Version 10 of schedstats includes support for sched_domains, which hit the mainline kernel in 2.6.7. Some counters make more sense to be per-runqueue; other to be per-domain. Note that domains (and their associated information) will only be pertinent and available on machines utilizing CONFIG_SMP. In version 10 of schedstat, there is at least one level of domain statistics for each cpu listed, and there may well be more than one domain. Domains have no particular names in this implementation, but the highest numbered one typically arbitrates balancing across all the cpus on the machine, while domain0 is the most tightly focused domain, sometimes balancing only between pairs of cpus. At this time, there are no architectures which need more than three domain levels. The first field in the domain stats is a bit map indicating which cpus are affected by that domain. These fields are counters, and only increment. Programs which make use of these will need to start with a baseline observation and then calculate the change in the counters at each subsequent observation. A perl script which does this for many of the fields is available at http://eaglet.rain.com/rick/linux/schedstat/ Note that any such script will necessarily be version-specific, as the main reason to change versions is changes in the output format. For those wishing to write their own scripts, the fields are described here. CPU statistics -------------- cpu<N> 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 NOTE: In the sched_yield() statistics, the active queue is considered empty if it has only one process in it, since obviously the process calling sched_yield() is that process. First four fields are sched_yield() statistics: 1) # of times both the active and the expired queue were empty 2) # of times just the active queue was empty 3) # of times just the expired queue was empty 4) # of times sched_yield() was called Next four are schedule() statistics: 5) # of times the active queue had at least one other process on it 6) # of times we switched to the expired queue and reused it 7) # of times schedule() was called 8) # of times schedule() left the processor idle Next four are active_load_balance() statistics: 9) # of times active_load_balance() was called 10) # of times active_load_balance() caused this cpu to gain a task 11) # of times active_load_balance() caused this cpu to lose a task 12) # of times active_load_balance() tried to move a task and failed Next three are try_to_wake_up() statistics: 13) # of times try_to_wake_up() was called 14) # of times try_to_wake_up() successfully moved the awakening task 15) # of times try_to_wake_up() attempted to move the awakening task Next two are wake_up_new_task() statistics: 16) # of times wake_up_new_task() was called 17) # of times wake_up_new_task() successfully moved the new task Next one is a sched_migrate_task() statistic: 18) # of times sched_migrate_task() was called Next one is a sched_balance_exec() statistic: 19) # of times sched_balance_exec() was called Next three are statistics describing scheduling latency: 20) sum of all time spent running by tasks on this processor (in ms) 21) sum of all time spent waiting to run by tasks on this processor (in ms) 22) # of tasks (not necessarily unique) given to the processor The last six are statistics dealing with pull_task(): 23) # of times pull_task() moved a task to this cpu when newly idle 24) # of times pull_task() stole a task from this cpu when another cpu was newly idle 25) # of times pull_task() moved a task to this cpu when idle 26) # of times pull_task() stole a task from this cpu when another cpu was idle 27) # of times pull_task() moved a task to this cpu when busy 28) # of times pull_task() stole a task from this cpu when another cpu was busy Domain statistics ----------------- One of these is produced per domain for each cpu described. (Note that if CONFIG_SMP is not defined, *no* domains are utilized and these lines will not appear in the output.) domain<N> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 The first field is a bit mask indicating what cpus this domain operates over. The next fifteen are a variety of load_balance() statistics: 1) # of times in this domain load_balance() was called when the cpu was idle 2) # of times in this domain load_balance() was called when the cpu was busy 3) # of times in this domain load_balance() was called when the cpu was just becoming idle 4) # of times in this domain load_balance() tried to move one or more tasks and failed, when the cpu was idle 5) # of times in this domain load_balance() tried to move one or more tasks and failed, when the cpu was busy 6) # of times in this domain load_balance() tried to move one or more tasks and failed, when the cpu was just becoming idle 7) sum of imbalances discovered (if any) with each call to load_balance() in this domain when the cpu was idle 8) sum of imbalances discovered (if any) with each call to load_balance() in this domain when the cpu was busy 9) sum of imbalances discovered (if any) with each call to load_balance() in this domain when the cpu was just becoming idle 10) # of times in this domain load_balance() was called but did not find a busier queue while the cpu was idle 11) # of times in this domain load_balance() was called but did not find a busier queue while the cpu was busy 12) # of times in this domain load_balance() was called but did not find a busier queue while the cpu was just becoming idle 13) # of times in this domain a busier queue was found while the cpu was idle but no busier group was found 14) # of times in this domain a busier queue was found while the cpu was busy but no busier group was found 15) # of times in this domain a busier queue was found while the cpu was just becoming idle but no busier group was found Next two are sched_balance_exec() statistics: 17) # of times in this domain sched_balance_exec() successfully pushed a task to a new cpu 18) # of times in this domain sched_balance_exec() tried but failed to push a task to a new cpu Next two are try_to_wake_up() statistics: 19) # of times in this domain try_to_wake_up() tried to move a task based on affinity and cache warmth 20) # of times in this domain try_to_wake_up() tried to move a task based on load balancing /proc/<pid>/schedstat ---------------- schedstats also adds a new /proc/<pid/schedstat file to include some of the same information on a per-process level. There are three fields in this file correlating to fields 20, 21, and 22 in the CPU fields, but they only apply for that process. A program could be easily written to make use of these extra fields to report on how well a particular process or set of processes is faring under the scheduler's policies. A simple version of such a program is available at http://eaglet.rain.com/rick/linux/schedstat/v10/latency.c