delay-accounting.rst (4692B)
1================ 2Delay accounting 3================ 4 5Tasks encounter delays in execution when they wait 6for some kernel resource to become available e.g. a 7runnable task may wait for a free CPU to run on. 8 9The per-task delay accounting functionality measures 10the delays experienced by a task while 11 12a) waiting for a CPU (while being runnable) 13b) completion of synchronous block I/O initiated by the task 14c) swapping in pages 15d) memory reclaim 16e) thrashing page cache 17f) direct compact 18g) write-protect copy 19 20and makes these statistics available to userspace through 21the taskstats interface. 22 23Such delays provide feedback for setting a task's cpu priority, 24io priority and rss limit values appropriately. Long delays for 25important tasks could be a trigger for raising its corresponding priority. 26 27The functionality, through its use of the taskstats interface, also provides 28delay statistics aggregated for all tasks (or threads) belonging to a 29thread group (corresponding to a traditional Unix process). This is a commonly 30needed aggregation that is more efficiently done by the kernel. 31 32Userspace utilities, particularly resource management applications, can also 33aggregate delay statistics into arbitrary groups. To enable this, delay 34statistics of a task are available both during its lifetime as well as on its 35exit, ensuring continuous and complete monitoring can be done. 36 37 38Interface 39--------- 40 41Delay accounting uses the taskstats interface which is described 42in detail in a separate document in this directory. Taskstats returns a 43generic data structure to userspace corresponding to per-pid and per-tgid 44statistics. The delay accounting functionality populates specific fields of 45this structure. See 46 47 include/uapi/linux/taskstats.h 48 49for a description of the fields pertaining to delay accounting. 50It will generally be in the form of counters returning the cumulative 51delay seen for cpu, sync block I/O, swapin, memory reclaim, thrash page 52cache, direct compact, write-protect copy etc. 53 54Taking the difference of two successive readings of a given 55counter (say cpu_delay_total) for a task will give the delay 56experienced by the task waiting for the corresponding resource 57in that interval. 58 59When a task exits, records containing the per-task statistics 60are sent to userspace without requiring a command. If it is the last exiting 61task of a thread group, the per-tgid statistics are also sent. More details 62are given in the taskstats interface description. 63 64The getdelays.c userspace utility in tools/accounting directory allows simple 65commands to be run and the corresponding delay statistics to be displayed. It 66also serves as an example of using the taskstats interface. 67 68Usage 69----- 70 71Compile the kernel with:: 72 73 CONFIG_TASK_DELAY_ACCT=y 74 CONFIG_TASKSTATS=y 75 76Delay accounting is disabled by default at boot up. 77To enable, add:: 78 79 delayacct 80 81to the kernel boot options. The rest of the instructions below assume this has 82been done. Alternatively, use sysctl kernel.task_delayacct to switch the state 83at runtime. Note however that only tasks started after enabling it will have 84delayacct information. 85 86After the system has booted up, use a utility 87similar to getdelays.c to access the delays 88seen by a given task or a task group (tgid). 89The utility also allows a given command to be 90executed and the corresponding delays to be 91seen. 92 93General format of the getdelays command:: 94 95 getdelays [-dilv] [-t tgid] [-p pid] 96 97Get delays, since system boot, for pid 10:: 98 99 # ./getdelays -d -p 10 100 (output similar to next case) 101 102Get sum of delays, since system boot, for all pids with tgid 5:: 103 104 # ./getdelays -d -t 5 105 print delayacct stats ON 106 TGID 5 107 108 109 CPU count real total virtual total delay total delay average 110 8 7000000 6872122 3382277 0.423ms 111 IO count delay total delay average 112 0 0 0ms 113 SWAP count delay total delay average 114 0 0 0ms 115 RECLAIM count delay total delay average 116 0 0 0ms 117 THRASHING count delay total delay average 118 0 0 0ms 119 COMPACT count delay total delay average 120 0 0 0ms 121 WPCOPY count delay total delay average 122 0 0 0ms 123 124Get IO accounting for pid 1, it works only with -p:: 125 126 # ./getdelays -i -p 1 127 printing IO accounting 128 linuxrc: read=65536, write=0, cancelled_write=0 129 130The above command can be used with -v to get more debug information.