From: Con Kolivas <kernel@kolivas.org>
To: linux-kernel <linux-kernel@vger.kernel.org>
Subject: [ANNOUNCE] Multiple Queue Skiplist Scheduler version 0.120
Date: Sat, 29 Oct 2016 16:41:56 +1100
Message-ID: <6544690.h1r1hWEHvO@hex>
User-Agent: KMail/5.1.3 (Linux/4.8.5-cktest+; KDE/5.18.0; x86_64; ; )
MIME-Version: 1.0
Content-Transfer-Encoding: 8BIT
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 21372
Lines: 431

This is to announce the first major ~stable public release of MuQSS (pronounced 
mux), the Multiple Queue Skiplist Scheduler. 

MuQSS for linux 4.8:
http://ck.kolivas.org/patches/muqss/4.0/4.8/4.8-sched-MuQSS_120.patch

MuQSS for linux 4.7:
http://ck.kolivas.org/patches/muqss/4.0/4.7/4.7-sched-MuQSS_120.patch

Git tree:
https://github.com/ckolivas/linux


MuQSS is the evolution and drop-in replacement for BFS, the Brain Fuck 
Scheduler, both of which were designed with responsiveness and interactivity 
as their primary goal. MuQSS is my response for requests for a more scalable 
version of BFS for our ever-increasing multicore hardware. It is a massive 
rewrite of BFS designed to maintain the same interactivity and responsiveness 
using the same algorithm for scheduling decisions, along with a very simple 
overall design that is easy to understand, model, and hack on, yet able to 
scale to hardware of virtually any size.

It is meant as a replacement for BFS primarily, and NOT a comprehensive 
replacement for the mainline scheduler since it lacks some of the features of 
mainline still (specifically cgroups and sched deadline support.) However 
outside of these feature requirements it performs better in latency and 
responsiveness areas and, while it performs poorer in some throughput 
benchmarks, it also performs better in many others. Additionally MuQSS has a 
number of unique features missing from mainline described in the full 
documentation below. Results from early 12x benchmarks below.

Interbench interactivity comparison:
http://ck-hack.blogspot.com/2016/10/interbench-benchmarks-for-muqss-116.html

Throughput comparison:
http://ck.kolivas.org/patches/muqss/Benchmarks/20161018/

Very little effort has gone into optimising throughput in its existing form, 
being effectively a first stable working version of the scheduler, though 
clearly it does well in some areas (postgres particularly.)

To be clear what my intentions are, this will be maintained out of mainline 
development - as BFS was - for those that wish an alternative scheduler with a 
different focus and set of features. There is plenty of interest in this 
outside of mainline development and there are some perverse things I do in the 
code that would make most mainline developers' skin crawl, but application and 
utility are far more important than technology in my opinion. I do not have 
the energy and time to engage with the mainline development process alone to 
try and make this replace the existing scheduler though I'm always happy to 
accept outside help and patches for this project.

Below is the complete documentation as included in the patch.

Enjoy!
お楽しみください
-- 
-ck

MuQSS - The Multiple Queue Skiplist Scheduler by Con Kolivas.

MuQSS is a per-cpu runqueue variant of the original BFS scheduler with
one 8 level skiplist per runqueue, and fine grained locking for much more
scalability.


Goals.

The goal of the Multiple Queue Skiplist Scheduler, referred to as MuQSS from
here on (pronounced mux) 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 MuQSS 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, while still being
scalable to many CPUs and processes.


Design summary.

MuQSS is best described as per-cpu multiple runqueue, O(log n) insertion, O(1)
lookup, earliest effective virtual deadline first tickless design, loosely 
based
on EEVDF (earliest eligible virtual deadline first) and my previous Staircase
Deadline scheduler, and evolved from the single runqueue O(n) BFS scheduler.
Each component shall be described in order to understand the significance of,
and reasoning for it.


Design reasoning.

In BFS, the use of a single runqueue across all CPUs meant that each CPU would
need to scan the entire runqueue looking for the process with the earliest
deadline and schedule that next, regardless of which CPU it originally came
from. This made BFS deterministic with respect to latency and provided
guaranteed latencies dependent on number of processes and CPUs. The single
runqueue, however, meant that all CPUs would compete for the single lock
protecting it, which would lead to increasing lock contention as the number of
CPUs rose and appeared to limit scalability of common workloads beyond 16
logical CPUs. Additionally, the O(n) lookup of the runqueue list obviously
increased overhead proportionate to the number of queued proecesses and led to
cache thrashing while iterating over the linked list.

MuQSS is an evolution of BFS, designed to maintain the same scheduling
decision mechanism and be virtually deterministic without relying on the
constrained design of the single runqueue by splitting out the single runqueue
to be per-CPU and use skiplists instead of linked lists.

The original reason for going back to a single runqueue design for BFS was 
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.

MuQSS works around the problems inherent in multiple runqueue designs by
making its skip lists priority ordered and through novel use of lockless
examination of each other runqueue it can decide if it should take the 
earliest
deadline task from another runqueue for latency reasons, or for CPU balancing
reasons. It still does not have a balancing system, choosing to allow the
next task scheduling decision and task wakeup CPU choice to allow balancing to
happen by virtue of its choices.


Design details.

Custom skip list implementation:

To avoid the overhead of building up and tearing down skip list structures,
the variant used by MuQSS has a number of optimisations making it specific for
its use case in the scheduler. It uses static arrays of 8 'levels' instead of
building up and tearing down structures dynamically. This makes each runqueue
only scale O(log N) up to 256 tasks. However as there is one runqueue per CPU
it means that it scales O(log N) up to 256 x number of logical CPUs which is
far beyond the realistic task limits each CPU could handle. By being 8 levels
it also makes the array exactly one cacheline in size. Additionally, each
skip list node is bidirectional making insertion and removal amortised O(1),
being O(k) where k is 1-8. Uniquely, we are only ever interested in the very
first entry in each list at all times with MuQSS, so there is never a need to
do a search and thus look up is always O(1).

Task insertion:

MuQSS inserts tasks into a per CPU runqueue as an O(log N) insertion into
a custom skip list as described above (based on the original design by William
Pugh). Insertion is ordered in such a way that there is never a need to do a
search by ordering tasks according to static priority primarily, and then
virtual deadline at the time of insertion.

Niffies:

Niffies are a monotonic forward moving timer not unlike the "jiffies" but are
of nanosecond resolution. Niffies are calculated per-runqueue from the high
resolution TSC timers, and in order to maintain fairness are synchronised
between CPUs whenever both runqueues are locked concurrently.

Virtual deadline:

The key to achieving low latency, scheduling fairness, and "nice level"
distribution in MuQSS is entirely in the virtual deadline mechanism. The one
tunable in MuQSS 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 niffies by this equation:

	niffies + (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.

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" (niffies) is
constantly moving.

Task lookup:

As tasks are already pre-ordered according to anticipated scheduling order in
the skip lists, lookup for the next suitable task per-runqueue is always a
matter of simply selecting the first task in the 0th level skip list entry.
In order to maintain optimal latency and fairness across CPUs, MuQSS does a
novel examination of every other runqueue in cache locality order, choosing 
the
best task across all runqueues. This provides near-determinism of how long any
task across the entire system may wait before receiving CPU time. The other
runqueues are first examine lockless and then trylocked to minimise the
potential lock contention if they are likely to have a suitable better task.
Each other runqueue lock is only held for as long as it takes to examine the
entry for suitability. In "interactive" mode, the default setting, MuQSS will
look for the best deadline task across all CPUs, while in !interactive mode,
it will only select a better deadline task from another CPU if it is more
heavily laden than the current one.

Lookup is therefore O(k) where k is number of CPUs.


Latency.

Through the use of virtual deadlines to govern the scheduling order of normal
tasks, queue-to-activation latency per runqueue is guaranteed to be bound by
the rr_interval tunable which is set to 6ms by default. This means that the
longest a CPU bound task will wait for more CPU is proportional to the number
of running tasks and in the common case of 0-2 running tasks per CPU, will be
under the 7ms threshold for human perception of jitter. Additionally, as newly
woken tasks will have an early deadline from their previous runtime, the very
tasks that are usually latency sensitive will have the shortest interval for
activation, usually preempting any existing CPU bound tasks.

Tickless expiry:

A feature of MuQSS is that it is not tied to the resolution of the chosen tick
rate in Hz, instead depending entirely on the high resolution timers where
possible for sub-millisecond accuracy on timeouts regarless of the underlying
tick rate. This allows MuQSS to be run with the low overhead of low Hz rates
such as 100 by default, benefiting from the improved throughput and lower
power usage it provides. Another advantage of this approach is that in
combination with the Full No HZ option, which disables ticks on running task
CPUs instead of just idle CPUs, the tick can be disabled at all times
regardless of how many tasks are running instead of being limited to just one
running task. Note that this option is NOT recommended for regular desktop
users.


Scalability and balancing.

Unlike traditional approaches where balancing is a combination of CPU 
selection
at task wakeup and intermittent balancing based on a vast array of rules set
according to architecture, busyness calculations and special case management,
MuQSS indirectly balances on the fly at task wakeup and next task selection.
During initialisation, MuQSS creates a cache coherency ordered list of CPUs 
for
each logical CPU and uses this to aid task/CPU selection when CPUs are busy.
Additionally it selects any idle CPUs, if they are available, at any time over
busy CPUs according to the following preference:

 * Same thread, idle or busy cache, idle or busy 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.
 * 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.

Mux is therefore SMT, MC and Numa aware without the need for extra
intermittent balancing to maintain CPUs busy and make the most of cache
coherency.


Features

As the initial prime target audience for MuQSS 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 3 optional tunables, and 2 extra scheduling policies. The 
rr_interval,
interactive, and iso_cpu tunables, and the SCHED_ISO and SCHED_IDLEPRIO
policies. In addition to this, MuQSS also uses sub-tick accounting. What MuQSS
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. However since some applications refuse to work without
cgroups, one can enable them with MuQSS as a stub and the filesystem will be
created which will allow the applications to work.

rr_interval:

	/proc/sys/kernel/rr_interval

The value is in milliseconds, and the default value is set to 6. 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. 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 MuQSS to other schedulers. Worst case latencies being
higher than 7ms are far worse than average latencies not being in the
microsecond range.

interactive:

	/proc/sys/kernel/interactive

The value is a simple boolean of 1 for on and 0 for off and is set to on by
default. Disabling this will disable the near-determinism of MuQSS when
selecting the next task by not examining all CPUs for the earliest deadline
task, or which CPU to wake to, instead prioritising CPU balancing for improved
throughput. Latency will still be bound by rr_interval, but on a per-CPU basis
instead of across the whole system.

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 CPU available per CPU, configurable as a percentage 
in
the following "resource handling" tunable (as opposed to a scheduler tunable):

iso_cpu:

	/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 MuQSS 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.


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. Once a task has
been scheduled as IDLEPRIO, it cannot be put back to SCHED_NORMAL without
superuser privileges since it is effectively a lower scheduling policy. 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. Mux uses sub-tick accounting from the TSC clock to determine real CPU
usage. Thus, the amount of CPU reported as being used by MuQSS 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. When comparing throughput of MuQSS 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".

Symmetric MultiThreading (SMT) aware nice:

SMT, a.k.a. hyperthreading, is a very common feature on modern CPUs. While the
logical CPU count rises by adding thread units to each CPU core, allowing more
than one task to be run simultaneously on the same core, the disadvantage of 
it
is that the CPU power is shared between the tasks, not summating to the power
of two CPUs. The practical upshot of this is that two tasks running on
separate threads of the same core run significantly slower than if they had one
core each to run on. While smart CPU selection allows each task to have a core
to itself whenever available (as is done on MuQSS), it cannot offset the
slowdown that occurs when the cores are all loaded and only a thread is left.
Most of the time this is harmless as the CPU is effectively overloaded at this
point and the extra thread is of benefit. However when running a niced task in
the presence of an un-niced task (say nice 19 v nice 0), the nice task gets
precisely the same amount of CPU power as the unniced one. MuQSS has an
optional configuration feature known as SMT-NICE which selectively idles the
secondary niced thread for a period proportional to the nice difference,
allowing CPU distribution according to nice level to be maintained, at the
expense of a small amount of extra overhead. If this is configured in on a
machine without SMT threads, the overhead is minimal.


Con Kolivas <kernel@kolivas.org> Sat, 29th October 2016