kernel-ark/Documentation/sched-stats.txt

Version 10 of schedstats includes support for sched_domains, which
hit the mainline kernel in 2.6.7.  Some counters make more sense to be
per-runqueue; other to be per-domain.  Note that domains (and their associated
information) will only be pertinent and available on machines utilizing
CONFIG_SMP.

In version 10 of schedstat, 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.  At this time, there
are no architectures which need more than three domain levels. The first
field in the domain stats is a bit map indicating which cpus are affected
by that domain.

These fields are counters, and only increment.  Programs which make use
of these will need to start with a baseline observation and then calculate
the change in the counters at each subsequent observation.  A perl script
which does this for many of the fields is available at

    http://eaglet.rain.com/rick/linux/schedstat/

Note that any such script will necessarily be version-specific, as the main
reason to change versions is changes in the output format.  For those wishing
to write their own scripts, the fields are described here.

CPU statistics
--------------
cpu<N> 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

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 four are schedule() statistics:
     5) # of times the active queue had at least one other process on it
     6) # of times we switched to the expired queue and reused it
     7) # of times schedule() was called
     8) # of times schedule() left the processor idle

Next four are active_load_balance() statistics:
     9) # of times active_load_balance() was called
    10) # of times active_load_balance() caused this cpu to gain a task
    11) # of times active_load_balance() caused this cpu to lose a task
    12) # of times active_load_balance() tried to move a task and failed

Next three are try_to_wake_up() statistics:
    13) # of times try_to_wake_up() was called
    14) # of times try_to_wake_up() successfully moved the awakening task
    15) # of times try_to_wake_up() attempted to move the awakening task

Next two are wake_up_new_task() statistics:
    16) # of times wake_up_new_task() was called
    17) # of times wake_up_new_task() successfully moved the new task

Next one is a sched_migrate_task() statistic:
    18) # of times sched_migrate_task() was called

Next one is a sched_balance_exec() statistic:
    19) # of times sched_balance_exec() was called

Next three are statistics describing scheduling latency:
    20) sum of all time spent running by tasks on this processor (in ms)
    21) sum of all time spent waiting to run by tasks on this processor (in ms)
    22) # of tasks (not necessarily unique) given to the processor

The last six are statistics dealing with pull_task():
    23) # of times pull_task() moved a task to this cpu when newly idle
    24) # of times pull_task() stole a task from this cpu when another cpu
	was newly idle
    25) # of times pull_task() moved a task to this cpu when idle
    26) # of times pull_task() stole a task from this cpu when another cpu
	was idle
    27) # of times pull_task() moved a task to this cpu when busy
    28) # of times pull_task() stole a task from this cpu when another cpu
	was busy


Domain statistics
-----------------
One of these is produced per domain for each cpu described. (Note that if
CONFIG_SMP is not defined, *no* domains are utilized and these lines
will not appear in the output.)

domain<N> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

The first field is a bit mask indicating what cpus this domain operates over.

The next fifteen are a variety of 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 when the cpu
	was busy
     3) # of times in this domain load_balance() was called when the cpu
	was just becoming idle
     4) # of times in this domain load_balance() tried to move one or more
	tasks and failed, when the cpu was idle
     5) # of times in this domain load_balance() tried to move one or more
	tasks and failed, when the cpu was busy
     6) # of times in this domain load_balance() tried to move one or more
	tasks and failed, when the cpu was just becoming idle
     7) sum of imbalances discovered (if any) with each call to
	load_balance() in this domain when the cpu was idle
     8) sum of imbalances discovered (if any) with each call to
	load_balance() in this domain when the cpu was busy
     9) sum of imbalances discovered (if any) with each call to
	load_balance() in this domain when the cpu was just becoming idle
    10) # of times in this domain load_balance() was called but did not find
	a busier queue while the cpu was idle
    11) # of times in this domain load_balance() was called but did not find
	a busier queue while the cpu was busy
    12) # of times in this domain load_balance() was called but did not find
	a busier queue while the cpu was just becoming idle
    13) # of times in this domain a busier queue was found while the cpu was
	idle but no busier group was found
    14) # of times in this domain a busier queue was found while the cpu was
	busy but no busier group was found
    15) # of times in this domain a busier queue was found while the cpu was
	just becoming idle but no busier group was found

Next two are sched_balance_exec() statistics:
    17) # of times in this domain sched_balance_exec() successfully pushed
	a task to a new cpu
    18) # of times in this domain sched_balance_exec() tried but failed to
	push a task to a new cpu

Next two are try_to_wake_up() statistics:
    19) # of times in this domain try_to_wake_up() tried to move a task based
	on affinity and cache warmth
    20) # of times in this domain try_to_wake_up() tried to move a task based
	on load balancing


/proc/<pid>/schedstat
----------------
schedstats also adds a new /proc/<pid/schedstat file to include some of
the same information on a per-process level.  There are three fields in
this file correlating to fields 20, 21, and 22 in the CPU fields, but
they only apply for that process.

A program could be easily written to make use of these extra fields to
report on how well a particular process or set of processes is faring
under the scheduler's policies.  A simple version of such a program is
available at
    http://eaglet.rain.com/rick/linux/schedstat/v10/latency.c