WSL2-Linux-Kernel/Documentation/sched-design-CFS.txt

120 строки
5.8 KiB
Plaintext

This is the CFS scheduler.
80% of CFS's design can be summed up in a single sentence: CFS basically
models an "ideal, precise multi-tasking CPU" on real hardware.
"Ideal multi-tasking CPU" is a (non-existent :-)) CPU that has 100%
physical power and which can run each task at precise equal speed, in
parallel, each at 1/nr_running speed. For example: if there are 2 tasks
running then it runs each at 50% physical power - totally in parallel.
On real hardware, we can run only a single task at once, so while that
one task runs, the other tasks that are waiting for the CPU are at a
disadvantage - the current task gets an unfair amount of CPU time. In
CFS this fairness imbalance is expressed and tracked via the per-task
p->wait_runtime (nanosec-unit) value. "wait_runtime" is the amount of
time the task should now run on the CPU for it to become completely fair
and balanced.
( small detail: on 'ideal' hardware, the p->wait_runtime value would
always be zero - no task would ever get 'out of balance' from the
'ideal' share of CPU time. )
CFS's task picking logic is based on this p->wait_runtime value and it
is thus very simple: it always tries to run the task with the largest
p->wait_runtime value. In other words, CFS tries to run the task with
the 'gravest need' for more CPU time. So CFS always tries to split up
CPU time between runnable tasks as close to 'ideal multitasking
hardware' as possible.
Most of the rest of CFS's design just falls out of this really simple
concept, with a few add-on embellishments like nice levels,
multiprocessing and various algorithm variants to recognize sleepers.
In practice it works like this: the system runs a task a bit, and when
the task schedules (or a scheduler tick happens) the task's CPU usage is
'accounted for': the (small) time it just spent using the physical CPU
is deducted from p->wait_runtime. [minus the 'fair share' it would have
gotten anyway]. Once p->wait_runtime gets low enough so that another
task becomes the 'leftmost task' of the time-ordered rbtree it maintains
(plus a small amount of 'granularity' distance relative to the leftmost
task so that we do not over-schedule tasks and trash the cache) then the
new leftmost task is picked and the current task is preempted.
The rq->fair_clock value tracks the 'CPU time a runnable task would have
fairly gotten, had it been runnable during that time'. So by using
rq->fair_clock values we can accurately timestamp and measure the
'expected CPU time' a task should have gotten. All runnable tasks are
sorted in the rbtree by the "rq->fair_clock - p->wait_runtime" key, and
CFS picks the 'leftmost' task and sticks to it. As the system progresses
forwards, newly woken tasks are put into the tree more and more to the
right - slowly but surely giving a chance for every task to become the
'leftmost task' and thus get on the CPU within a deterministic amount of
time.
Some implementation details:
- the introduction of Scheduling Classes: an extensible hierarchy of
scheduler modules. These modules encapsulate scheduling policy
details and are handled by the scheduler core without the core
code assuming about them too much.
- sched_fair.c implements the 'CFS desktop scheduler': it is a
replacement for the vanilla scheduler's SCHED_OTHER interactivity
code.
I'd like to give credit to Con Kolivas for the general approach here:
he has proven via RSDL/SD that 'fair scheduling' is possible and that
it results in better desktop scheduling. Kudos Con!
The CFS patch uses a completely different approach and implementation
from RSDL/SD. My goal was to make CFS's interactivity quality exceed
that of RSDL/SD, which is a high standard to meet :-) Testing
feedback is welcome to decide this one way or another. [ and, in any
case, all of SD's logic could be added via a kernel/sched_sd.c module
as well, if Con is interested in such an approach. ]
CFS's design is quite radical: it does not use runqueues, it uses a
time-ordered rbtree to build a 'timeline' of future task execution,
and thus has no 'array switch' artifacts (by which both the vanilla
scheduler and RSDL/SD are affected).
CFS uses nanosecond granularity accounting and does not rely on any
jiffies or other HZ detail. Thus the CFS scheduler has no notion of
'timeslices' and has no heuristics whatsoever. There is only one
central tunable:
/proc/sys/kernel/sched_granularity_ns
which can be used to tune the scheduler from 'desktop' (low
latencies) to 'server' (good batching) workloads. It defaults to a
setting suitable for desktop workloads. SCHED_BATCH is handled by the
CFS scheduler module too.
Due to its design, the CFS scheduler is not prone to any of the
'attacks' that exist today against the heuristics of the stock
scheduler: fiftyp.c, thud.c, chew.c, ring-test.c, massive_intr.c all
work fine and do not impact interactivity and produce the expected
behavior.
the CFS scheduler has a much stronger handling of nice levels and
SCHED_BATCH: both types of workloads should be isolated much more
agressively than under the vanilla scheduler.
( another detail: due to nanosec accounting and timeline sorting,
sched_yield() support is very simple under CFS, and in fact under
CFS sched_yield() behaves much better than under any other
scheduler i have tested so far. )
- sched_rt.c implements SCHED_FIFO and SCHED_RR semantics, in a simpler
way than the vanilla scheduler does. It uses 100 runqueues (for all
100 RT priority levels, instead of 140 in the vanilla scheduler)
and it needs no expired array.
- reworked/sanitized SMP load-balancing: the runqueue-walking
assumptions are gone from the load-balancing code now, and
iterators of the scheduling modules are used. The balancing code got
quite a bit simpler as a result.