1 NO_HZ: Reducing Scheduling-Clock Ticks
4 This document describes Kconfig options and boot parameters that can
5 reduce the number of scheduling-clock interrupts, thereby improving energy
6 efficiency and reducing OS jitter. Reducing OS jitter is important for
7 some types of computationally intensive high-performance computing (HPC)
8 applications and for real-time applications.
10 There are two main contexts in which the number of scheduling-clock
11 interrupts can be reduced compared to the old-school approach of sending
12 a scheduling-clock interrupt to all CPUs every jiffy whether they need
13 it or not (CONFIG_HZ_PERIODIC=y or CONFIG_NO_HZ=n for older kernels):
15 1. Idle CPUs (CONFIG_NO_HZ_IDLE=y or CONFIG_NO_HZ=y for older kernels).
17 2. CPUs having only one runnable task (CONFIG_NO_HZ_FULL=y).
19 These two cases are described in the following two sections, followed
20 by a third section on RCU-specific considerations and a fourth and final
21 section listing known issues.
26 If a CPU is idle, there is little point in sending it a scheduling-clock
27 interrupt. After all, the primary purpose of a scheduling-clock interrupt
28 is to force a busy CPU to shift its attention among multiple duties,
29 and an idle CPU has no duties to shift its attention among.
31 The CONFIG_NO_HZ_IDLE=y Kconfig option causes the kernel to avoid sending
32 scheduling-clock interrupts to idle CPUs, which is critically important
33 both to battery-powered devices and to highly virtualized mainframes.
34 A battery-powered device running a CONFIG_HZ_PERIODIC=y kernel would
35 drain its battery very quickly, easily 2-3 times as fast as would the
36 same device running a CONFIG_NO_HZ_IDLE=y kernel. A mainframe running
37 1,500 OS instances might find that half of its CPU time was consumed by
38 unnecessary scheduling-clock interrupts. In these situations, there
39 is strong motivation to avoid sending scheduling-clock interrupts to
40 idle CPUs. That said, dyntick-idle mode is not free:
42 1. It increases the number of instructions executed on the path
43 to and from the idle loop.
45 2. On many architectures, dyntick-idle mode also increases the
46 number of expensive clock-reprogramming operations.
48 Therefore, systems with aggressive real-time response constraints often
49 run CONFIG_HZ_PERIODIC=y kernels (or CONFIG_NO_HZ=n for older kernels)
50 in order to avoid degrading from-idle transition latencies.
52 An idle CPU that is not receiving scheduling-clock interrupts is said to
53 be "dyntick-idle", "in dyntick-idle mode", "in nohz mode", or "running
54 tickless". The remainder of this document will use "dyntick-idle mode".
56 There is also a boot parameter "nohz=" that can be used to disable
57 dyntick-idle mode in CONFIG_NO_HZ_IDLE=y kernels by specifying "nohz=off".
58 By default, CONFIG_NO_HZ_IDLE=y kernels boot with "nohz=on", enabling
62 CPUs WITH ONLY ONE RUNNABLE TASK
64 If a CPU has only one runnable task, there is little point in sending it
65 a scheduling-clock interrupt because there is no other task to switch to.
67 The CONFIG_NO_HZ_FULL=y Kconfig option causes the kernel to avoid
68 sending scheduling-clock interrupts to CPUs with a single runnable task,
69 and such CPUs are said to be "adaptive-ticks CPUs". This is important
70 for applications with aggressive real-time response constraints because
71 it allows them to improve their worst-case response times by the maximum
72 duration of a scheduling-clock interrupt. It is also important for
73 computationally intensive short-iteration workloads: If any CPU is
74 delayed during a given iteration, all the other CPUs will be forced to
75 wait idle while the delayed CPU finishes. Thus, the delay is multiplied
76 by one less than the number of CPUs. In these situations, there is
77 again strong motivation to avoid sending scheduling-clock interrupts.
79 By default, no CPU will be an adaptive-ticks CPU. The "nohz_full="
80 boot parameter specifies the adaptive-ticks CPUs. For example,
81 "nohz_full=1,6-8" says that CPUs 1, 6, 7, and 8 are to be adaptive-ticks
82 CPUs. Note that you are prohibited from marking all of the CPUs as
83 adaptive-tick CPUs: At least one non-adaptive-tick CPU must remain
84 online to handle timekeeping tasks in order to ensure that system calls
85 like gettimeofday() returns accurate values on adaptive-tick CPUs.
86 (This is not an issue for CONFIG_NO_HZ_IDLE=y because there are no
87 running user processes to observe slight drifts in clock rate.)
88 Therefore, the boot CPU is prohibited from entering adaptive-ticks
89 mode. Specifying a "nohz_full=" mask that includes the boot CPU will
90 result in a boot-time error message, and the boot CPU will be removed
93 Alternatively, the CONFIG_NO_HZ_FULL_ALL=y Kconfig parameter specifies
94 that all CPUs other than the boot CPU are adaptive-ticks CPUs. This
95 Kconfig parameter will be overridden by the "nohz_full=" boot parameter,
96 so that if both the CONFIG_NO_HZ_FULL_ALL=y Kconfig parameter and
97 the "nohz_full=1" boot parameter is specified, the boot parameter will
98 prevail so that only CPU 1 will be an adaptive-ticks CPU.
100 Finally, adaptive-ticks CPUs must have their RCU callbacks offloaded.
101 This is covered in the "RCU IMPLICATIONS" section below.
103 Normally, a CPU remains in adaptive-ticks mode as long as possible.
104 In particular, transitioning to kernel mode does not automatically change
105 the mode. Instead, the CPU will exit adaptive-ticks mode only if needed,
106 for example, if that CPU enqueues an RCU callback.
108 Just as with dyntick-idle mode, the benefits of adaptive-tick mode do
111 1. CONFIG_NO_HZ_FULL selects CONFIG_NO_HZ_COMMON, so you cannot run
112 adaptive ticks without also running dyntick idle. This dependency
113 extends down into the implementation, so that all of the costs
114 of CONFIG_NO_HZ_IDLE are also incurred by CONFIG_NO_HZ_FULL.
116 2. The user/kernel transitions are slightly more expensive due
117 to the need to inform kernel subsystems (such as RCU) about
120 3. POSIX CPU timers on adaptive-tick CPUs may miss their deadlines
121 (perhaps indefinitely) because they currently rely on
122 scheduling-tick interrupts. This will likely be fixed in
123 one of two ways: (1) Prevent CPUs with POSIX CPU timers from
124 entering adaptive-tick mode, or (2) Use hrtimers or other
125 adaptive-ticks-immune mechanism to cause the POSIX CPU timer to
128 4. If there are more perf events pending than the hardware can
129 accommodate, they are normally round-robined so as to collect
130 all of them over time. Adaptive-tick mode may prevent this
131 round-robining from happening. This will likely be fixed by
132 preventing CPUs with large numbers of perf events pending from
133 entering adaptive-tick mode.
135 5. Scheduler statistics for adaptive-tick CPUs may be computed
136 slightly differently than those for non-adaptive-tick CPUs.
137 This might in turn perturb load-balancing of real-time tasks.
139 6. The LB_BIAS scheduler feature is disabled by adaptive ticks.
141 Although improvements are expected over time, adaptive ticks is quite
142 useful for many types of real-time and compute-intensive applications.
143 However, the drawbacks listed above mean that adaptive ticks should not
144 (yet) be enabled by default.
149 There are situations in which idle CPUs cannot be permitted to
150 enter either dyntick-idle mode or adaptive-tick mode, the most
151 common being when that CPU has RCU callbacks pending.
153 The CONFIG_RCU_FAST_NO_HZ=y Kconfig option may be used to cause such CPUs
154 to enter dyntick-idle mode or adaptive-tick mode anyway. In this case,
155 a timer will awaken these CPUs every four jiffies in order to ensure
156 that the RCU callbacks are processed in a timely fashion.
158 Another approach is to offload RCU callback processing to "rcuo" kthreads
159 using the CONFIG_RCU_NOCB_CPU=y Kconfig option. The specific CPUs to
160 offload may be selected via several methods:
162 1. One of three mutually exclusive Kconfig options specify a
163 build-time default for the CPUs to offload:
165 a. The CONFIG_RCU_NOCB_CPU_NONE=y Kconfig option results in
166 no CPUs being offloaded.
168 b. The CONFIG_RCU_NOCB_CPU_ZERO=y Kconfig option causes
169 CPU 0 to be offloaded.
171 c. The CONFIG_RCU_NOCB_CPU_ALL=y Kconfig option causes all
172 CPUs to be offloaded. Note that the callbacks will be
173 offloaded to "rcuo" kthreads, and that those kthreads
174 will in fact run on some CPU. However, this approach
175 gives fine-grained control on exactly which CPUs the
176 callbacks run on, along with their scheduling priority
177 (including the default of SCHED_OTHER), and it further
178 allows this control to be varied dynamically at runtime.
180 2. The "rcu_nocbs=" kernel boot parameter, which takes a comma-separated
181 list of CPUs and CPU ranges, for example, "1,3-5" selects CPUs 1,
182 3, 4, and 5. The specified CPUs will be offloaded in addition to
183 any CPUs specified as offloaded by CONFIG_RCU_NOCB_CPU_ZERO=y or
184 CONFIG_RCU_NOCB_CPU_ALL=y. This means that the "rcu_nocbs=" boot
185 parameter has no effect for kernels built with RCU_NOCB_CPU_ALL=y.
187 The offloaded CPUs will never queue RCU callbacks, and therefore RCU
188 never prevents offloaded CPUs from entering either dyntick-idle mode
189 or adaptive-tick mode. That said, note that it is up to userspace to
190 pin the "rcuo" kthreads to specific CPUs if desired. Otherwise, the
191 scheduler will decide where to run them, which might or might not be
192 where you want them to run.
197 o Dyntick-idle slows transitions to and from idle slightly.
198 In practice, this has not been a problem except for the most
199 aggressive real-time workloads, which have the option of disabling
200 dyntick-idle mode, an option that most of them take. However,
201 some workloads will no doubt want to use adaptive ticks to
202 eliminate scheduling-clock interrupt latencies. Here are some
203 options for these workloads:
205 a. Use PMQOS from userspace to inform the kernel of your
206 latency requirements (preferred).
208 b. On x86 systems, use the "idle=mwait" boot parameter.
210 c. On x86 systems, use the "intel_idle.max_cstate=" to limit
211 ` the maximum C-state depth.
213 d. On x86 systems, use the "idle=poll" boot parameter.
214 However, please note that use of this parameter can cause
215 your CPU to overheat, which may cause thermal throttling
216 to degrade your latencies -- and that this degradation can
217 be even worse than that of dyntick-idle. Furthermore,
218 this parameter effectively disables Turbo Mode on Intel
219 CPUs, which can significantly reduce maximum performance.
221 o Adaptive-ticks slows user/kernel transitions slightly.
222 This is not expected to be a problem for computationally intensive
223 workloads, which have few such transitions. Careful benchmarking
224 will be required to determine whether or not other workloads
225 are significantly affected by this effect.
227 o Adaptive-ticks does not do anything unless there is only one
228 runnable task for a given CPU, even though there are a number
229 of other situations where the scheduling-clock tick is not
230 needed. To give but one example, consider a CPU that has one
231 runnable high-priority SCHED_FIFO task and an arbitrary number
232 of low-priority SCHED_OTHER tasks. In this case, the CPU is
233 required to run the SCHED_FIFO task until it either blocks or
234 some other higher-priority task awakens on (or is assigned to)
235 this CPU, so there is no point in sending a scheduling-clock
236 interrupt to this CPU. However, the current implementation
237 nevertheless sends scheduling-clock interrupts to CPUs having a
238 single runnable SCHED_FIFO task and multiple runnable SCHED_OTHER
239 tasks, even though these interrupts are unnecessary.
241 Better handling of these sorts of situations is future work.
243 o A reboot is required to reconfigure both adaptive idle and RCU
244 callback offloading. Runtime reconfiguration could be provided
245 if needed, however, due to the complexity of reconfiguring RCU at
246 runtime, there would need to be an earthshakingly good reason.
247 Especially given that you have the straightforward option of
248 simply offloading RCU callbacks from all CPUs and pinning them
249 where you want them whenever you want them pinned.
251 o Additional configuration is required to deal with other sources
252 of OS jitter, including interrupts and system-utility tasks
253 and processes. This configuration normally involves binding
254 interrupts and tasks to particular CPUs.
256 o Some sources of OS jitter can currently be eliminated only by
257 constraining the workload. For example, the only way to eliminate
258 OS jitter due to global TLB shootdowns is to avoid the unmapping
259 operations (such as kernel module unload operations) that
260 result in these shootdowns. For another example, page faults
261 and TLB misses can be reduced (and in some cases eliminated) by
262 using huge pages and by constraining the amount of memory used
263 by the application. Pre-faulting the working set can also be
264 helpful, especially when combined with the mlock() and mlockall()
267 o Unless all CPUs are idle, at least one CPU must keep the
268 scheduling-clock interrupt going in order to support accurate
271 o If there are adaptive-ticks CPUs, there will be at least one
272 CPU keeping the scheduling-clock interrupt going, even if all
273 CPUs are otherwise idle.