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
25Management
26----------
27Quota and period are managed within the cpu subsystem via cgroupfs.
28
29.. note::
30   The cgroupfs files described in this section are only applicable
31   to cgroup v1. For cgroup v2, see
32   :ref:`Documentation/admin-guide/cgroup-v2.rst <cgroup-v2-cpu>`.
33
34- cpu.cfs_quota_us: the total available run-time within a period (in
35  microseconds)
36- cpu.cfs_period_us: the length of a period (in microseconds)
37- cpu.stat: exports throttling statistics [explained further below]
38
39The default values are::
40
41	cpu.cfs_period_us=100ms
42	cpu.cfs_quota=-1
43
44A value of -1 for cpu.cfs_quota_us indicates that the group does not have any
45bandwidth restriction in place, such a group is described as an unconstrained
46bandwidth group. This represents the traditional work-conserving behavior for
47CFS.
48
49Writing any (valid) positive value(s) will enact the specified bandwidth limit.
50The minimum quota allowed for the quota or period is 1ms. There is also an
51upper bound on the period length of 1s. Additional restrictions exist when
52bandwidth limits are used in a hierarchical fashion, these are explained in
53more detail below.
54
55Writing any negative value to cpu.cfs_quota_us will remove the bandwidth limit
56and return the group to an unconstrained state once more.
57
58Any updates to a group's bandwidth specification will result in it becoming
59unthrottled if it is in a constrained state.
60
61System wide settings
62--------------------
63For efficiency run-time is transferred between the global pool and CPU local
64"silos" in a batch fashion. This greatly reduces global accounting pressure
65on large systems. The amount transferred each time such an update is required
66is described as the "slice".
67
68This is tunable via procfs::
69
70	/proc/sys/kernel/sched_cfs_bandwidth_slice_us (default=5ms)
71
72Larger slice values will reduce transfer overheads, while smaller values allow
73for more fine-grained consumption.
74
75Statistics
76----------
77A group's bandwidth statistics are exported via 3 fields in cpu.stat.
78
79cpu.stat:
80
81- nr_periods: Number of enforcement intervals that have elapsed.
82- nr_throttled: Number of times the group has been throttled/limited.
83- throttled_time: The total time duration (in nanoseconds) for which entities
84  of the group have been throttled.
85
86This interface is read-only.
87
88Hierarchical considerations
89---------------------------
90The interface enforces that an individual entity's bandwidth is always
91attainable, that is: max(c_i) <= C. However, over-subscription in the
92aggregate case is explicitly allowed to enable work-conserving semantics
93within a hierarchy:
94
95  e.g. \Sum (c_i) may exceed C
96
97[ Where C is the parent's bandwidth, and c_i its children ]
98
99
100There are two ways in which a group may become throttled:
101
102	a. it fully consumes its own quota within a period
103	b. a parent's quota is fully consumed within its period
104
105In case b) above, even though the child may have runtime remaining it will not
106be allowed to until the parent's runtime is refreshed.
107
108CFS Bandwidth Quota Caveats
109---------------------------
110Once a slice is assigned to a cpu it does not expire.  However all but 1ms of
111the slice may be returned to the global pool if all threads on that cpu become
112unrunnable. This is configured at compile time by the min_cfs_rq_runtime
113variable. This is a performance tweak that helps prevent added contention on
114the global lock.
115
116The fact that cpu-local slices do not expire results in some interesting corner
117cases that should be understood.
118
119For cgroup cpu constrained applications that are cpu limited this is a
120relatively moot point because they will naturally consume the entirety of their
121quota as well as the entirety of each cpu-local slice in each period. As a
122result it is expected that nr_periods roughly equal nr_throttled, and that
123cpuacct.usage will increase roughly equal to cfs_quota_us in each period.
124
125For highly-threaded, non-cpu bound applications this non-expiration nuance
126allows applications to briefly burst past their quota limits by the amount of
127unused slice on each cpu that the task group is running on (typically at most
1281ms per cpu or as defined by min_cfs_rq_runtime).  This slight burst only
129applies if quota had been assigned to a cpu and then not fully used or returned
130in previous periods. This burst amount will not be transferred between cores.
131As a result, this mechanism still strictly limits the task group to quota
132average usage, albeit over a longer time window than a single period.  This
133also limits the burst ability to no more than 1ms per cpu.  This provides
134better more predictable user experience for highly threaded applications with
135small quota limits on high core count machines. It also eliminates the
136propensity to throttle these applications while simultanously using less than
137quota amounts of cpu. Another way to say this, is that by allowing the unused
138portion of a slice to remain valid across periods we have decreased the
139possibility of wastefully expiring quota on cpu-local silos that don't need a
140full slice's amount of cpu time.
141
142The interaction between cpu-bound and non-cpu-bound-interactive applications
143should also be considered, especially when single core usage hits 100%. If you
144gave each of these applications half of a cpu-core and they both got scheduled
145on the same CPU it is theoretically possible that the non-cpu bound application
146will use up to 1ms additional quota in some periods, thereby preventing the
147cpu-bound application from fully using its quota by that same amount. In these
148instances it will be up to the CFS algorithm (see sched-design-CFS.rst) to
149decide which application is chosen to run, as they will both be runnable and
150have remaining quota. This runtime discrepancy will be made up in the following
151periods when the interactive application idles.
152
153Examples
154--------
1551. Limit a group to 1 CPU worth of runtime::
156
157	If period is 250ms and quota is also 250ms, the group will get
158	1 CPU worth of runtime every 250ms.
159
160	# echo 250000 > cpu.cfs_quota_us /* quota = 250ms */
161	# echo 250000 > cpu.cfs_period_us /* period = 250ms */
162
1632. Limit a group to 2 CPUs worth of runtime on a multi-CPU machine
164
165   With 500ms period and 1000ms quota, the group can get 2 CPUs worth of
166   runtime every 500ms::
167
168	# echo 1000000 > cpu.cfs_quota_us /* quota = 1000ms */
169	# echo 500000 > cpu.cfs_period_us /* period = 500ms */
170
171	The larger period here allows for increased burst capacity.
172
1733. Limit a group to 20% of 1 CPU.
174
175   With 50ms period, 10ms quota will be equivalent to 20% of 1 CPU::
176
177	# echo 10000 > cpu.cfs_quota_us /* quota = 10ms */
178	# echo 50000 > cpu.cfs_period_us /* period = 50ms */
179
180   By using a small period here we are ensuring a consistent latency
181   response at the expense of burst capacity.
182