1===================== 2CFS Bandwidth Control 3===================== 4 5[ This document only discusses CPU bandwidth control for SCHED_NORMAL. 6 The SCHED_RT case is covered in Documentation/scheduler/sched-rt-group.rst ] 7 8CFS bandwidth control is a CONFIG_FAIR_GROUP_SCHED extension which allows the 9specification of the maximum CPU bandwidth available to a group or hierarchy. 10 11The bandwidth allowed for a group is specified using a quota and period. Within 12each given "period" (microseconds), a task group is allocated up to "quota" 13microseconds of CPU time. That quota is assigned to per-cpu run queues in 14slices as threads in the cgroup become runnable. Once all quota has been 15assigned any additional requests for quota will result in those threads being 16throttled. Throttled threads will not be able to run again until the next 17period when the quota is replenished. 18 19A group's unassigned quota is globally tracked, being refreshed back to 20cfs_quota units at each period boundary. As threads consume this bandwidth it 21is transferred to cpu-local "silos" on a demand basis. The amount transferred 22within each of these updates is tunable and described as the "slice". 23 24Management 25---------- 26Quota and period are managed within the cpu subsystem via cgroupfs. 27 28cpu.cfs_quota_us: the total available run-time within a period (in microseconds) 29cpu.cfs_period_us: the length of a period (in microseconds) 30cpu.stat: exports throttling statistics [explained further below] 31 32The default values are:: 33 34 cpu.cfs_period_us=100ms 35 cpu.cfs_quota=-1 36 37A value of -1 for cpu.cfs_quota_us indicates that the group does not have any 38bandwidth restriction in place, such a group is described as an unconstrained 39bandwidth group. This represents the traditional work-conserving behavior for 40CFS. 41 42Writing any (valid) positive value(s) will enact the specified bandwidth limit. 43The minimum quota allowed for the quota or period is 1ms. There is also an 44upper bound on the period length of 1s. Additional restrictions exist when 45bandwidth limits are used in a hierarchical fashion, these are explained in 46more detail below. 47 48Writing any negative value to cpu.cfs_quota_us will remove the bandwidth limit 49and return the group to an unconstrained state once more. 50 51Any updates to a group's bandwidth specification will result in it becoming 52unthrottled if it is in a constrained state. 53 54System wide settings 55-------------------- 56For efficiency run-time is transferred between the global pool and CPU local 57"silos" in a batch fashion. This greatly reduces global accounting pressure 58on large systems. The amount transferred each time such an update is required 59is described as the "slice". 60 61This is tunable via procfs:: 62 63 /proc/sys/kernel/sched_cfs_bandwidth_slice_us (default=5ms) 64 65Larger slice values will reduce transfer overheads, while smaller values allow 66for more fine-grained consumption. 67 68Statistics 69---------- 70A group's bandwidth statistics are exported via 3 fields in cpu.stat. 71 72cpu.stat: 73 74- nr_periods: Number of enforcement intervals that have elapsed. 75- nr_throttled: Number of times the group has been throttled/limited. 76- throttled_time: The total time duration (in nanoseconds) for which entities 77 of the group have been throttled. 78 79This interface is read-only. 80 81Hierarchical considerations 82--------------------------- 83The interface enforces that an individual entity's bandwidth is always 84attainable, that is: max(c_i) <= C. However, over-subscription in the 85aggregate case is explicitly allowed to enable work-conserving semantics 86within a hierarchy: 87 88 e.g. \Sum (c_i) may exceed C 89 90[ Where C is the parent's bandwidth, and c_i its children ] 91 92 93There are two ways in which a group may become throttled: 94 95 a. it fully consumes its own quota within a period 96 b. a parent's quota is fully consumed within its period 97 98In case b) above, even though the child may have runtime remaining it will not 99be allowed to until the parent's runtime is refreshed. 100 101CFS Bandwidth Quota Caveats 102--------------------------- 103Once a slice is assigned to a cpu it does not expire. However all but 1ms of 104the slice may be returned to the global pool if all threads on that cpu become 105unrunnable. This is configured at compile time by the min_cfs_rq_runtime 106variable. This is a performance tweak that helps prevent added contention on 107the global lock. 108 109The fact that cpu-local slices do not expire results in some interesting corner 110cases that should be understood. 111 112For cgroup cpu constrained applications that are cpu limited this is a 113relatively moot point because they will naturally consume the entirety of their 114quota as well as the entirety of each cpu-local slice in each period. As a 115result it is expected that nr_periods roughly equal nr_throttled, and that 116cpuacct.usage will increase roughly equal to cfs_quota_us in each period. 117 118For highly-threaded, non-cpu bound applications this non-expiration nuance 119allows applications to briefly burst past their quota limits by the amount of 120unused slice on each cpu that the task group is running on (typically at most 1211ms per cpu or as defined by min_cfs_rq_runtime). This slight burst only 122applies if quota had been assigned to a cpu and then not fully used or returned 123in previous periods. This burst amount will not be transferred between cores. 124As a result, this mechanism still strictly limits the task group to quota 125average usage, albeit over a longer time window than a single period. This 126also limits the burst ability to no more than 1ms per cpu. This provides 127better more predictable user experience for highly threaded applications with 128small quota limits on high core count machines. It also eliminates the 129propensity to throttle these applications while simultanously using less than 130quota amounts of cpu. Another way to say this, is that by allowing the unused 131portion of a slice to remain valid across periods we have decreased the 132possibility of wastefully expiring quota on cpu-local silos that don't need a 133full slice's amount of cpu time. 134 135The interaction between cpu-bound and non-cpu-bound-interactive applications 136should also be considered, especially when single core usage hits 100%. If you 137gave each of these applications half of a cpu-core and they both got scheduled 138on the same CPU it is theoretically possible that the non-cpu bound application 139will use up to 1ms additional quota in some periods, thereby preventing the 140cpu-bound application from fully using its quota by that same amount. In these 141instances it will be up to the CFS algorithm (see sched-design-CFS.rst) to 142decide which application is chosen to run, as they will both be runnable and 143have remaining quota. This runtime discrepancy will be made up in the following 144periods when the interactive application idles. 145 146Examples 147-------- 1481. Limit a group to 1 CPU worth of runtime:: 149 150 If period is 250ms and quota is also 250ms, the group will get 151 1 CPU worth of runtime every 250ms. 152 153 # echo 250000 > cpu.cfs_quota_us /* quota = 250ms */ 154 # echo 250000 > cpu.cfs_period_us /* period = 250ms */ 155 1562. Limit a group to 2 CPUs worth of runtime on a multi-CPU machine 157 158 With 500ms period and 1000ms quota, the group can get 2 CPUs worth of 159 runtime every 500ms:: 160 161 # echo 1000000 > cpu.cfs_quota_us /* quota = 1000ms */ 162 # echo 500000 > cpu.cfs_period_us /* period = 500ms */ 163 164 The larger period here allows for increased burst capacity. 165 1663. Limit a group to 20% of 1 CPU. 167 168 With 50ms period, 10ms quota will be equivalent to 20% of 1 CPU:: 169 170 # echo 10000 > cpu.cfs_quota_us /* quota = 10ms */ 171 # echo 50000 > cpu.cfs_period_us /* period = 50ms */ 172 173 By using a small period here we are ensuring a consistent latency 174 response at the expense of burst capacity. 175