| NO_HZ: Reducing Scheduling-Clock Ticks
|This document describes Kconfig options and boot parameters that can
|reduce the number of scheduling-clock interrupts, thereby improving energy
|efficiency and reducing OS jitter. Reducing OS jitter is important for
|some types of computationally intensive high-performance computing (HPC)
|applications and for real-time applications.
|There are two main contexts in which the number of scheduling-clock
|interrupts can be reduced compared to the old-school approach of sending
|a scheduling-clock interrupt to all CPUs every jiffy whether they need
|it or not (CONFIG_HZ_PERIODIC=y or CONFIG_NO_HZ=n for older kernels):
|1. Idle CPUs (CONFIG_NO_HZ_IDLE=y or CONFIG_NO_HZ=y for older kernels).
|2. CPUs having only one runnable task (CONFIG_NO_HZ_FULL=y).
|These two cases are described in the following two sections, followed
|by a third section on RCU-specific considerations and a fourth and final
|section listing known issues.
|If a CPU is idle, there is little point in sending it a scheduling-clock
|interrupt. After all, the primary purpose of a scheduling-clock interrupt
|is to force a busy CPU to shift its attention among multiple duties,
|and an idle CPU has no duties to shift its attention among.
|The CONFIG_NO_HZ_IDLE=y Kconfig option causes the kernel to avoid sending
|scheduling-clock interrupts to idle CPUs, which is critically important
|both to battery-powered devices and to highly virtualized mainframes.
|A battery-powered device running a CONFIG_HZ_PERIODIC=y kernel would
|drain its battery very quickly, easily 2-3 times as fast as would the
|same device running a CONFIG_NO_HZ_IDLE=y kernel. A mainframe running
|1,500 OS instances might find that half of its CPU time was consumed by
|unnecessary scheduling-clock interrupts. In these situations, there
|is strong motivation to avoid sending scheduling-clock interrupts to
|idle CPUs. That said, dyntick-idle mode is not free:
|1. It increases the number of instructions executed on the path
| to and from the idle loop.
|2. On many architectures, dyntick-idle mode also increases the
| number of expensive clock-reprogramming operations.
|Therefore, systems with aggressive real-time response constraints often
|run CONFIG_HZ_PERIODIC=y kernels (or CONFIG_NO_HZ=n for older kernels)
|in order to avoid degrading from-idle transition latencies.
|An idle CPU that is not receiving scheduling-clock interrupts is said to
|be "dyntick-idle", "in dyntick-idle mode", "in nohz mode", or "running
|tickless". The remainder of this document will use "dyntick-idle mode".
|There is also a boot parameter "nohz=" that can be used to disable
|dyntick-idle mode in CONFIG_NO_HZ_IDLE=y kernels by specifying "nohz=off".
|By default, CONFIG_NO_HZ_IDLE=y kernels boot with "nohz=on", enabling
|CPUs WITH ONLY ONE RUNNABLE TASK
|If a CPU has only one runnable task, there is little point in sending it
|a scheduling-clock interrupt because there is no other task to switch to.
|The CONFIG_NO_HZ_FULL=y Kconfig option causes the kernel to avoid
|sending scheduling-clock interrupts to CPUs with a single runnable task,
|and such CPUs are said to be "adaptive-ticks CPUs". This is important
|for applications with aggressive real-time response constraints because
|it allows them to improve their worst-case response times by the maximum
|duration of a scheduling-clock interrupt. It is also important for
|computationally intensive short-iteration workloads: If any CPU is
|delayed during a given iteration, all the other CPUs will be forced to
|wait idle while the delayed CPU finishes. Thus, the delay is multiplied
|by one less than the number of CPUs. In these situations, there is
|again strong motivation to avoid sending scheduling-clock interrupts.
|By default, no CPU will be an adaptive-ticks CPU. The "nohz_full="
|boot parameter specifies the adaptive-ticks CPUs. For example,
|"nohz_full=1,6-8" says that CPUs 1, 6, 7, and 8 are to be adaptive-ticks
|CPUs. Note that you are prohibited from marking all of the CPUs as
|adaptive-tick CPUs: At least one non-adaptive-tick CPU must remain
|online to handle timekeeping tasks in order to ensure that system calls
|like gettimeofday() returns accurate values on adaptive-tick CPUs.
|(This is not an issue for CONFIG_NO_HZ_IDLE=y because there are no
|running user processes to observe slight drifts in clock rate.)
|Therefore, the boot CPU is prohibited from entering adaptive-ticks
|mode. Specifying a "nohz_full=" mask that includes the boot CPU will
|result in a boot-time error message, and the boot CPU will be removed
|from the mask.
|Alternatively, the CONFIG_NO_HZ_FULL_ALL=y Kconfig parameter specifies
|that all CPUs other than the boot CPU are adaptive-ticks CPUs. This
|Kconfig parameter will be overridden by the "nohz_full=" boot parameter,
|so that if both the CONFIG_NO_HZ_FULL_ALL=y Kconfig parameter and
|the "nohz_full=1" boot parameter is specified, the boot parameter will
|prevail so that only CPU 1 will be an adaptive-ticks CPU.
|Finally, adaptive-ticks CPUs must have their RCU callbacks offloaded.
|This is covered in the "RCU IMPLICATIONS" section below.
|Normally, a CPU remains in adaptive-ticks mode as long as possible.
|In particular, transitioning to kernel mode does not automatically change
|the mode. Instead, the CPU will exit adaptive-ticks mode only if needed,
|for example, if that CPU enqueues an RCU callback.
|Just as with dyntick-idle mode, the benefits of adaptive-tick mode do
|not come for free:
|1. CONFIG_NO_HZ_FULL selects CONFIG_NO_HZ_COMMON, so you cannot run
| adaptive ticks without also running dyntick idle. This dependency
| extends down into the implementation, so that all of the costs
| of CONFIG_NO_HZ_IDLE are also incurred by CONFIG_NO_HZ_FULL.
|2. The user/kernel transitions are slightly more expensive due
| to the need to inform kernel subsystems (such as RCU) about
| the change in mode.
|3. POSIX CPU timers on adaptive-tick CPUs may miss their deadlines
| (perhaps indefinitely) because they currently rely on
| scheduling-tick interrupts. This will likely be fixed in
| one of two ways: (1) Prevent CPUs with POSIX CPU timers from
| entering adaptive-tick mode, or (2) Use hrtimers or other
| adaptive-ticks-immune mechanism to cause the POSIX CPU timer to
| fire properly.
|4. If there are more perf events pending than the hardware can
| accommodate, they are normally round-robined so as to collect
| all of them over time. Adaptive-tick mode may prevent this
| round-robining from happening. This will likely be fixed by
| preventing CPUs with large numbers of perf events pending from
| entering adaptive-tick mode.
|5. Scheduler statistics for adaptive-tick CPUs may be computed
| slightly differently than those for non-adaptive-tick CPUs.
| This might in turn perturb load-balancing of real-time tasks.
|6. The LB_BIAS scheduler feature is disabled by adaptive ticks.
|Although improvements are expected over time, adaptive ticks is quite
|useful for many types of real-time and compute-intensive applications.
|However, the drawbacks listed above mean that adaptive ticks should not
|(yet) be enabled by default.
|There are situations in which idle CPUs cannot be permitted to
|enter either dyntick-idle mode or adaptive-tick mode, the most
|common being when that CPU has RCU callbacks pending.
|The CONFIG_RCU_FAST_NO_HZ=y Kconfig option may be used to cause such CPUs
|to enter dyntick-idle mode or adaptive-tick mode anyway. In this case,
|a timer will awaken these CPUs every four jiffies in order to ensure
|that the RCU callbacks are processed in a timely fashion.
|Another approach is to offload RCU callback processing to "rcuo" kthreads
|using the CONFIG_RCU_NOCB_CPU=y Kconfig option. The specific CPUs to
|offload may be selected via several methods:
|1. One of three mutually exclusive Kconfig options specify a
| build-time default for the CPUs to offload:
| a. The CONFIG_RCU_NOCB_CPU_NONE=y Kconfig option results in
| no CPUs being offloaded.
| b. The CONFIG_RCU_NOCB_CPU_ZERO=y Kconfig option causes
| CPU 0 to be offloaded.
| c. The CONFIG_RCU_NOCB_CPU_ALL=y Kconfig option causes all
| CPUs to be offloaded. Note that the callbacks will be
| offloaded to "rcuo" kthreads, and that those kthreads
| will in fact run on some CPU. However, this approach
| gives fine-grained control on exactly which CPUs the
| callbacks run on, along with their scheduling priority
| (including the default of SCHED_OTHER), and it further
| allows this control to be varied dynamically at runtime.
|2. The "rcu_nocbs=" kernel boot parameter, which takes a comma-separated
| list of CPUs and CPU ranges, for example, "1,3-5" selects CPUs 1,
| 3, 4, and 5. The specified CPUs will be offloaded in addition to
| any CPUs specified as offloaded by CONFIG_RCU_NOCB_CPU_ZERO=y or
| CONFIG_RCU_NOCB_CPU_ALL=y. This means that the "rcu_nocbs=" boot
| parameter has no effect for kernels built with RCU_NOCB_CPU_ALL=y.
|The offloaded CPUs will never queue RCU callbacks, and therefore RCU
|never prevents offloaded CPUs from entering either dyntick-idle mode
|or adaptive-tick mode. That said, note that it is up to userspace to
|pin the "rcuo" kthreads to specific CPUs if desired. Otherwise, the
|scheduler will decide where to run them, which might or might not be
|where you want them to run.
|o Dyntick-idle slows transitions to and from idle slightly.
| In practice, this has not been a problem except for the most
| aggressive real-time workloads, which have the option of disabling
| dyntick-idle mode, an option that most of them take. However,
| some workloads will no doubt want to use adaptive ticks to
| eliminate scheduling-clock interrupt latencies. Here are some
| options for these workloads:
| a. Use PMQOS from userspace to inform the kernel of your
| latency requirements (preferred).
| b. On x86 systems, use the "idle=mwait" boot parameter.
| c. On x86 systems, use the "intel_idle.max_cstate=" to limit
| ` the maximum C-state depth.
| d. On x86 systems, use the "idle=poll" boot parameter.
| However, please note that use of this parameter can cause
| your CPU to overheat, which may cause thermal throttling
| to degrade your latencies -- and that this degradation can
| be even worse than that of dyntick-idle. Furthermore,
| this parameter effectively disables Turbo Mode on Intel
| CPUs, which can significantly reduce maximum performance.
|o Adaptive-ticks slows user/kernel transitions slightly.
| This is not expected to be a problem for computationally intensive
| workloads, which have few such transitions. Careful benchmarking
| will be required to determine whether or not other workloads
| are significantly affected by this effect.
|o Adaptive-ticks does not do anything unless there is only one
| runnable task for a given CPU, even though there are a number
| of other situations where the scheduling-clock tick is not
| needed. To give but one example, consider a CPU that has one
| runnable high-priority SCHED_FIFO task and an arbitrary number
| of low-priority SCHED_OTHER tasks. In this case, the CPU is
| required to run the SCHED_FIFO task until it either blocks or
| some other higher-priority task awakens on (or is assigned to)
| this CPU, so there is no point in sending a scheduling-clock
| interrupt to this CPU. However, the current implementation
| nevertheless sends scheduling-clock interrupts to CPUs having a
| single runnable SCHED_FIFO task and multiple runnable SCHED_OTHER
| tasks, even though these interrupts are unnecessary.
| Better handling of these sorts of situations is future work.
|o A reboot is required to reconfigure both adaptive idle and RCU
| callback offloading. Runtime reconfiguration could be provided
| if needed, however, due to the complexity of reconfiguring RCU at
| runtime, there would need to be an earthshakingly good reason.
| Especially given that you have the straightforward option of
| simply offloading RCU callbacks from all CPUs and pinning them
| where you want them whenever you want them pinned.
|o Additional configuration is required to deal with other sources
| of OS jitter, including interrupts and system-utility tasks
| and processes. This configuration normally involves binding
| interrupts and tasks to particular CPUs.
|o Some sources of OS jitter can currently be eliminated only by
| constraining the workload. For example, the only way to eliminate
| OS jitter due to global TLB shootdowns is to avoid the unmapping
| operations (such as kernel module unload operations) that
| result in these shootdowns. For another example, page faults
| and TLB misses can be reduced (and in some cases eliminated) by
| using huge pages and by constraining the amount of memory used
| by the application. Pre-faulting the working set can also be
| helpful, especially when combined with the mlock() and mlockall()
| system calls.
|o Unless all CPUs are idle, at least one CPU must keep the
| scheduling-clock interrupt going in order to support accurate
|o If there are adaptive-ticks CPUs, there will be at least one
| CPU keeping the scheduling-clock interrupt going, even if all
| CPUs are otherwise idle.