1CFS Bandwidth Control 2===================== 3 4[ This document only discusses CPU bandwidth control for SCHED_NORMAL. 5 The SCHED_RT case is covered in Documentation/scheduler/sched-rt-group.txt ] 6 7CFS bandwidth control is a CONFIG_FAIR_GROUP_SCHED extension which allows the 8specification of the maximum CPU bandwidth available to a group or hierarchy. 9 10The bandwidth allowed for a group is specified using a quota and period. Within 11each given "period" (microseconds), a group is allowed to consume only up to 12"quota" microseconds of CPU time. When the CPU bandwidth consumption of a 13group exceeds this limit (for that period), the tasks belonging to its 14hierarchy will be throttled and are not allowed to run again until the next 15period. 16 17A group's unused runtime is globally tracked, being refreshed with quota units 18above at each period boundary. As threads consume this bandwidth it is 19transferred to cpu-local "silos" on a demand basis. The amount transferred 20within each of these updates is tunable and described as the "slice". 21 22Management 23---------- 24Quota and period are managed within the cpu subsystem via cgroupfs. 25 26cpu.cfs_quota_us: the total available run-time within a period (in microseconds) 27cpu.cfs_period_us: the length of a period (in microseconds) 28cpu.stat: exports throttling statistics [explained further below] 29 30The default values are: 31 cpu.cfs_period_us=100ms 32 cpu.cfs_quota=-1 33 34A value of -1 for cpu.cfs_quota_us indicates that the group does not have any 35bandwidth restriction in place, such a group is described as an unconstrained 36bandwidth group. This represents the traditional work-conserving behavior for 37CFS. 38 39Writing any (valid) positive value(s) will enact the specified bandwidth limit. 40The minimum quota allowed for the quota or period is 1ms. There is also an 41upper bound on the period length of 1s. Additional restrictions exist when 42bandwidth limits are used in a hierarchical fashion, these are explained in 43more detail below. 44 45Writing any negative value to cpu.cfs_quota_us will remove the bandwidth limit 46and return the group to an unconstrained state once more. 47 48Any updates to a group's bandwidth specification will result in it becoming 49unthrottled if it is in a constrained state. 50 51System wide settings 52-------------------- 53For efficiency run-time is transferred between the global pool and CPU local 54"silos" in a batch fashion. This greatly reduces global accounting pressure 55on large systems. The amount transferred each time such an update is required 56is described as the "slice". 57 58This is tunable via procfs: 59 /proc/sys/kernel/sched_cfs_bandwidth_slice_us (default=5ms) 60 61Larger slice values will reduce transfer overheads, while smaller values allow 62for more fine-grained consumption. 63 64Statistics 65---------- 66A group's bandwidth statistics are exported via 3 fields in cpu.stat. 67 68cpu.stat: 69- nr_periods: Number of enforcement intervals that have elapsed. 70- nr_throttled: Number of times the group has been throttled/limited. 71- throttled_time: The total time duration (in nanoseconds) for which entities 72 of the group have been throttled. 73 74This interface is read-only. 75 76Hierarchical considerations 77--------------------------- 78The interface enforces that an individual entity's bandwidth is always 79attainable, that is: max(c_i) <= C. However, over-subscription in the 80aggregate case is explicitly allowed to enable work-conserving semantics 81within a hierarchy. 82 e.g. \Sum (c_i) may exceed C 83[ Where C is the parent's bandwidth, and c_i its children ] 84 85 86There are two ways in which a group may become throttled: 87 a. it fully consumes its own quota within a period 88 b. a parent's quota is fully consumed within its period 89 90In case b) above, even though the child may have runtime remaining it will not 91be allowed to until the parent's runtime is refreshed. 92 93CFS Bandwidth Quota Caveats 94--------------------------- 95Once a slice is assigned to a cpu it does not expire. However all but 1ms of 96the slice may be returned to the global pool if all threads on that cpu become 97unrunnable. This is configured at compile time by the min_cfs_rq_runtime 98variable. This is a performance tweak that helps prevent added contention on 99the global lock. 100 101The fact that cpu-local slices do not expire results in some interesting corner 102cases that should be understood. 103 104For cgroup cpu constrained applications that are cpu limited this is a 105relatively moot point because they will naturally consume the entirety of their 106quota as well as the entirety of each cpu-local slice in each period. As a 107result it is expected that nr_periods roughly equal nr_throttled, and that 108cpuacct.usage will increase roughly equal to cfs_quota_us in each period. 109 110For highly-threaded, non-cpu bound applications this non-expiration nuance 111allows applications to briefly burst past their quota limits by the amount of 112unused slice on each cpu that the task group is running on (typically at most 1131ms per cpu or as defined by min_cfs_rq_runtime). This slight burst only 114applies if quota had been assigned to a cpu and then not fully used or returned 115in previous periods. This burst amount will not be transferred between cores. 116As a result, this mechanism still strictly limits the task group to quota 117average usage, albeit over a longer time window than a single period. This 118also limits the burst ability to no more than 1ms per cpu. This provides 119better more predictable user experience for highly threaded applications with 120small quota limits on high core count machines. It also eliminates the 121propensity to throttle these applications while simultanously using less than 122quota amounts of cpu. Another way to say this, is that by allowing the unused 123portion of a slice to remain valid across periods we have decreased the 124possibility of wastefully expiring quota on cpu-local silos that don't need a 125full slice's amount of cpu time. 126 127The interaction between cpu-bound and non-cpu-bound-interactive applications 128should also be considered, especially when single core usage hits 100%. If you 129gave each of these applications half of a cpu-core and they both got scheduled 130on the same CPU it is theoretically possible that the non-cpu bound application 131will use up to 1ms additional quota in some periods, thereby preventing the 132cpu-bound application from fully using its quota by that same amount. In these 133instances it will be up to the CFS algorithm (see sched-design-CFS.rst) to 134decide which application is chosen to run, as they will both be runnable and 135have remaining quota. This runtime discrepancy will be made up in the following 136periods when the interactive application idles. 137 138Examples 139-------- 1401. Limit a group to 1 CPU worth of runtime. 141 142 If period is 250ms and quota is also 250ms, the group will get 143 1 CPU worth of runtime every 250ms. 144 145 # echo 250000 > cpu.cfs_quota_us /* quota = 250ms */ 146 # echo 250000 > cpu.cfs_period_us /* period = 250ms */ 147 1482. Limit a group to 2 CPUs worth of runtime on a multi-CPU machine. 149 150 With 500ms period and 1000ms quota, the group can get 2 CPUs worth of 151 runtime every 500ms. 152 153 # echo 1000000 > cpu.cfs_quota_us /* quota = 1000ms */ 154 # echo 500000 > cpu.cfs_period_us /* period = 500ms */ 155 156 The larger period here allows for increased burst capacity. 157 1583. Limit a group to 20% of 1 CPU. 159 160 With 50ms period, 10ms quota will be equivalent to 20% of 1 CPU. 161 162 # echo 10000 > cpu.cfs_quota_us /* quota = 10ms */ 163 # echo 50000 > cpu.cfs_period_us /* period = 50ms */ 164 165 By using a small period here we are ensuring a consistent latency 166 response at the expense of burst capacity. 167 168