1==================== 2Scheduler Statistics 3==================== 4 5Version 15 of schedstats dropped counters for some sched_yield: 6yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is 7identical to version 14. 8 9Version 14 of schedstats includes support for sched_domains, which hit the 10mainline kernel in 2.6.20 although it is identical to the stats from version 1112 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel 12release). Some counters make more sense to be per-runqueue; other to be 13per-domain. Note that domains (and their associated information) will only 14be pertinent and available on machines utilizing CONFIG_SMP. 15 16In version 14 of schedstat, there is at least one level of domain 17statistics for each cpu listed, and there may well be more than one 18domain. Domains have no particular names in this implementation, but 19the highest numbered one typically arbitrates balancing across all the 20cpus on the machine, while domain0 is the most tightly focused domain, 21sometimes balancing only between pairs of cpus. At this time, there 22are no architectures which need more than three domain levels. The first 23field in the domain stats is a bit map indicating which cpus are affected 24by that domain. 25 26These fields are counters, and only increment. Programs which make use 27of these will need to start with a baseline observation and then calculate 28the change in the counters at each subsequent observation. A perl script 29which does this for many of the fields is available at 30 31 http://eaglet.pdxhosts.com/rick/linux/schedstat/ 32 33Note that any such script will necessarily be version-specific, as the main 34reason to change versions is changes in the output format. For those wishing 35to write their own scripts, the fields are described here. 36 37CPU statistics 38-------------- 39cpu<N> 1 2 3 4 5 6 7 8 9 40 41First field is a sched_yield() statistic: 42 43 1) # of times sched_yield() was called 44 45Next three are schedule() statistics: 46 47 2) This field is a legacy array expiration count field used in the O(1) 48 scheduler. We kept it for ABI compatibility, but it is always set to zero. 49 3) # of times schedule() was called 50 4) # of times schedule() left the processor idle 51 52Next two are try_to_wake_up() statistics: 53 54 5) # of times try_to_wake_up() was called 55 6) # of times try_to_wake_up() was called to wake up the local cpu 56 57Next three are statistics describing scheduling latency: 58 59 7) sum of all time spent running by tasks on this processor (in nanoseconds) 60 8) sum of all time spent waiting to run by tasks on this processor (in 61 nanoseconds) 62 9) # of timeslices run on this cpu 63 64 65Domain statistics 66----------------- 67One of these is produced per domain for each cpu described. (Note that if 68CONFIG_SMP is not defined, *no* domains are utilized and these lines 69will not appear in the output.) 70 71domain<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 72 73The first field is a bit mask indicating what cpus this domain operates over. 74 75The next 24 are a variety of load_balance() statistics in grouped into types 76of idleness (idle, busy, and newly idle): 77 78 1) # of times in this domain load_balance() was called when the 79 cpu was idle 80 2) # of times in this domain load_balance() checked but found 81 the load did not require balancing when the cpu was idle 82 3) # of times in this domain load_balance() tried to move one or 83 more tasks and failed, when the cpu was idle 84 4) sum of imbalances discovered (if any) with each call to 85 load_balance() in this domain when the cpu was idle 86 5) # of times in this domain pull_task() was called when the cpu 87 was idle 88 6) # of times in this domain pull_task() was called even though 89 the target task was cache-hot when idle 90 7) # of times in this domain load_balance() was called but did 91 not find a busier queue while the cpu was idle 92 8) # of times in this domain a busier queue was found while the 93 cpu was idle but no busier group was found 94 9) # of times in this domain load_balance() was called when the 95 cpu was busy 96 10) # of times in this domain load_balance() checked but found the 97 load did not require balancing when busy 98 11) # of times in this domain load_balance() tried to move one or 99 more tasks and failed, when the cpu was busy 100 12) sum of imbalances discovered (if any) with each call to 101 load_balance() in this domain when the cpu was busy 102 13) # of times in this domain pull_task() was called when busy 103 14) # of times in this domain pull_task() was called even though the 104 target task was cache-hot when busy 105 15) # of times in this domain load_balance() was called but did not 106 find a busier queue while the cpu was busy 107 16) # of times in this domain a busier queue was found while the cpu 108 was busy but no busier group was found 109 110 17) # of times in this domain load_balance() was called when the 111 cpu was just becoming idle 112 18) # of times in this domain load_balance() checked but found the 113 load did not require balancing when the cpu was just becoming idle 114 19) # of times in this domain load_balance() tried to move one or more 115 tasks and failed, when the cpu was just becoming idle 116 20) sum of imbalances discovered (if any) with each call to 117 load_balance() in this domain when the cpu was just becoming idle 118 21) # of times in this domain pull_task() was called when newly idle 119 22) # of times in this domain pull_task() was called even though the 120 target task was cache-hot when just becoming idle 121 23) # of times in this domain load_balance() was called but did not 122 find a busier queue while the cpu was just becoming idle 123 24) # of times in this domain a busier queue was found while the cpu 124 was just becoming idle but no busier group was found 125 126 Next three are active_load_balance() statistics: 127 128 25) # of times active_load_balance() was called 129 26) # of times active_load_balance() tried to move a task and failed 130 27) # of times active_load_balance() successfully moved a task 131 132 Next three are sched_balance_exec() statistics: 133 134 28) sbe_cnt is not used 135 29) sbe_balanced is not used 136 30) sbe_pushed is not used 137 138 Next three are sched_balance_fork() statistics: 139 140 31) sbf_cnt is not used 141 32) sbf_balanced is not used 142 33) sbf_pushed is not used 143 144 Next three are try_to_wake_up() statistics: 145 146 34) # of times in this domain try_to_wake_up() awoke a task that 147 last ran on a different cpu in this domain 148 35) # of times in this domain try_to_wake_up() moved a task to the 149 waking cpu because it was cache-cold on its own cpu anyway 150 36) # of times in this domain try_to_wake_up() started passive balancing 151 152/proc/<pid>/schedstat 153--------------------- 154schedstats also adds a new /proc/<pid>/schedstat file to include some of 155the same information on a per-process level. There are three fields in 156this file correlating for that process to: 157 158 1) time spent on the cpu (in nanoseconds) 159 2) time spent waiting on a runqueue (in nanoseconds) 160 3) # of timeslices run on this cpu 161 162A program could be easily written to make use of these extra fields to 163report on how well a particular process or set of processes is faring 164under the scheduler's policies. A simple version of such a program is 165available at 166 167 http://eaglet.pdxhosts.com/rick/linux/schedstat/v12/latency.c 168