1=====================
2CFS Bandwidth Control
3=====================
4
5.. note::
6   This document only discusses CPU bandwidth control for SCHED_NORMAL.
7   The SCHED_RT case is covered in Documentation/scheduler/sched-rt-group.rst
8
9CFS bandwidth control is a CONFIG_FAIR_GROUP_SCHED extension which allows the
10specification of the maximum CPU bandwidth available to a group or hierarchy.
11
12The bandwidth allowed for a group is specified using a quota and period. Within
13each given "period" (microseconds), a task group is allocated up to "quota"
14microseconds of CPU time. That quota is assigned to per-cpu run queues in
15slices as threads in the cgroup become runnable. Once all quota has been
16assigned any additional requests for quota will result in those threads being
17throttled. Throttled threads will not be able to run again until the next
18period when the quota is replenished.
19
20A group's unassigned quota is globally tracked, being refreshed back to
21cfs_quota units at each period boundary. As threads consume this bandwidth it
22is transferred to cpu-local "silos" on a demand basis. The amount transferred
23within each of these updates is tunable and described as the "slice".
24
25Burst feature
26-------------
27This feature borrows time now against our future underrun, at the cost of
28increased interference against the other system users. All nicely bounded.
29
30Traditional (UP-EDF) bandwidth control is something like:
31
32  (U = \Sum u_i) <= 1
33
34This guaranteeds both that every deadline is met and that the system is
35stable. After all, if U were > 1, then for every second of walltime,
36we'd have to run more than a second of program time, and obviously miss
37our deadline, but the next deadline will be further out still, there is
38never time to catch up, unbounded fail.
39
40The burst feature observes that a workload doesn't always executes the full
41quota; this enables one to describe u_i as a statistical distribution.
42
43For example, have u_i = {x,e}_i, where x is the p(95) and x+e p(100)
44(the traditional WCET). This effectively allows u to be smaller,
45increasing the efficiency (we can pack more tasks in the system), but at
46the cost of missing deadlines when all the odds line up. However, it
47does maintain stability, since every overrun must be paired with an
48underrun as long as our x is above the average.
49
50That is, suppose we have 2 tasks, both specify a p(95) value, then we
51have a p(95)*p(95) = 90.25% chance both tasks are within their quota and
52everything is good. At the same time we have a p(5)p(5) = 0.25% chance
53both tasks will exceed their quota at the same time (guaranteed deadline
54fail). Somewhere in between there's a threshold where one exceeds and
55the other doesn't underrun enough to compensate; this depends on the
56specific CDFs.
57
58At the same time, we can say that the worst case deadline miss, will be
59\Sum e_i; that is, there is a bounded tardiness (under the assumption
60that x+e is indeed WCET).
61
62The interferenece when using burst is valued by the possibilities for
63missing the deadline and the average WCET. Test results showed that when
64there many cgroups or CPU is under utilized, the interference is
65limited. More details are shown in:
66https://lore.kernel.org/lkml/5371BD36-55AE-4F71-B9D7-B86DC32E3D2B@linux.alibaba.com/
67
68Management
69----------
70Quota, period and burst are managed within the cpu subsystem via cgroupfs.
71
72.. note::
73   The cgroupfs files described in this section are only applicable
74   to cgroup v1. For cgroup v2, see
75   :ref:`Documentation/admin-guide/cgroup-v2.rst <cgroup-v2-cpu>`.
76
77- cpu.cfs_quota_us: run-time replenished within a period (in microseconds)
78- cpu.cfs_period_us: the length of a period (in microseconds)
79- cpu.stat: exports throttling statistics [explained further below]
80- cpu.cfs_burst_us: the maximum accumulated run-time (in microseconds)
81
82The default values are::
83
84	cpu.cfs_period_us=100ms
85	cpu.cfs_quota_us=-1
86	cpu.cfs_burst_us=0
87
88A value of -1 for cpu.cfs_quota_us indicates that the group does not have any
89bandwidth restriction in place, such a group is described as an unconstrained
90bandwidth group. This represents the traditional work-conserving behavior for
91CFS.
92
93Writing any (valid) positive value(s) no smaller than cpu.cfs_burst_us will
94enact the specified bandwidth limit. The minimum quota allowed for the quota or
95period is 1ms. There is also an upper bound on the period length of 1s.
96Additional restrictions exist when bandwidth limits are used in a hierarchical
97fashion, these are explained in more detail below.
98
99Writing any negative value to cpu.cfs_quota_us will remove the bandwidth limit
100and return the group to an unconstrained state once more.
101
102A value of 0 for cpu.cfs_burst_us indicates that the group can not accumulate
103any unused bandwidth. It makes the traditional bandwidth control behavior for
104CFS unchanged. Writing any (valid) positive value(s) no larger than
105cpu.cfs_quota_us into cpu.cfs_burst_us will enact the cap on unused bandwidth
106accumulation.
107
108Any updates to a group's bandwidth specification will result in it becoming
109unthrottled if it is in a constrained state.
110
111System wide settings
112--------------------
113For efficiency run-time is transferred between the global pool and CPU local
114"silos" in a batch fashion. This greatly reduces global accounting pressure
115on large systems. The amount transferred each time such an update is required
116is described as the "slice".
117
118This is tunable via procfs::
119
120	/proc/sys/kernel/sched_cfs_bandwidth_slice_us (default=5ms)
121
122Larger slice values will reduce transfer overheads, while smaller values allow
123for more fine-grained consumption.
124
125Statistics
126----------
127A group's bandwidth statistics are exported via 5 fields in cpu.stat.
128
129cpu.stat:
130
131- nr_periods: Number of enforcement intervals that have elapsed.
132- nr_throttled: Number of times the group has been throttled/limited.
133- throttled_time: The total time duration (in nanoseconds) for which entities
134  of the group have been throttled.
135- nr_bursts: Number of periods burst occurs.
136- burst_time: Cumulative wall-time (in nanoseconds) that any CPUs has used
137  above quota in respective periods.
138
139This interface is read-only.
140
141Hierarchical considerations
142---------------------------
143The interface enforces that an individual entity's bandwidth is always
144attainable, that is: max(c_i) <= C. However, over-subscription in the
145aggregate case is explicitly allowed to enable work-conserving semantics
146within a hierarchy:
147
148  e.g. \Sum (c_i) may exceed C
149
150[ Where C is the parent's bandwidth, and c_i its children ]
151
152
153There are two ways in which a group may become throttled:
154
155	a. it fully consumes its own quota within a period
156	b. a parent's quota is fully consumed within its period
157
158In case b) above, even though the child may have runtime remaining it will not
159be allowed to until the parent's runtime is refreshed.
160
161CFS Bandwidth Quota Caveats
162---------------------------
163Once a slice is assigned to a cpu it does not expire.  However all but 1ms of
164the slice may be returned to the global pool if all threads on that cpu become
165unrunnable. This is configured at compile time by the min_cfs_rq_runtime
166variable. This is a performance tweak that helps prevent added contention on
167the global lock.
168
169The fact that cpu-local slices do not expire results in some interesting corner
170cases that should be understood.
171
172For cgroup cpu constrained applications that are cpu limited this is a
173relatively moot point because they will naturally consume the entirety of their
174quota as well as the entirety of each cpu-local slice in each period. As a
175result it is expected that nr_periods roughly equal nr_throttled, and that
176cpuacct.usage will increase roughly equal to cfs_quota_us in each period.
177
178For highly-threaded, non-cpu bound applications this non-expiration nuance
179allows applications to briefly burst past their quota limits by the amount of
180unused slice on each cpu that the task group is running on (typically at most
1811ms per cpu or as defined by min_cfs_rq_runtime).  This slight burst only
182applies if quota had been assigned to a cpu and then not fully used or returned
183in previous periods. This burst amount will not be transferred between cores.
184As a result, this mechanism still strictly limits the task group to quota
185average usage, albeit over a longer time window than a single period.  This
186also limits the burst ability to no more than 1ms per cpu.  This provides
187better more predictable user experience for highly threaded applications with
188small quota limits on high core count machines. It also eliminates the
189propensity to throttle these applications while simultanously using less than
190quota amounts of cpu. Another way to say this, is that by allowing the unused
191portion of a slice to remain valid across periods we have decreased the
192possibility of wastefully expiring quota on cpu-local silos that don't need a
193full slice's amount of cpu time.
194
195The interaction between cpu-bound and non-cpu-bound-interactive applications
196should also be considered, especially when single core usage hits 100%. If you
197gave each of these applications half of a cpu-core and they both got scheduled
198on the same CPU it is theoretically possible that the non-cpu bound application
199will use up to 1ms additional quota in some periods, thereby preventing the
200cpu-bound application from fully using its quota by that same amount. In these
201instances it will be up to the CFS algorithm (see sched-design-CFS.rst) to
202decide which application is chosen to run, as they will both be runnable and
203have remaining quota. This runtime discrepancy will be made up in the following
204periods when the interactive application idles.
205
206Examples
207--------
2081. Limit a group to 1 CPU worth of runtime::
209
210	If period is 250ms and quota is also 250ms, the group will get
211	1 CPU worth of runtime every 250ms.
212
213	# echo 250000 > cpu.cfs_quota_us /* quota = 250ms */
214	# echo 250000 > cpu.cfs_period_us /* period = 250ms */
215
2162. Limit a group to 2 CPUs worth of runtime on a multi-CPU machine
217
218   With 500ms period and 1000ms quota, the group can get 2 CPUs worth of
219   runtime every 500ms::
220
221	# echo 1000000 > cpu.cfs_quota_us /* quota = 1000ms */
222	# echo 500000 > cpu.cfs_period_us /* period = 500ms */
223
224	The larger period here allows for increased burst capacity.
225
2263. Limit a group to 20% of 1 CPU.
227
228   With 50ms period, 10ms quota will be equivalent to 20% of 1 CPU::
229
230	# echo 10000 > cpu.cfs_quota_us /* quota = 10ms */
231	# echo 50000 > cpu.cfs_period_us /* period = 50ms */
232
233   By using a small period here we are ensuring a consistent latency
234   response at the expense of burst capacity.
235
2364. Limit a group to 40% of 1 CPU, and allow accumulate up to 20% of 1 CPU
237   additionally, in case accumulation has been done.
238
239   With 50ms period, 20ms quota will be equivalent to 40% of 1 CPU.
240   And 10ms burst will be equivalent to 20% of 1 CPU::
241
242	# echo 20000 > cpu.cfs_quota_us /* quota = 20ms */
243	# echo 50000 > cpu.cfs_period_us /* period = 50ms */
244	# echo 10000 > cpu.cfs_burst_us /* burst = 10ms */
245
246   Larger buffer setting (no larger than quota) allows greater burst capacity.
247