Documentation/scheduler/sched-stats.txt

   1 Version 15 of schedstats dropped counters for some sched_yield:
   2 yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is
   3 identical to version 14.
   4
   5 Version 14 of schedstats includes support for sched_domains, which hit the
   6 mainline kernel in 2.6.20 although it is identical to the stats from version
   7 12 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
   8 release).  Some counters make more sense to be per-runqueue; other to be
   9 per-domain.  Note that domains (and their associated information) will only
  10 be pertinent and available on machines utilizing CONFIG_SMP.
  11
  12 In version 14 of schedstat, there is at least one level of domain
  13 statistics for each cpu listed, and there may well be more than one
  14 domain.  Domains have no particular names in this implementation, but
  15 the highest numbered one typically arbitrates balancing across all the
  16 cpus on the machine, while domain0 is the most tightly focused domain,
  17 sometimes balancing only between pairs of cpus.  At this time, there
  18 are no architectures which need more than three domain levels. The first
  19 field in the domain stats is a bit map indicating which cpus are affected
  20 by that domain.
  21
  22 These fields are counters, and only increment.  Programs which make use
  23 of these will need to start with a baseline observation and then calculate
  24 the change in the counters at each subsequent observation.  A perl script
  25 which does this for many of the fields is available at
  26
  27     http://eaglet.rain.com/rick/linux/schedstat/
  28
  29 Note that any such script will necessarily be version-specific, as the main
  30 reason to change versions is changes in the output format.  For those wishing
  31 to write their own scripts, the fields are described here.
  32
  33 CPU statistics
  34 --------------
  35 cpu<N> 1 2 3 4 5 6 7 8 9
  36
  37 First field is a sched_yield() statistic:
  38      1) # of times sched_yield() was called
  39
  40 Next three are schedule() statistics:
  41      2) # of times we switched to the expired queue and reused it
  42      3) # of times schedule() was called
  43      4) # of times schedule() left the processor idle
  44
  45 Next two are try_to_wake_up() statistics:
  46      5) # of times try_to_wake_up() was called
  47      6) # of times try_to_wake_up() was called to wake up the local cpu
  48
  49 Next three are statistics describing scheduling latency:
  50      7) sum of all time spent running by tasks on this processor (in jiffies)
  51      8) sum of all time spent waiting to run by tasks on this processor (in
  52         jiffies)
  53      9) # of timeslices run on this cpu
  54
  55
  56 Domain statistics
  57 -----------------
  58 One of these is produced per domain for each cpu described. (Note that if
  59 CONFIG_SMP is not defined, *no* domains are utilized and these lines
  60 will not appear in the output.)
  61
  62 domain<N> <cpumask> 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
  63
  64 The first field is a bit mask indicating what cpus this domain operates over.
  65
  66 The next 24 are a variety of load_balance() statistics in grouped into types
  67 of idleness (idle, busy, and newly idle):
  68
  69      1) # of times in this domain load_balance() was called when the
  70         cpu was idle
  71      2) # of times in this domain load_balance() checked but found
  72         the load did not require balancing when the cpu was idle
  73      3) # of times in this domain load_balance() tried to move one or
  74         more tasks and failed, when the cpu was idle
  75      4) sum of imbalances discovered (if any) with each call to
  76         load_balance() in this domain when the cpu was idle
  77      5) # of times in this domain pull_task() was called when the cpu
  78         was idle
  79      6) # of times in this domain pull_task() was called even though
  80         the target task was cache-hot when idle
  81      7) # of times in this domain load_balance() was called but did
  82         not find a busier queue while the cpu was idle
  83      8) # of times in this domain a busier queue was found while the
  84         cpu was idle but no busier group was found
  85
  86      9) # of times in this domain load_balance() was called when the
  87         cpu was busy
  88     10) # of times in this domain load_balance() checked but found the
  89         load did not require balancing when busy
  90     11) # of times in this domain load_balance() tried to move one or
  91         more tasks and failed, when the cpu was busy
  92     12) sum of imbalances discovered (if any) with each call to
  93         load_balance() in this domain when the cpu was busy
  94     13) # of times in this domain pull_task() was called when busy
  95     14) # of times in this domain pull_task() was called even though the
  96         target task was cache-hot when busy
  97     15) # of times in this domain load_balance() was called but did not
  98         find a busier queue while the cpu was busy
  99     16) # of times in this domain a busier queue was found while the cpu
 100         was busy but no busier group was found
 101
 102     17) # of times in this domain load_balance() was called when the
 103         cpu was just becoming idle
 104     18) # of times in this domain load_balance() checked but found the
 105         load did not require balancing when the cpu was just becoming idle
 106     19) # of times in this domain load_balance() tried to move one or more
 107         tasks and failed, when the cpu was just becoming idle
 108     20) sum of imbalances discovered (if any) with each call to
 109         load_balance() in this domain when the cpu was just becoming idle
 110     21) # of times in this domain pull_task() was called when newly idle
 111     22) # of times in this domain pull_task() was called even though the
 112         target task was cache-hot when just becoming idle
 113     23) # of times in this domain load_balance() was called but did not
 114         find a busier queue while the cpu was just becoming idle
 115     24) # of times in this domain a busier queue was found while the cpu
 116         was just becoming idle but no busier group was found
 117
 118    Next three are active_load_balance() statistics:
 119     25) # of times active_load_balance() was called
 120     26) # of times active_load_balance() tried to move a task and failed
 121     27) # of times active_load_balance() successfully moved a task
 122
 123    Next three are sched_balance_exec() statistics:
 124     28) sbe_cnt is not used
 125     29) sbe_balanced is not used
 126     30) sbe_pushed is not used
 127
 128    Next three are sched_balance_fork() statistics:
 129     31) sbf_cnt is not used
 130     32) sbf_balanced is not used
 131     33) sbf_pushed is not used
 132
 133    Next three are try_to_wake_up() statistics:
 134     34) # of times in this domain try_to_wake_up() awoke a task that
 135         last ran on a different cpu in this domain
 136     35) # of times in this domain try_to_wake_up() moved a task to the
 137         waking cpu because it was cache-cold on its own cpu anyway
 138     36) # of times in this domain try_to_wake_up() started passive balancing
 139
 140 /proc/<pid>/schedstat
 141 ----------------
 142 schedstats also adds a new /proc/<pid>/schedstat file to include some of
 143 the same information on a per-process level.  There are three fields in
 144 this file correlating for that process to:
 145      1) time spent on the cpu
 146      2) time spent waiting on a runqueue
 147      3) # of timeslices run on this cpu
 148
 149 A program could be easily written to make use of these extra fields to
 150 report on how well a particular process or set of processes is faring
 151 under the scheduler's policies.  A simple version of such a program is
 152 available at
 153     http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c