173 lines · plain
1====================2Scheduler Statistics3====================4 5Version 16 of schedstats changed the order of definitions within6'enum cpu_idle_type', which changed the order of [CPU_MAX_IDLE_TYPES]7columns in show_schedstat(). In particular the position of CPU_IDLE8and __CPU_NOT_IDLE changed places. The size of the array is unchanged.9 10Version 15 of schedstats dropped counters for some sched_yield:11yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is12identical to version 14.13 14Version 14 of schedstats includes support for sched_domains, which hit the15mainline kernel in 2.6.20 although it is identical to the stats from version1612 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel17release). Some counters make more sense to be per-runqueue; other to be18per-domain. Note that domains (and their associated information) will only19be pertinent and available on machines utilizing CONFIG_SMP.20 21In version 14 of schedstat, there is at least one level of domain22statistics for each cpu listed, and there may well be more than one23domain. Domains have no particular names in this implementation, but24the highest numbered one typically arbitrates balancing across all the25cpus on the machine, while domain0 is the most tightly focused domain,26sometimes balancing only between pairs of cpus. At this time, there27are no architectures which need more than three domain levels. The first28field in the domain stats is a bit map indicating which cpus are affected29by that domain.30 31These fields are counters, and only increment. Programs which make use32of these will need to start with a baseline observation and then calculate33the change in the counters at each subsequent observation. A perl script34which does this for many of the fields is available at35 36 http://eaglet.pdxhosts.com/rick/linux/schedstat/37 38Note that any such script will necessarily be version-specific, as the main39reason to change versions is changes in the output format. For those wishing40to write their own scripts, the fields are described here.41 42CPU statistics43--------------44cpu<N> 1 2 3 4 5 6 7 8 945 46First field is a sched_yield() statistic:47 48 1) # of times sched_yield() was called49 50Next three are schedule() statistics:51 52 2) This field is a legacy array expiration count field used in the O(1)53 scheduler. We kept it for ABI compatibility, but it is always set to zero.54 3) # of times schedule() was called55 4) # of times schedule() left the processor idle56 57Next two are try_to_wake_up() statistics:58 59 5) # of times try_to_wake_up() was called60 6) # of times try_to_wake_up() was called to wake up the local cpu61 62Next three are statistics describing scheduling latency:63 64 7) sum of all time spent running by tasks on this processor (in nanoseconds)65 8) sum of all time spent waiting to run by tasks on this processor (in66 nanoseconds)67 9) # of timeslices run on this cpu68 69 70Domain statistics71-----------------72One of these is produced per domain for each cpu described. (Note that if73CONFIG_SMP is not defined, *no* domains are utilized and these lines74will not appear in the output.)75 76domain<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 3677 78The first field is a bit mask indicating what cpus this domain operates over.79 80The next 24 are a variety of sched_balance_rq() statistics in grouped into types81of idleness (idle, busy, and newly idle):82 83 1) # of times in this domain sched_balance_rq() was called when the84 cpu was idle85 2) # of times in this domain sched_balance_rq() checked but found86 the load did not require balancing when the cpu was idle87 3) # of times in this domain sched_balance_rq() tried to move one or88 more tasks and failed, when the cpu was idle89 4) sum of imbalances discovered (if any) with each call to90 sched_balance_rq() in this domain when the cpu was idle91 5) # of times in this domain pull_task() was called when the cpu92 was idle93 6) # of times in this domain pull_task() was called even though94 the target task was cache-hot when idle95 7) # of times in this domain sched_balance_rq() was called but did96 not find a busier queue while the cpu was idle97 8) # of times in this domain a busier queue was found while the98 cpu was idle but no busier group was found99 9) # of times in this domain sched_balance_rq() was called when the100 cpu was busy101 10) # of times in this domain sched_balance_rq() checked but found the102 load did not require balancing when busy103 11) # of times in this domain sched_balance_rq() tried to move one or104 more tasks and failed, when the cpu was busy105 12) sum of imbalances discovered (if any) with each call to106 sched_balance_rq() in this domain when the cpu was busy107 13) # of times in this domain pull_task() was called when busy108 14) # of times in this domain pull_task() was called even though the109 target task was cache-hot when busy110 15) # of times in this domain sched_balance_rq() was called but did not111 find a busier queue while the cpu was busy112 16) # of times in this domain a busier queue was found while the cpu113 was busy but no busier group was found114 115 17) # of times in this domain sched_balance_rq() was called when the116 cpu was just becoming idle117 18) # of times in this domain sched_balance_rq() checked but found the118 load did not require balancing when the cpu was just becoming idle119 19) # of times in this domain sched_balance_rq() tried to move one or more120 tasks and failed, when the cpu was just becoming idle121 20) sum of imbalances discovered (if any) with each call to122 sched_balance_rq() in this domain when the cpu was just becoming idle123 21) # of times in this domain pull_task() was called when newly idle124 22) # of times in this domain pull_task() was called even though the125 target task was cache-hot when just becoming idle126 23) # of times in this domain sched_balance_rq() was called but did not127 find a busier queue while the cpu was just becoming idle128 24) # of times in this domain a busier queue was found while the cpu129 was just becoming idle but no busier group was found130 131 Next three are active_load_balance() statistics:132 133 25) # of times active_load_balance() was called134 26) # of times active_load_balance() tried to move a task and failed135 27) # of times active_load_balance() successfully moved a task136 137 Next three are sched_balance_exec() statistics:138 139 28) sbe_cnt is not used140 29) sbe_balanced is not used141 30) sbe_pushed is not used142 143 Next three are sched_balance_fork() statistics:144 145 31) sbf_cnt is not used146 32) sbf_balanced is not used147 33) sbf_pushed is not used148 149 Next three are try_to_wake_up() statistics:150 151 34) # of times in this domain try_to_wake_up() awoke a task that152 last ran on a different cpu in this domain153 35) # of times in this domain try_to_wake_up() moved a task to the154 waking cpu because it was cache-cold on its own cpu anyway155 36) # of times in this domain try_to_wake_up() started passive balancing156 157/proc/<pid>/schedstat158---------------------159schedstats also adds a new /proc/<pid>/schedstat file to include some of160the same information on a per-process level. There are three fields in161this file correlating for that process to:162 163 1) time spent on the cpu (in nanoseconds)164 2) time spent waiting on a runqueue (in nanoseconds)165 3) # of timeslices run on this cpu166 167A program could be easily written to make use of these extra fields to168report on how well a particular process or set of processes is faring169under the scheduler's policies. A simple version of such a program is170available at171 172 http://eaglet.pdxhosts.com/rick/linux/schedstat/v12/latency.c173