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) This field is a legacy array expiration count field used in the O(1)
  42         scheduler. We kept it for ABI compatibility, but it is always set to zero.
  43      3) # of times schedule() was called
  44      4) # of times schedule() left the processor idle
  45
  46 Next two are try_to_wake_up() statistics:
  47      5) # of times try_to_wake_up() was called
  48      6) # of times try_to_wake_up() was called to wake up the local cpu
  49
  50 Next three are statistics describing scheduling latency:
  51      7) sum of all time spent running by tasks on this processor (in jiffies)
  52      8) sum of all time spent waiting to run by tasks on this processor (in
  53         jiffies)
  54      9) # of timeslices run on this cpu
  55
  56
  57 Domain statistics
  58 -----------------
  59 One of these is produced per domain for each cpu described. (Note that if
  60 CONFIG_SMP is not defined, *no* domains are utilized and these lines
  61 will not appear in the output.)
  62
  63 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
  64
  65 The first field is a bit mask indicating what cpus this domain operates over.
  66
  67 The next 24 are a variety of load_balance() statistics in grouped into types
  68 of idleness (idle, busy, and newly idle):
  69
  70      1) # of times in this domain load_balance() was called when the
  71         cpu was idle
  72      2) # of times in this domain load_balance() checked but found
  73         the load did not require balancing when the cpu was idle
  74      3) # of times in this domain load_balance() tried to move one or
  75         more tasks and failed, when the cpu was idle
  76      4) sum of imbalances discovered (if any) with each call to
  77         load_balance() in this domain when the cpu was idle
  78      5) # of times in this domain pull_task() was called when the cpu
  79         was idle
  80      6) # of times in this domain pull_task() was called even though
  81         the target task was cache-hot when idle
  82      7) # of times in this domain load_balance() was called but did
  83         not find a busier queue while the cpu was idle
  84      8) # of times in this domain a busier queue was found while the
  85         cpu was idle but no busier group was found
  86
  87      9) # of times in this domain load_balance() was called when the
  88         cpu was busy
  89     10) # of times in this domain load_balance() checked but found the
  90         load did not require balancing when busy
  91     11) # of times in this domain load_balance() tried to move one or
  92         more tasks and failed, when the cpu was busy
  93     12) sum of imbalances discovered (if any) with each call to
  94         load_balance() in this domain when the cpu was busy
  95     13) # of times in this domain pull_task() was called when busy
  96     14) # of times in this domain pull_task() was called even though the
  97         target task was cache-hot when busy
  98     15) # of times in this domain load_balance() was called but did not
  99         find a busier queue while the cpu was busy
 100     16) # of times in this domain a busier queue was found while the cpu
 101         was busy but no busier group was found
 102
 103     17) # of times in this domain load_balance() was called when the
 104         cpu was just becoming idle
 105     18) # of times in this domain load_balance() checked but found the
 106         load did not require balancing when the cpu was just becoming idle
 107     19) # of times in this domain load_balance() tried to move one or more
 108         tasks and failed, when the cpu was just becoming idle
 109     20) sum of imbalances discovered (if any) with each call to
 110         load_balance() in this domain when the cpu was just becoming idle
 111     21) # of times in this domain pull_task() was called when newly idle
 112     22) # of times in this domain pull_task() was called even though the
 113         target task was cache-hot when just becoming idle
 114     23) # of times in this domain load_balance() was called but did not
 115         find a busier queue while the cpu was just becoming idle
 116     24) # of times in this domain a busier queue was found while the cpu
 117         was just becoming idle but no busier group was found
 118
 119    Next three are active_load_balance() statistics:
 120     25) # of times active_load_balance() was called
 121     26) # of times active_load_balance() tried to move a task and failed
 122     27) # of times active_load_balance() successfully moved a task
 123
 124    Next three are sched_balance_exec() statistics:
 125     28) sbe_cnt is not used
 126     29) sbe_balanced is not used
 127     30) sbe_pushed is not used
 128
 129    Next three are sched_balance_fork() statistics:
 130     31) sbf_cnt is not used
 131     32) sbf_balanced is not used
 132     33) sbf_pushed is not used
 133
 134    Next three are try_to_wake_up() statistics:
 135     34) # of times in this domain try_to_wake_up() awoke a task that
 136         last ran on a different cpu in this domain
 137     35) # of times in this domain try_to_wake_up() moved a task to the
 138         waking cpu because it was cache-cold on its own cpu anyway
 139     36) # of times in this domain try_to_wake_up() started passive balancing
 140
 141 /proc/<pid>/schedstat
 142 ----------------
 143 schedstats also adds a new /proc/<pid>/schedstat file to include some of
 144 the same information on a per-process level.  There are three fields in
 145 this file correlating for that process to:
 146      1) time spent on the cpu
 147      2) time spent waiting on a runqueue
 148      3) # of timeslices run on this cpu
 149
 150 A program could be easily written to make use of these extra fields to
 151 report on how well a particular process or set of processes is faring
 152 under the scheduler's policies.  A simple version of such a program is
 153 available at
 154     http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c