1Version 15 of schedstats dropped counters for some sched_yield: 2yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is 3identical to version 14. 4 5Version 14 of schedstats includes support for sched_domains, which hit the 6mainline kernel in 2.6.20 although it is identical to the stats from version 712 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel 8release). Some counters make more sense to be per-runqueue; other to be 9per-domain. Note that domains (and their associated information) will only 10be pertinent and available on machines utilizing CONFIG_SMP. 11 12In version 14 of schedstat, there is at least one level of domain 13statistics for each cpu listed, and there may well be more than one 14domain. Domains have no particular names in this implementation, but 15the highest numbered one typically arbitrates balancing across all the 16cpus on the machine, while domain0 is the most tightly focused domain, 17sometimes balancing only between pairs of cpus. At this time, there 18are no architectures which need more than three domain levels. The first 19field in the domain stats is a bit map indicating which cpus are affected 20by that domain. 21 22These fields are counters, and only increment. Programs which make use 23of these will need to start with a baseline observation and then calculate 24the change in the counters at each subsequent observation. A perl script 25which does this for many of the fields is available at 26 27 http://eaglet.rain.com/rick/linux/schedstat/ 28 29Note that any such script will necessarily be version-specific, as the main 30reason to change versions is changes in the output format. For those wishing 31to write their own scripts, the fields are described here. 32 33CPU statistics 34-------------- 35cpu<N> 1 2 3 4 5 6 7 8 9 36 37First field is a sched_yield() statistic: 38 1) # of times sched_yield() was called 39 40Next 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 45Next 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 49Next 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 56Domain statistics 57----------------- 58One of these is produced per domain for each cpu described. (Note that if 59CONFIG_SMP is not defined, *no* domains are utilized and these lines 60will not appear in the output.) 61 62domain<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 64The first field is a bit mask indicating what cpus this domain operates over. 65 66The next 24 are a variety of load_balance() statistics in grouped into types 67of 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---------------- 142schedstats also adds a new /proc/<pid>/schedstat file to include some of 143the same information on a per-process level. There are three fields in 144this 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 149A program could be easily written to make use of these extra fields to 150report on how well a particular process or set of processes is faring 151under the scheduler's policies. A simple version of such a program is 152available at 153 http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c 154