as well as open inodes, the table often appears to be fully utilized. When a new entry is needed, one can usually be found by reusing one of the recently closed inode entries. On HP-UX, file system directory activity is done through inodes that are stored on disk. The kernel keeps a memory cache of active and recently accessed inodes to reduce disk IOs. When a file is opened through a pathname, the kernel converts the pathname to an inode number and attempts to obtain the inode information from the cache based on the filesystem type. If the inode entry is not in the cache, the inode is read from disk into the inode cache. On HP-UX, the number of used entries in the inode caches are usually at or near the capacity. This does not necessarily indicate that the configured sizes are too small because the tables may contain recently used inodes and inodes referenced by entries in the directory name lookup cache. When a new inode cache entry is required and a free entry does not exist, inactive entries referenced by the directory name cache are used. If after freeing inode entries only referenced by the directory name cache does not create enough free space, the message “inode: table is full” message may appear on the console. If this occurs, increase the size of the kernel parameter, ninode. Low directory name cache hit ratios may also indicate an underconfigured inode cache. On HP-UX, the default formula for the ninode size is: ninode = ((nproc+16+maxusers)+32+ (2*npty)+(4*num_clients)) On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. TBL_MSG_TABLE_AVAIL ---------------------------------- The configured maximum number of message queues that can be allocated on the system. A message queue is allocated by a program using the msgget(2) call. Refer to the ipcs(1) man page for more information. On SUN, the InterProcess Communication facilities are dynamically loadable. If the amount available is zero, this facility was not loaded when data collection began, and its data is not obtainable. The data collector is unable to determine that a facility has been loaded once data collection has started. If you know a new facility has been loaded, restart the data collection, and the data for that facility will be collected. See ipcs(1) to report on interprocess communication resources. TBL_MSG_TABLE_UTIL ---------------------------------- The percentage of configured message queues currently in use. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. TBL_PROC_TABLE_AVAIL ---------------------------------- The configured maximum number of the proc table entries used by the kernel to manage processes. This number includes both free and used entries. On HP-UX, this is set by the NPROC value during system generation. AIX has a “dynamic” proc table, which means that AVAIL has been set higher than should ever be needed. On AIX System WPARs, this metric is NA. TBL_PROC_TABLE_UTIL ---------------------------------- The percentage of proc table entries currently used by processes. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. On Solaris non-global zones, this metric is N/A. TBL_SEM_TABLE_AVAIL ---------------------------------- The configured number of semaphore identifiers (sets) that can be allocated on the system. On SUN, the InterProcess Communication facilities are dynamically loadable. If the amount available is zero, this facility was not loaded when data collection began, and its data is not obtainable. The data collector is unable to determine that a facility has been loaded once data collection has started. If you know a new facility has been loaded, restart the data collection, and the data for that facility will be collected. See ipcs(1) to report on interprocess communication resources. TBL_SEM_TABLE_UTIL ---------------------------------- The percentage of configured semaphores identifiers currently in use. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. TBL_SHMEM_TABLE_AVAIL ---------------------------------- The configured number of shared memory segments that can be allocated on the system. On SUN, the InterProcess Communication facilities are dynamically loadable. If the amount available is zero, this facility was not loaded when data collection began, and its data is not obtainable. The data collector is unable to determine that a facility has been loaded once data collection has started. If you know a new facility has been loaded, restart the data collection, and the data for that facility will be collected. See ipcs(1) to report on interprocess communication resources. TBL_SHMEM_TABLE_UTIL ---------------------------------- The percentage of configured shared memory segments currently in use. On Unix systems, this metric is updated every 30 seconds or the sampling interval, whichever is greater. TIME ---------------------------------- The local time of day for the start of the interval. The time is an ASCII field in hh:mm:ss 24-hour format. This field will always contain 8 characters in ASCII files. The three subfields (hh, mm, ss) will contain a leading zero if the value is less than 10. This metric is extracted from GBL_STATTIME, which is obtained using the time() system call at the start of the interval. This field responds to language localization. In binary files this field contains four byte size subfields. The most significant byte contains the hour, the next most significant byte contains the minute, then the seconds and finally the tenths of a second. The left two bytes can be isolated by dividing by 65536. HHMM = TIME/65536. Then HOUR = HHMM/256 and MINUTE = HHMM mod 256. SSTS = TIME mod 65536. Then SECOND = SSTS/256. TTBIN_TRANS_COUNT_1 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_10 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_2 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_3 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_4 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_5 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_6 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_7 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_8 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_TRANS_COUNT_9 ---------------------------------- The number of completed transactions in this range during the last interval. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_1 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_10 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_2 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_3 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_4 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_5 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_6 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_7 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_8 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TTBIN_UPPER_RANGE_9 ---------------------------------- The upper range (transaction time) for this bin. There are a maximum of nine user-defined transaction response time bins (TTBIN_UPPER_RANGE). The last bin, which is not specified in the transaction configuration file (ttdconf.mwc on Windows or ttd.conf on UNIX platforms), is the overflow bin and will always have a value of -2 (overflow). Note that the values specified in the transaction configuration file cannot exceed 2147483.6, which is the number of seconds in 24.85 days. If the user specifies any values greater than 2147483.6, the numbers reported for those bins or Service Level Objectives (SLO) will be -2. On SUN systems, this metric is only available on 5.X or later. TT_ABORT ---------------------------------- The number of aborted transactions during the last interval for this transaction. TT_ABORT_WALL_TIME_PER_TRAN ---------------------------------- The average time, in seconds, per aborted transaction during the last interval. On SUN systems, this metric is only available on 5.X or later. TT_APP_NAME ---------------------------------- The registered ARM Application name. TT_APP_TRAN_NAME ---------------------------------- A concatenation of TT_APP_NAME and TT_NAME. This provides a way to uniquely identify a specific transaction. The field is limited to 60 characters. TT_CLIENT_ADDRESS ---------------------------------- The correlator address. This is the address where the child transaction originated. TT_CLIENT_ADDRESS_FORMAT ---------------------------------- The correlator address format. This shows the protocol family for the client network address. Refer to the ARM API Guide for the list and description of supported address formats. TT_CLIENT_TRAN_ID ---------------------------------- A numerical ID that uniquely identifies the transaction class in this correlator. TT_COUNT ---------------------------------- The number of completed transactions during the last interval for this transaction. TT_CPU_TOTAL_TIME_PER_TRAN ---------------------------------- The average total CPU time, in seconds, consumed by each completed instance of the transaction during the interval. Total CPU time is the sum of the CPU time components for a process or kernel thread, including system, user, context switch, interrupt processing, realtime, and nice utilization values. Per-transaction performance resource metrics represent an average for all completed instances of the given transaction during the interval. If there are no completed transaction instances during an interval, then there are no resources accounted, even though there may be in-progress transactions using resources which have not completed. Resource metrics for in-progress transactions will be shown in the interval after they complete (that is, after the process has called arm_stop). If there is only one completed transaction instance during an interval, then the resources attributed to the transaction will represent the resources used by the process between its call to arm_start and arm_stop, even if arm_start was called before the current interval. Thus, the resource usage time or wall time per transaction can exceed the current collection interval time. If there are several completed transaction instances during an interval for a given transaction, then the resources attributed to the transaction will represent an average for all completed instances during the interval. To obtain the total accumulated resource consumption for all completed transactions during an interval, multiply the resource metric by the number of completed transaction instances during the interval (TT_COUNT). TT_DISK_LOGL_IO_PER_TRAN ---------------------------------- The average number of logical IOs made by (or for) each completed instance of the transaction during the interval. NFS mounted disks are not included in this list. “Disk” refers to a physical drive (that is, “spindle”), not a partition on a drive (unless the partition occupies the entire physical disk). On many Unix systems, logical disk IOs are measured by counting the read and write system calls that are directed to disk devices. Also counted are read and write system calls made indirectly through other system calls, including readv, recvfrom, recv, recvmsg, ipcrecvcn, recfrom, writev, send, sento, sendmsg, and ipcsend. Per-transaction performance resource metrics represent an average for all completed instances of the given transaction during the interval. If there are no completed transaction instances during an interval, then there are no resources accounted, even though there may be in-progress transactions using resources which have not completed. Resource metrics for in-progress transactions will be shown in the interval after they complete (that is, after the process has called arm_stop). If there is only one completed transaction instance during an interval, then the resources attributed to the transaction will represent the resources used by the process between its call to arm_start and arm_stop, even if arm_start was called before the current interval. Thus, the resource usage time or wall time per transaction can exceed the current collection interval time. If there are several completed transaction instances during an interval for a given transaction, then the resources attributed to the transaction will represent an average for all completed instances during the interval. To obtain the total accumulated resource consumption for all completed transactions during an interval, multiply the resource metric by the number of completed transaction instances during the interval (TT_COUNT). TT_DISK_PHYS_IO_PER_TRAN ---------------------------------- The average number of physical disk IOs per second made by each completed instance of the transaction during the interval. For transactions which run for less than the measurement interval, this metric is normalized over the measurement interval. For example, a transaction ran for 1 second and did 50 IOs during its life. If the measurement interval is 5 seconds, it is reported as having done 10 IOs per second. If the measurement interval is 60 seconds, it is reported as having done 50/60 or 0.83 IOs per second. “Disk” in this instance refers to any locally attached physical disk drives (that is, “spindles”) that may hold file systems and/or swap. NFS mounted disks are not included in this list. On HP-UX, since this value is reported by the drivers, multiple physical requests that have been collapsed to a single physical operation (due to driver IO merging) are only counted once. Per-transaction performance resource metrics represent an average for all completed instances of the given transaction during the interval. If there are no completed transaction instances during an interval, then there are no resources accounted, even though there may be in-progress transactions using resources which have not completed. Resource metrics for in-progress transactions will be shown in the interval after they complete (that is, after the process has called arm_stop). If there is only one completed transaction instance during an interval, then the resources attributed to the transaction will represent the resources used by the process between its call to arm_start and arm_stop, even if arm_start was called before the current interval. Thus, the resource usage time or wall time per transaction can exceed the current collection interval time. If there are several completed transaction instances during an interval for a given transaction, then the resources attributed to the transaction will represent an average for all completed instances during the interval. To obtain the total accumulated resource consumption for all completed transactions during an interval, multiply the resource metric by the number of completed transaction instances during the interval (TT_COUNT). TT_FAILED ---------------------------------- The number of Failed transactions during the last interval for this transaction name. TT_INFO ---------------------------------- The registered ARM Transaction Information for this transaction. TT_NAME ---------------------------------- The registered transaction name for this transaction. TT_NUM_BINS ---------------------------------- The number of distribution ranges. On SUN systems, this metric is only available on 5.X or later. TT_SLO_COUNT ---------------------------------- The number of completed transactions that violated the defined Service Level Objective (SLO) by exceeding the SLO threshold time during the interval. TT_SLO_PERCENT ---------------------------------- The percentage of transactions which violate service level objectives. TT_SLO_THRESHOLD ---------------------------------- The upper range (transaction time) of the Service Level Objective (SLO) threshold value. This value is used to count the number of transactions that exceed this user-supplied transaction time value. TT_TERM_TRAN_1_HR_RATE ---------------------------------- For this transaction name, the number of completed transactions calculated to a 1 hour rate. For example, if you completed five of these transactions in a 5 minute window, the rate is 60 transactions per hour. On SUN systems, this metric is only available on 5.X or later. TT_TRAN_1_MIN_RATE ---------------------------------- For this transaction name, the number of completed transactions calculated to a 1 minute rate. For example, if you completed five of these transactions in a 5 minute window, the rate is one transaction per minute. TT_TRAN_ID ---------------------------------- The registered ARM Transaction ID for this transaction class as returned by arm_getid(). A unique transaction id is returned for a unique application id (returned by arm_init), tran name, and meta data buffer contents. TT_UNAME ---------------------------------- The registered ARM Transaction User Name for this transaction. If the arm_init function has NULL for the appl_user_id field, then the user name is blank. Otherwise, if “*” was specified, then the user name is displayed. For example, to show the user name for the armsample1 program, use: appl_id = arm_init(“armsample1”,”*”,0,0,0); To ignore the user name for the armsample1 program, use: appl_id = arm_init(“armsample1”,NULL,0,0,0); TT_USER_MEASUREMENT_AVG ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the average counter differences of the transaction or transaction instance during the last interval. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this returns the average of the values passed on any ARM call for the transaction or transaction instance during the last interval. TT_USER_MEASUREMENT_AVG_2 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the average counter differences of the transaction or transaction instance during the last interval. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this returns the average of the values passed on any ARM call for the transaction or transaction instance during the last interval. TT_USER_MEASUREMENT_AVG_3 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the average counter differences of the transaction or transaction instance during the last interval. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this returns the average of the values passed on any ARM call for the transaction or transaction instance during the last interval. TT_USER_MEASUREMENT_AVG_4 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the average counter differences of the transaction or transaction instance during the last interval. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this returns the average of the values passed on any ARM call for the transaction or transaction instance during the last interval. TT_USER_MEASUREMENT_AVG_5 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the average counter differences of the transaction or transaction instance during the last interval. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this returns the average of the values passed on any ARM call for the transaction or transaction instance during the last interval. TT_USER_MEASUREMENT_AVG_6 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the average counter differences of the transaction or transaction instance during the last interval. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this returns the average of the values passed on any ARM call for the transaction or transaction instance during the last interval. TT_USER_MEASUREMENT_COUNT ---------------------------------- This returns the total number of times the associated user defined metric (UDM) was sampled during the last interval. TT_USER_MEASUREMENT_COUNT_2 ---------------------------------- This returns the total number of times the associated user defined metric (UDM) was sampled during the last interval. TT_USER_MEASUREMENT_COUNT_3 ---------------------------------- This returns the total number of times the associated user defined metric (UDM) was sampled during the last interval. TT_USER_MEASUREMENT_COUNT_4 ---------------------------------- This returns the total number of times the associated user defined metric (UDM) was sampled during the last interval. TT_USER_MEASUREMENT_COUNT_5 ---------------------------------- This returns the total number of times the associated user defined metric (UDM) was sampled during the last interval. TT_USER_MEASUREMENT_COUNT_6 ---------------------------------- This returns the total number of times the associated user defined metric (UDM) was sampled during the last interval. TT_USER_MEASUREMENT_MAX ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the highest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the highest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MAX_2 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the highest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the highest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MAX_3 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the highest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the highest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MAX_4 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the highest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the highest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MAX_5 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the highest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the highest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MAX_6 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the highest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the highest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MIN ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the lowest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the lowest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MIN_2 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the lowest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the lowest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MIN_3 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the lowest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the lowest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MIN_4 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the lowest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the lowest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MIN_5 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the lowest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the lowest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_MIN_6 ---------------------------------- If the measurement type is a numeric or a string, this metric returns “na”. If the measurement type is a counter, this metric returns the lowest measured counter value over the life of the transaction or transaction instance. The counter value is the difference observed from a counter between the start and the stop (or last update) of a transaction. If the measurement type is a gauge, this metric returns the lowest value passed on any ARM call over the life of the transaction or transaction instance. TT_USER_MEASUREMENT_NAME ---------------------------------- The name of the user defined transactional measurement. The length of the string complies with the ARM 2.0 standard, which is 44 characters long (there are 43 usable characters since this is a NULL terminated character string). TT_USER_MEASUREMENT_NAME_2 ---------------------------------- The name of the user defined transactional measurement. The length of the string complies with the ARM 2.0 standard, which is 44 characters long (there are 43 usable characters since this is a NULL terminated character string). TT_USER_MEASUREMENT_NAME_3 ---------------------------------- The name of the user defined transactional measurement. The length of the string complies with the ARM 2.0 standard, which is 44 characters long (there are 43 usable characters since this is a NULL terminated character string). TT_USER_MEASUREMENT_NAME_4 ---------------------------------- The name of the user defined transactional measurement. The length of the string complies with the ARM 2.0 standard, which is 44 characters long (there are 43 usable characters since this is a NULL terminated character string). TT_USER_MEASUREMENT_NAME_5 ---------------------------------- The name of the user defined transactional measurement. The length of the string complies with the ARM 2.0 standard, which is 44 characters long (there are 43 usable characters since this is a NULL terminated character string). TT_USER_MEASUREMENT_NAME_6 ---------------------------------- The name of the user defined transactional measurement. The length of the string complies with the ARM 2.0 standard, which is 44 characters long (there are 43 usable characters since this is a NULL terminated character string). TT_WALL_TIME_PER_TRAN ---------------------------------- The average transaction time, in seconds, during the last interval for this transaction. YEAR ---------------------------------- The year, including the century, the data in this record was captured. This metric will contain 4 digits, such as 2002.