Kyle Swenson | 8d8f654 | 2021-03-15 11:02:55 -0600 | [diff] [blame^] | 1 | Version 15 of schedstats dropped counters for some sched_yield: |
| 2 | yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is |
| 3 | identical to version 14. |
| 4 | |
| 5 | Version 14 of schedstats includes support for sched_domains, which hit the |
| 6 | mainline kernel in 2.6.20 although it is identical to the stats from version |
| 7 | 12 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel |
| 8 | release). Some counters make more sense to be per-runqueue; other to be |
| 9 | per-domain. Note that domains (and their associated information) will only |
| 10 | be pertinent and available on machines utilizing CONFIG_SMP. |
| 11 | |
| 12 | In version 14 of schedstat, there is at least one level of domain |
| 13 | statistics for each cpu listed, and there may well be more than one |
| 14 | domain. Domains have no particular names in this implementation, but |
| 15 | the highest numbered one typically arbitrates balancing across all the |
| 16 | cpus on the machine, while domain0 is the most tightly focused domain, |
| 17 | sometimes balancing only between pairs of cpus. At this time, there |
| 18 | are no architectures which need more than three domain levels. The first |
| 19 | field in the domain stats is a bit map indicating which cpus are affected |
| 20 | by that domain. |
| 21 | |
| 22 | These fields are counters, and only increment. Programs which make use |
| 23 | of these will need to start with a baseline observation and then calculate |
| 24 | the change in the counters at each subsequent observation. A perl script |
| 25 | which does this for many of the fields is available at |
| 26 | |
| 27 | http://eaglet.rain.com/rick/linux/schedstat/ |
| 28 | |
| 29 | Note that any such script will necessarily be version-specific, as the main |
| 30 | reason to change versions is changes in the output format. For those wishing |
| 31 | to write their own scripts, the fields are described here. |
| 32 | |
| 33 | CPU statistics |
| 34 | -------------- |
| 35 | cpu<N> 1 2 3 4 5 6 7 8 9 |
| 36 | |
| 37 | First field is a sched_yield() statistic: |
| 38 | 1) # of times sched_yield() was called |
| 39 | |
| 40 | Next 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 | |
| 46 | Next 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 | |
| 50 | Next 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 | |
| 57 | Domain statistics |
| 58 | ----------------- |
| 59 | One of these is produced per domain for each cpu described. (Note that if |
| 60 | CONFIG_SMP is not defined, *no* domains are utilized and these lines |
| 61 | will not appear in the output.) |
| 62 | |
| 63 | 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 |
| 64 | |
| 65 | The first field is a bit mask indicating what cpus this domain operates over. |
| 66 | |
| 67 | The next 24 are a variety of load_balance() statistics in grouped into types |
| 68 | of 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 | ---------------- |
| 143 | schedstats also adds a new /proc/<pid>/schedstat file to include some of |
| 144 | the same information on a per-process level. There are three fields in |
| 145 | this 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 | |
| 150 | A program could be easily written to make use of these extra fields to |
| 151 | report on how well a particular process or set of processes is faring |
| 152 | under the scheduler's policies. A simple version of such a program is |
| 153 | available at |
| 154 | http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c |