path: root/Documentation/scheduler/sched-BFS.txt

BFS - The Brain Fuck Scheduler by Con Kolivas.

Goals.

The goal of the Brain Fuck Scheduler, referred to as BFS from here on, is to
completely do away with the complex designs of the past for the cpu process
scheduler and instead implement one that is very simple in basic design.
The main focus of BFS is to achieve excellent desktop interactivity and
responsiveness without heuristics and tuning knobs that are difficult to
understand, impossible to model and predict the effect of, and when tuned to
one workload cause massive detriment to another.


Design summary.

BFS is best described as a single runqueue, O(n) lookup, earliest effective
virtual deadline first design, loosely based on EEVDF (earliest eligible virtual
deadline first) and my previous Staircase Deadline scheduler. Each component
shall be described in order to understand the significance of, and reasoning for
it. The codebase when the first stable version was released was approximately
9000 lines less code than the existing mainline linux kernel scheduler (in
2.6.31). This does not even take into account the removal of documentation and
the cgroups code that is not used.

Design reasoning.

The single runqueue refers to the queued but not running processes for the
entire system, regardless of the number of CPUs. The reason for going back to
a single runqueue design is that once multiple runqueues are introduced,
per-CPU or otherwise, there will be complex interactions as each runqueue will
be responsible for the scheduling latency and fairness of the tasks only on its
own runqueue, and to achieve fairness and low latency across multiple CPUs, any
advantage in throughput of having CPU local tasks causes other disadvantages.
This is due to requiring a very complex balancing system to at best achieve some
semblance of fairness across CPUs and can only maintain relatively low latency
for tasks bound to the same CPUs, not across them. To increase said fairness
and latency across CPUs, the advantage of local runqueue locking, which makes
for better scalability, is lost due to having to grab multiple locks.

A significant feature of BFS is that all accounting is done purely based on CPU
used and nowhere is sleep time used in any way to determine entitlement or
interactivity. Interactivity "estimators" that use some kind of sleep/run
algorithm are doomed to fail to detect all interactive tasks, and to falsely tag
tasks that aren't interactive as being so. The reason for this is that it is
close to impossible to determine that when a task is sleeping, whether it is
doing it voluntarily, as in a userspace application waiting for input in the
form of a mouse click or otherwise, or involuntarily, because it is waiting for
another thread, process, I/O, kernel activity or whatever. Thus, such an
estimator will introduce corner cases, and more heuristics will be required to
cope with those corner cases, introducing more corner cases and failed
interactivity detection and so on. Interactivity in BFS is built into the design
by virtue of the fact that tasks that are waking up have not used up their quota
of CPU time, and have earlier effective deadlines, thereby making it very likely
they will preempt any CPU bound task of equivalent nice level. See below for
more information on the virtual deadline mechanism. Even if they do not preempt
a running task, because the rr interval is guaranteed to have a bound upper
limit on how long a task will wait for, it will be scheduled within a timeframe
that will not cause visible interface jitter.


Design details.

Task insertion.

BFS inserts tasks into each relevant queue as an O(1) insertion into a double
linked list. On insertion, *every* running queue is checked to see if the newly
queued task can run on any idle queue, or preempt the lowest running task on the
system. This is how the cross-CPU scheduling of BFS achieves significantly lower
latency per extra CPU the system has. In this case the lookup is, in the worst
case scenario, O(n) where n is the number of CPUs on the system.

Data protection.

BFS has one single lock protecting the process local data of every task in the
global queue. Thus every insertion, removal and modification of task data in the
global runqueue needs to grab the global lock. However, once a task is taken by
a CPU, the CPU has its own local data copy of the running process' accounting
information which only that CPU accesses and modifies (such as during a
timer tick) thus allowing the accounting data to be updated lockless. Once a
CPU has taken a task to run, it removes it from the global queue. Thus the
global queue only ever has, at most,

	(number of tasks requesting cpu time) - (number of logical CPUs) + 1

tasks in the global queue. This value is relevant for the time taken to look up
tasks during scheduling. This will increase if many tasks with CPU affinity set
in their policy to limit which CPUs they're allowed to run on if they outnumber
the number of CPUs. The +1 is because when rescheduling a task, the CPU's
currently running task is put back on the queue. Lookup will be described after
the virtual deadline mechanism is explained.

Virtual deadline.

The key to achieving low latency, scheduling fairness, and "nice level"
distribution in BFS is entirely in the virtual deadline mechanism. The one
tunable in BFS is the rr_interval, or "round robin interval". This is the
maximum time two SCHED_OTHER (or SCHED_NORMAL, the common scheduling policy)
tasks of the same nice level will be running for, or looking at it the other
way around, the longest duration two tasks of the same nice level will be
delayed for. When a task requests cpu time, it is given a quota (time_slice)
equal to the rr_interval and a virtual deadline. The virtual deadline is
offset from the current time in jiffies by this equation:

	jiffies + (prio_ratio * rr_interval)

The prio_ratio is determined as a ratio compared to the baseline of nice -20
and increases by 10% per nice level. The deadline is a virtual one only in that
no guarantee is placed that a task will actually be scheduled by this time, but
it is used to compare which task should go next. There are three components to
how a task is next chosen. First is time_slice expiration. If a task runs out
of its time_slice, it is descheduled, the time_slice is refilled, and the
deadline reset to that formula above. Second is sleep, where a task no longer
is requesting CPU for whatever reason. The time_slice and deadline are _not_
adjusted in this case and are just carried over for when the task is next
scheduled. Third is preemption, and that is when a newly waking task is deemed
higher priority than a currently running task on any cpu by virtue of the fact
that it has an earlier virtual deadline than the currently running task. The
earlier deadline is the key to which task is next chosen for the first and
second cases. Once a task is descheduled, it is put back on the queue, and an
O(n) lookup of all queued-but-not-running tasks is done to determine which has
the earliest deadline and that task is chosen to receive CPU next.

The CPU proportion of different nice tasks works out to be approximately the

	(prio_ratio difference)^2

The reason it is squared is that a task's deadline does not change while it is
running unless it runs out of time_slice. Thus, even if the time actually
passes the deadline of another task that is queued, it will not get CPU time
unless the current running task deschedules, and the time "base" (jiffies) is
constantly moving.

Task lookup.

BFS has 103 priority queues. 100 of these are dedicated to the static priority
of realtime tasks, and the remaining 3 are, in order of best to worst priority,
SCHED_ISO (isochronous), SCHED_NORMAL, and SCHED_IDLEPRIO (idle priority
scheduling). When a task of these priorities is queued, a bitmap of running
priorities is set showing which of these priorities has tasks waiting for CPU
time. When a CPU is made to reschedule, the lookup for the next task to get
CPU time is performed in the following way:

First the bitmap is checked to see what static priority tasks are queued. If
any realtime priorities are found, the corresponding queue is checked and the
first task listed there is taken (provided CPU affinity is suitable) and lookup
is complete. If the priority corresponds to a SCHED_ISO task, they are also
taken in FIFO order (as they behave like SCHED_RR). If the priority corresponds
to either SCHED_NORMAL or SCHED_IDLEPRIO, then the lookup becomes O(n). At this
stage, every task in the runlist that corresponds to that priority is checked
to see which has the earliest set deadline, and (provided it has suitable CPU
affinity) it is taken off the runqueue and given the CPU. If a task has an
expired deadline, it is taken and the rest of the lookup aborted (as they are
chosen in FIFO order).

Thus, the lookup is O(n) in the worst case only, where n is as described
earlier, as tasks may be chosen before the whole task list is looked over.


Scalability.

The major limitations of BFS will be that of scalability, as the separate
runqueue designs will have less lock contention as the number of CPUs rises.
However they do not scale linearly even with separate runqueues as multiple
runqueues will need to be locked concurrently on such designs to be able to
achieve fair CPU balancing, to try and achieve some sort of nice-level fairness
across CPUs, and to achieve low enough latency for tasks on a busy CPU when
other CPUs would be more suited. BFS has the advantage that it requires no
balancing algorithm whatsoever, as balancing occurs by proxy simply because
all CPUs draw off the global runqueue, in priority and deadline order. Despite
the fact that scalability is _not_ the prime concern of BFS, it both shows very
good scalability to smaller numbers of CPUs and is likely a more scalable design
at these numbers of CPUs.

It also has some very low overhead scalability features built into the design
when it has been deemed their overhead is so marginal that they're worth adding.
The first is the local copy of the running process' data to the CPU it's running
on to allow that data to be updated lockless where possible. Then there is
deference paid to the last CPU a task was running on, by trying that CPU first
when looking for an idle CPU to use the next time it's scheduled. Finally there
is the notion of "sticky" tasks that are flagged when they are involuntarily
descheduled, meaning they still want further CPU time. This sticky flag is
used to bias heavily against those tasks being scheduled on a different CPU
unless that CPU would be otherwise idle. When a cpu frequency governor is used
that scales with CPU load, such as ondemand, sticky tasks are not scheduled
on a different CPU at all, preferring instead to go idle. This means the CPU
they were bound to is more likely to increase its speed while the other CPU
will go idle, thus speeding up total task execution time and likely decreasing
power usage. This is the only scenario where BFS will allow a CPU to go idle
in preference to scheduling a task on the earliest available spare CPU.

The real cost of migrating a task from one CPU to another is entirely dependant
on the cache footprint of the task, how cache intensive the task is, how long
it's been running on that CPU to take up the bulk of its cache, how big the CPU
cache is, how fast and how layered the CPU cache is, how fast a context switch
is... and so on. In other words, it's close to random in the real world where we
do more than just one sole workload. The only thing we can be sure of is that
it's not free. So BFS uses the principle that an idle CPU is a wasted CPU and
utilising idle CPUs is more important than cache locality, and cache locality
only plays a part after that.

When choosing an idle CPU for a waking task, the cache locality is determined
according to where the task last ran and then idle CPUs are ranked from best
to worst to choose the most suitable idle CPU based on cache locality, NUMA
node locality and hyperthread sibling business. They are chosen in the
following preference (if idle):

* Same core, idle or busy cache, idle threads
* Other core, same cache, idle or busy cache, idle threads.
* Same node, other CPU, idle cache, idle threads.
* Same node, other CPU, busy cache, idle threads.
* Same core, busy threads.
* Other core, same cache, busy threads.
* Same node, other CPU, busy threads.
* Other node, other CPU, idle cache, idle threads.
* Other node, other CPU, busy cache, idle threads.
* Other node, other CPU, busy threads.

This shows the SMT or "hyperthread" awareness in the design as well which will
choose a real idle core first before a logical SMT sibling which already has
tasks on the physical CPU.

Early benchmarking of BFS suggested scalability dropped off at the 16 CPU mark.
However this benchmarking was performed on an earlier design that was far less
scalable than the current one so it's hard to know how scalable it is in terms
of both CPUs (due to the global runqueue) and heavily loaded machines (due to
O(n) lookup) at this stage. Note that in terms of scalability, the number of
_logical_ CPUs matters, not the number of _physical_ CPUs. Thus, a dual (2x)
quad core (4X) hyperthreaded (2X) machine is effectively a 16X. Newer benchmark
results are very promising indeed, without needing to tweak any knobs, features
or options. Benchmark contributions are most welcome.


Features

As the initial prime target audience for BFS was the average desktop user, it
was designed to not need tweaking, tuning or have features set to obtain benefit
from it. Thus the number of knobs and features has been kept to an absolute
minimum and should not require extra user input for the vast majority of cases.
There are precisely 2 tunables, and 2 extra scheduling policies. The rr_interval
and iso_cpu tunables, and the SCHED_ISO and SCHED_IDLEPRIO policies. In addition
to this, BFS also uses sub-tick accounting. What BFS does _not_ now feature is
support for CGROUPS. The average user should neither need to know what these
are, nor should they need to be using them to have good desktop behaviour.

rr_interval

There is only one "scheduler" tunable, the round robin interval. This can be
accessed in

	/proc/sys/kernel/rr_interval

The value is in milliseconds, and the default value is set to 6ms. Valid values
are from 1 to 1000. Decreasing the value will decrease latencies at the cost of
decreasing throughput, while increasing it will improve throughput, but at the
cost of worsening latencies. The accuracy of the rr interval is limited by HZ
resolution of the kernel configuration. Thus, the worst case latencies are
usually slightly higher than this actual value. BFS uses "dithering" to try and
minimise the effect the Hz limitation has. The default value of 6 is not an
arbitrary one. It is based on the fact that humans can detect jitter at
approximately 7ms, so aiming for much lower latencies is pointless under most
circumstances. It is worth noting this fact when comparing the latency
performance of BFS to other schedulers. Worst case latencies being higher than
7ms are far worse than average latencies not being in the microsecond range.
Experimentation has shown that rr intervals being increased up to 300 can
improve throughput but beyond that, scheduling noise from elsewhere prevents
further demonstrable throughput.

Isochronous scheduling.

Isochronous scheduling is a unique scheduling policy designed to provide
near-real-time performance to unprivileged (ie non-root) users without the
ability to starve the machine indefinitely. Isochronous tasks (which means
"same time") are set using, for example, the schedtool application like so:

	schedtool -I -e amarok

This will start the audio application "amarok" as SCHED_ISO. How SCHED_ISO works
is that it has a priority level between true realtime tasks and SCHED_NORMAL
which would allow them to preempt all normal tasks, in a SCHED_RR fashion (ie,
if multiple SCHED_ISO tasks are running, they purely round robin at rr_interval
rate). However if ISO tasks run for more than a tunable finite amount of time,
they are then demoted back to SCHED_NORMAL scheduling. This finite amount of
time is the percentage of _total CPU_ available across the machine, configurable
as a percentage in the following "resource handling" tunable (as opposed to a
scheduler tunable):

	/proc/sys/kernel/iso_cpu

and is set to 70% by default. It is calculated over a rolling 5 second average
Because it is the total CPU available, it means that on a multi CPU machine, it
is possible to have an ISO task running as realtime scheduling indefinitely on
just one CPU, as the other CPUs will be available. Setting this to 100 is the
equivalent of giving all users SCHED_RR access and setting it to 0 removes the
ability to run any pseudo-realtime tasks.

A feature of BFS is that it detects when an application tries to obtain a
realtime policy (SCHED_RR or SCHED_FIFO) and the caller does not have the
appropriate privileges to use those policies. When it detects this, it will
give the task SCHED_ISO policy instead. Thus it is transparent to the user.
Because some applications constantly set their policy as well as their nice
level, there is potential for them to undo the override specified by the user
on the command line of setting the policy to SCHED_ISO. To counter this, once
a task has been set to SCHED_ISO policy, it needs superuser privileges to set
it back to SCHED_NORMAL. This will ensure the task remains ISO and all child
processes and threads will also inherit the ISO policy.

Idleprio scheduling.

Idleprio scheduling is a scheduling policy designed to give out CPU to a task
_only_ when the CPU would be otherwise idle. The idea behind this is to allow
ultra low priority tasks to be run in the background that have virtually no
effect on the foreground tasks. This is ideally suited to distributed computing
clients (like setiathome, folding, mprime etc) but can also be used to start
a video encode or so on without any slowdown of other tasks. To avoid this
policy from grabbing shared resources and holding them indefinitely, if it
detects a state where the task is waiting on I/O, the machine is about to
suspend to ram and so on, it will transiently schedule them as SCHED_NORMAL. As
per the Isochronous task management, once a task has been scheduled as IDLEPRIO,
it cannot be put back to SCHED_NORMAL without superuser privileges. Tasks can
be set to start as SCHED_IDLEPRIO with the schedtool command like so:

	schedtool -D -e ./mprime

Subtick accounting.

It is surprisingly difficult to get accurate CPU accounting, and in many cases,
the accounting is done by simply determining what is happening at the precise
moment a timer tick fires off. This becomes increasingly inaccurate as the
timer tick frequency (HZ) is lowered. It is possible to create an application
which uses almost 100% CPU, yet by being descheduled at the right time, records
zero CPU usage. While the main problem with this is that there are possible
security implications, it is also difficult to determine how much CPU a task
really does use. BFS tries to use the sub-tick accounting from the TSC clock,
where possible, to determine real CPU usage. This is not entirely reliable, but
is far more likely to produce accurate CPU usage data than the existing designs
and will not show tasks as consuming no CPU usage when they actually are. Thus,
the amount of CPU reported as being used by BFS will more accurately represent
how much CPU the task itself is using (as is shown for example by the 'time'
application), so the reported values may be quite different to other schedulers.
Values reported as the 'load' are more prone to problems with this design, but
per process values are closer to real usage. When comparing throughput of BFS
to other designs, it is important to compare the actual completed work in terms
of total wall clock time taken and total work done, rather than the reported
"cpu usage".


Con Kolivas <kernel@kolivas.org> Tue, 5 Apr 2011