Documentation/sched-stats.txt

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