blob: c4c70e1aada3cbff8359c3d5f2ffcc29fc90fbc5 [file] [log] [blame] [edit]
======================================
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 three main ways of managing scheduling-clock interrupts
(also known as "scheduling-clock ticks" or simply "ticks"):
1. Never omit scheduling-clock ticks (CONFIG_HZ_PERIODIC=y or
CONFIG_NO_HZ=n for older kernels). You normally will -not-
want to choose this option.
2. Omit scheduling-clock ticks on idle CPUs (CONFIG_NO_HZ_IDLE=y or
CONFIG_NO_HZ=y for older kernels). This is the most common
approach, and should be the default.
3. Omit scheduling-clock ticks on CPUs that are either idle or that
have only one runnable task (CONFIG_NO_HZ_FULL=y). Unless you
are running realtime applications or certain types of HPC
workloads, you will normally -not- want this option.
These three cases are described in the following three sections, followed
by a third section on RCU-specific considerations, a fourth section
discussing testing, and a fifth and final section listing known issues.
Never Omit Scheduling-Clock Ticks
=================================
Very old versions of Linux from the 1990s and the very early 2000s
are incapable of omitting scheduling-clock ticks. It turns out that
there are some situations where this old-school approach is still the
right approach, for example, in heavy workloads with lots of tasks
that use short bursts of CPU, where there are very frequent idle
periods, but where these idle periods are also quite short (tens or
hundreds of microseconds). For these types of workloads, scheduling
clock interrupts will normally be delivered any way because there
will frequently be multiple runnable tasks per CPU. In these cases,
attempting to turn off the scheduling clock interrupt will have no effect
other than increasing the overhead of switching to and from idle and
transitioning between user and kernel execution.
This mode of operation can be selected using CONFIG_HZ_PERIODIC=y (or
CONFIG_NO_HZ=n for older kernels).
However, if you are instead running a light workload with long idle
periods, failing to omit scheduling-clock interrupts will result in
excessive power consumption. This is especially bad on battery-powered
devices, where it results in extremely short battery lifetimes. If you
are running light workloads, you should therefore read the following
section.
In addition, if you are running either a real-time workload or an HPC
workload with short iterations, the scheduling-clock interrupts can
degrade your applications performance. If this describes your workload,
you should read the following two sections.
Omit Scheduling-Clock Ticks For Idle CPUs
=========================================
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
dyntick-idle mode.
Omit Scheduling-Clock Ticks For 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.
Note that omitting scheduling-clock ticks for CPUs with only one runnable
task implies also omitting them for idle CPUs.
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. Note that
this means that your system must have at least two CPUs in order for
CONFIG_NO_HZ_FULL=y to do anything for you.
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 prevent CPUs from entering adaptive-tick mode.
Real-time applications needing to take actions based on CPU time
consumption need to use other means of doing so.
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.
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.
RCU Implications
================
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 using 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 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.
Testing
=======
So you enable all the OS-jitter features described in this document,
but do not see any change in your workload's behavior. Is this because
your workload isn't affected that much by OS jitter, or is it because
something else is in the way? This section helps answer this question
by providing a simple OS-jitter test suite, which is available on branch
master of the following git archive:
git://git.kernel.org/pub/scm/linux/kernel/git/frederic/dynticks-testing.git
Clone this archive and follow the instructions in the README file.
This test procedure will produce a trace that will allow you to evaluate
whether or not you have succeeded in removing OS jitter from your system.
If this trace shows that you have removed OS jitter as much as is
possible, then you can conclude that your workload is not all that
sensitive to OS jitter.
Note: this test requires that your system have at least two CPUs.
We do not currently have a good way to remove OS jitter from single-CPU
systems.
Known Issues
============
* 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.
* 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.
* 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.
And even when there are multiple runnable tasks on a given CPU,
there is little point in interrupting that CPU until the current
running task's timeslice expires, which is almost always way
longer than the time of the next scheduling-clock interrupt.
Better handling of these sorts of situations is future work.
* 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.
* 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.
* 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.
* Unless all CPUs are idle, at least one CPU must keep the
scheduling-clock interrupt going in order to support accurate
timekeeping.
* If there might potentially be some adaptive-ticks CPUs, there
will be at least one CPU keeping the scheduling-clock interrupt
going, even if all CPUs are otherwise idle.
Better handling of this situation is ongoing work.
* Some process-handling operations still require the occasional
scheduling-clock tick. These operations include calculating CPU
load, maintaining sched average, computing CFS entity vruntime,
computing avenrun, and carrying out load balancing. They are
currently accommodated by scheduling-clock tick every second
or so. On-going work will eliminate the need even for these
infrequent scheduling-clock ticks.