Re: [PATCH 03/15] sched/fair: Add lag based placement

From: Peter Zijlstra <peterz@infradead.org>
To: Benjamin Segall <bsegall@google.com>
Cc: mingo@kernel.org, vincent.guittot@linaro.org,
	linux-kernel@vger.kernel.org, juri.lelli@redhat.com,
	dietmar.eggemann@arm.com, rostedt@goodmis.org, mgorman@suse.de,
	bristot@redhat.com, corbet@lwn.net, qyousef@layalina.io,
	chris.hyser@oracle.com, patrick.bellasi@matbug.net,
	pjt@google.com, pavel@ucw.cz, qperret@google.com,
	tim.c.chen@linux.intel.com, joshdon@google.com, timj@gnu.org,
	kprateek.nayak@amd.com, yu.c.chen@intel.com,
	youssefesmat@chromium.org, joel@joelfernandes.org, efault@gmx.de,
	tglx@linutronix.de
Subject: Re: [PATCH 03/15] sched/fair: Add lag based placement
Date: Fri, 13 Oct 2023 16:34:14 +0200	[thread overview]
Message-ID: <20231013143414.GA36211@noisy.programming.kicks-ass.net> (raw)
In-Reply-To: <xm26fs2fhcu7.fsf@bsegall-linux.svl.corp.google.com>

On Thu, Oct 12, 2023 at 12:15:12PM -0700, Benjamin Segall wrote:
> Peter Zijlstra <peterz@infradead.org> writes:
> 
> > @@ -4853,49 +4872,119 @@ static void
> >  place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
> >  {
> >  	u64 vruntime = avg_vruntime(cfs_rq);
> > +	s64 lag = 0;
> >  
> > -	/* sleeps up to a single latency don't count. */
> > -	if (!initial) {
> > -		unsigned long thresh;
> > +	/*
> > +	 * Due to how V is constructed as the weighted average of entities,
> > +	 * adding tasks with positive lag, or removing tasks with negative lag
> > +	 * will move 'time' backwards, this can screw around with the lag of
> > +	 * other tasks.
> > +	 *
> > +	 * EEVDF: placement strategy #1 / #2
> > +	 */
> 
> So the big problem with EEVDF #1 compared to #2/#3 and CFS (hacky though
> it is) is that it creates a significant perverse incentive to yield or
> spin until you see yourself be preempted, rather than just sleep (if you
> have any competition on the cpu). If you go to sleep immediately after
> doing work and happen to do so near the end of a slice (arguably what
> you _want_ to have happen overall), then you have to pay that negative
> lag in wakeup latency later, because it is maintained through any amount
> of sleep. (#1 or similar is good for reweight/migrate of course)
> 
> #2 in theory could be abused by micro-sleeping right before you are
> preempted, but that isn't something tasks can really predict, unlike
> seeing more "don't go to sleep, just spin, the latency numbers are so
> much better" nonsense.

For giggles (cyclictest vs hackbench):

$ echo PLACE_LAG > /debug/sched/features
$ ./doit-latency-slice.sh
# Running 'sched/messaging' benchmark:
slice 30000000
# /dev/cpu_dma_latency set to 0us
# Min Latencies: 00051
# Avg Latencies: 00819
# Max Latencies: 172558
slice 3000000
# /dev/cpu_dma_latency set to 0us
# Min Latencies: 00033
# Avg Latencies: 00407
# Max Latencies: 12024
slice 300000
# /dev/cpu_dma_latency set to 0us
# Min Latencies: 00055
# Avg Latencies: 00395
# Max Latencies: 11780

$ echo NO_PLACE_LAG > /debug/sched/features
$ ./doit-latency-slice.sh
# Running 'sched/messaging' benchmark:
slice 30000000
# /dev/cpu_dma_latency set to 0us
# Min Latencies: 00069
# Avg Latencies: 69071
# Max Latencies: 1492250
slice 3000000
# /dev/cpu_dma_latency set to 0us
# Min Latencies: 00062
# Avg Latencies: 10215
# Max Latencies: 21209
slice 300000
# /dev/cpu_dma_latency set to 0us
# Min Latencies: 00055
# Avg Latencies: 00060
# Max Latencies: 03088

IOW, insanely worse latencies in most cases. This is because when
everybody starts at 0-lag, everybody is always eligible, and 'fairness'
goes out the window fast.

Placement strategy #1 only really works when you have well behaving
tasks (eg. conforming to the periodic task model -- not waking up before
its time and all that).