0001 ====================
0002 Scheduler Statistics
0003 ====================
0004
0005 Version 15 of schedstats dropped counters for some sched_yield:
0006 yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is
0007 identical to version 14.
0008
0009 Version 14 of schedstats includes support for sched_domains, which hit the
0010 mainline kernel in 2.6.20 although it is identical to the stats from version
0011 12 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
0012 release). Some counters make more sense to be per-runqueue; other to be
0013 per-domain. Note that domains (and their associated information) will only
0014 be pertinent and available on machines utilizing CONFIG_SMP.
0015
0016 In version 14 of schedstat, there is at least one level of domain
0017 statistics for each cpu listed, and there may well be more than one
0018 domain. Domains have no particular names in this implementation, but
0019 the highest numbered one typically arbitrates balancing across all the
0020 cpus on the machine, while domain0 is the most tightly focused domain,
0021 sometimes balancing only between pairs of cpus. At this time, there
0022 are no architectures which need more than three domain levels. The first
0023 field in the domain stats is a bit map indicating which cpus are affected
0024 by that domain.
0025
0026 These fields are counters, and only increment. Programs which make use
0027 of these will need to start with a baseline observation and then calculate
0028 the change in the counters at each subsequent observation. A perl script
0029 which does this for many of the fields is available at
0030
0031 http://eaglet.pdxhosts.com/rick/linux/schedstat/
0032
0033 Note that any such script will necessarily be version-specific, as the main
0034 reason to change versions is changes in the output format. For those wishing
0035 to write their own scripts, the fields are described here.
0036
0037 CPU statistics
0038 --------------
0039 cpu<N> 1 2 3 4 5 6 7 8 9
0040
0041 First field is a sched_yield() statistic:
0042
0043 1) # of times sched_yield() was called
0044
0045 Next three are schedule() statistics:
0046
0047 2) This field is a legacy array expiration count field used in the O(1)
0048 scheduler. We kept it for ABI compatibility, but it is always set to zero.
0049 3) # of times schedule() was called
0050 4) # of times schedule() left the processor idle
0051
0052 Next two are try_to_wake_up() statistics:
0053
0054 5) # of times try_to_wake_up() was called
0055 6) # of times try_to_wake_up() was called to wake up the local cpu
0056
0057 Next three are statistics describing scheduling latency:
0058
0059 7) sum of all time spent running by tasks on this processor (in nanoseconds)
0060 8) sum of all time spent waiting to run by tasks on this processor (in
0061 nanoseconds)
0062 9) # of timeslices run on this cpu
0063
0064
0065 Domain statistics
0066 -----------------
0067 One of these is produced per domain for each cpu described. (Note that if
0068 CONFIG_SMP is not defined, *no* domains are utilized and these lines
0069 will not appear in the output.)
0070
0071 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
0072
0073 The first field is a bit mask indicating what cpus this domain operates over.
0074
0075 The next 24 are a variety of load_balance() statistics in grouped into types
0076 of idleness (idle, busy, and newly idle):
0077
0078 1) # of times in this domain load_balance() was called when the
0079 cpu was idle
0080 2) # of times in this domain load_balance() checked but found
0081 the load did not require balancing when the cpu was idle
0082 3) # of times in this domain load_balance() tried to move one or
0083 more tasks and failed, when the cpu was idle
0084 4) sum of imbalances discovered (if any) with each call to
0085 load_balance() in this domain when the cpu was idle
0086 5) # of times in this domain pull_task() was called when the cpu
0087 was idle
0088 6) # of times in this domain pull_task() was called even though
0089 the target task was cache-hot when idle
0090 7) # of times in this domain load_balance() was called but did
0091 not find a busier queue while the cpu was idle
0092 8) # of times in this domain a busier queue was found while the
0093 cpu was idle but no busier group was found
0094 9) # of times in this domain load_balance() was called when the
0095 cpu was busy
0096 10) # of times in this domain load_balance() checked but found the
0097 load did not require balancing when busy
0098 11) # of times in this domain load_balance() tried to move one or
0099 more tasks and failed, when the cpu was busy
0100 12) sum of imbalances discovered (if any) with each call to
0101 load_balance() in this domain when the cpu was busy
0102 13) # of times in this domain pull_task() was called when busy
0103 14) # of times in this domain pull_task() was called even though the
0104 target task was cache-hot when busy
0105 15) # of times in this domain load_balance() was called but did not
0106 find a busier queue while the cpu was busy
0107 16) # of times in this domain a busier queue was found while the cpu
0108 was busy but no busier group was found
0109
0110 17) # of times in this domain load_balance() was called when the
0111 cpu was just becoming idle
0112 18) # of times in this domain load_balance() checked but found the
0113 load did not require balancing when the cpu was just becoming idle
0114 19) # of times in this domain load_balance() tried to move one or more
0115 tasks and failed, when the cpu was just becoming idle
0116 20) sum of imbalances discovered (if any) with each call to
0117 load_balance() in this domain when the cpu was just becoming idle
0118 21) # of times in this domain pull_task() was called when newly idle
0119 22) # of times in this domain pull_task() was called even though the
0120 target task was cache-hot when just becoming idle
0121 23) # of times in this domain load_balance() was called but did not
0122 find a busier queue while the cpu was just becoming idle
0123 24) # of times in this domain a busier queue was found while the cpu
0124 was just becoming idle but no busier group was found
0125
0126 Next three are active_load_balance() statistics:
0127
0128 25) # of times active_load_balance() was called
0129 26) # of times active_load_balance() tried to move a task and failed
0130 27) # of times active_load_balance() successfully moved a task
0131
0132 Next three are sched_balance_exec() statistics:
0133
0134 28) sbe_cnt is not used
0135 29) sbe_balanced is not used
0136 30) sbe_pushed is not used
0137
0138 Next three are sched_balance_fork() statistics:
0139
0140 31) sbf_cnt is not used
0141 32) sbf_balanced is not used
0142 33) sbf_pushed is not used
0143
0144 Next three are try_to_wake_up() statistics:
0145
0146 34) # of times in this domain try_to_wake_up() awoke a task that
0147 last ran on a different cpu in this domain
0148 35) # of times in this domain try_to_wake_up() moved a task to the
0149 waking cpu because it was cache-cold on its own cpu anyway
0150 36) # of times in this domain try_to_wake_up() started passive balancing
0151
0152 /proc/<pid>/schedstat
0153 ---------------------
0154 schedstats also adds a new /proc/<pid>/schedstat file to include some of
0155 the same information on a per-process level. There are three fields in
0156 this file correlating for that process to:
0157
0158 1) time spent on the cpu (in nanoseconds)
0159 2) time spent waiting on a runqueue (in nanoseconds)
0160 3) # of timeslices run on this cpu
0161
0162 A program could be easily written to make use of these extra fields to
0163 report on how well a particular process or set of processes is faring
0164 under the scheduler's policies. A simple version of such a program is
0165 available at
0166
0167 http://eaglet.pdxhosts.com/rick/linux/schedstat/v12/latency.c
