Linux Scheduler Statistics
/proc/schedstat format and /proc/<pid>/schedstat description
/proc/schedstat format and /proc/<pid>/schedstat description version 12

Support for sched_domains, which hit the mainline kernel in 2.6.7, requires that some counters be per-runqueue; others to be per-domain. This are detailed below.

There is at least one level of domain statistics for each cpu listed, and there may well be more than one domain. Domains have no particular names in this implementation, but the highest numbered one typically arbitrates balancing across all the cpus on the machine, while domain0 is the most tightly focused domain, sometimes balancing only between pairs of cpus. The first field in the domain stats is a bit map indicating which cpus are affected by that domain.

CPU statistics

cpuN 1 2 3 4 5 6 7 8 9 10 11 12
NOTE: In the sched_yield() statistics, the active queue is considered empty if it has only one process in it, since obviously the process calling sched_yield() is that process.
First four fields are sched_yield() statistics:
  1. # of times both the active and the expired queue were empty
  2. # of times just the active queue was empty
  3. # of times just the expired queue was empty
  4. # of times sched_yield() was called


Next three are schedule() statistics:

  1. # of times the active queue had at least one other process on it.
  2. # of times schedule() was called
  3. # of times schedule() left the processor idle
Next two are try_to_wake_up() statistics:
  1. # of times try_to_wake_up() was called
  2. # of times try_to_wake_up() discovered the process being awakened last ran on the waking cpu


Next three are statistics describing scheduling latency:

  1. sum of all time spent running by tasks on this processor (in ms)
  2. sum of all time spent waiting to run by tasks on this processor (in ms)
  3. # of tasks (not necessarily unique) given to the processor



Domain statistics

domainN 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
The first field is a bit mask indicating what cpus this domain operates over.


The next twenty-four are load_balance() statistics:

  1. # of times in this domain load_balance() was called when the cpu was idle
  2. # of times in this domain load_balance() was called while idle but found no imbalance
  3. # of times in this domain load_balance() tried to move one or more tasks and failed, when the cpu was idle
  4. sum of imbalances discovered (if any) with each call to load_balance() in this domain when the cpu was idle
  5. # of times in this domain load_balance() pulled tasks to this cpu while idle
  6. # of times in this domain load_balance() pulled a task to this cpueven though it was hot on its original cpu
  7. # of times in this domain load_balance() was called but did not find a busier queue while the cpu was idle
  8. # of times in this domain a busier queue was found while the cpu was idle but no busier group was found

  9. # of times in this domain load_balance() was called when the cpu was just becoming idle
  10. # of times in this domain load_balance() was called when just becoming idle but found no imbalance
  11. # of times in this domain load_balance() tried to move one or more tasks and failed, when the cpu was just becoming idle
  12. sum of imbalances discovered (if any) with each call to load_balance() in this domain when the cpu was just becoming idle
  13. # of times in this domain load_balance() pulled tasks to this cpu when newly idle
  14. # oh times in this domain load_balance()pulled tasks to this cpu even thought the tasks were hot in their original cpus.t
  15. # of times in this domain load_balance() was called but did not find a busier queue while the cpu was just becoming idle
  16. # of times in this domain a busier queue was found while the cpu was just becoming idle but no busier group was found
  1. # of times in this domain load_balance() was called when the cpu was busy
  2. # of times in this domain load_balance() was called while busy but found no imbalance
  3. # of times in this domain load_balance() tried to move one or more tasks and failed, when the cpu was busy
  4. sum of imbalances discovered (if any) with each call to load_balance() in this domain when the cpu was busy
  5. # of times in this domain load_balance() pulled tasks to this cpu while busy
  6. # of times in this domain load_balance() was called but did not find a busier queue while the cpu was busy
  7. # of times in this domain a busier queue was found while the cpu was busy but no busier group was found


Next three are active_load_balance() statistics:

  1. # of times active_load_balance() was called
  2. # of times active_load_balance() tried to move a task and failed
  3. # of times active_load_balance() successfully pushed a task off a procssor.


Next six are always zero.

  1. 0
  2. 0
  3. 0
  4. 0
  5. 0
  6. 0


Next three are try_to_wake_up() statistics:

  1. # of times in this domain try_to_wake_up() woke a process on a remote cpu
  2. # of times in this domain try_to_wake_up() decided to choose a cpu based on SD_WAKE_AFFINE.
  3. # of times in this domain try_to_wake_up() decided to choose a cpu based on SD_WAKE_BALANCE.


/proc/<pid>/schedstat

Schedstats creates a file in the /proc/<pid>/ directory named schedstat which contains statistics for the individual processes. It includes the same information that fields 10-12 do in the cpu stats, above, but apply only for that process. The program latency.c makes use of these extra fields to report on how well a particular process is faring under the scheduler's policies. The example below utilizes that program on a particular process, rather than examining runqueue statistics through sampling /proc/schedstat. The program observed, loadtest, is a simply-written cpu-intensive program and thus uses up most of its allocated timeslice without voluntarily pausing for I/O. Processes such as cc and bash may well pause for other I/O events and give up the cpu at times, and thus appear to have much smaller timeslices. But it is important to remember that avgrun only tells us, on average, how long we were on the cpu each time, not what our given timeslice was. 

% latency 25611
25611 (loadtest) avgrun=60.36ms avgwait=0.00ms
25611 (loadtest) avgrun=92.56ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.96ms avgwait=0.00ms
25611 (loadtest) avgrun=99.96ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.02ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.94ms avgwait=0.00ms
25611 (loadtest) avgrun=99.92ms avgwait=0.02ms
25611 (loadtest) avgrun=99.96ms avgwait=0.00ms
Process 25611 (loadtest) has exited.
%

Since the above test was done on an unloaded, multiple-cpu machine, loadtest pretty much had a cpu to itself, was granted about a 100ms timeslice, and used virtually all of it before giving up the cpu. Renicing loadtest can show dramatically how the timeslice changes with the priority of the process. Running N+1 loadtests, where N is the number of processors on the machine, introduces contention and the avgwait field starts to go up significantly.
Questions to ricklind@us.ibm.com