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) 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 46Next 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 50Next 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 57Domain statistics 58----------------- 59One of these is produced per domain for each cpu described. (Note that if 60CONFIG_SMP is not defined, *no* domains are utilized and these lines 61will not appear in the output.) 62 63domain<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 65The first field is a bit mask indicating what cpus this domain operates over. 66 67The next 24 are a variety of load_balance() statistics in grouped into types 68of 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---------------- 143schedstats also adds a new /proc/<pid>/schedstat file to include some of 144the same information on a per-process level. There are three fields in 145this 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 150A program could be easily written to make use of these extra fields to 151report on how well a particular process or set of processes is faring 152under the scheduler's policies. A simple version of such a program is 153available at 154 http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c 155