Message-ID: <20230411020945.GA65214@maniforge>
Date:   Mon, 10 Apr 2023 21:09:45 -0500
From:   David Vernet <void@...ifault.com>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     mingo@...nel.org, vincent.guittot@...aro.org,
        linux-kernel@...r.kernel.org, juri.lelli@...hat.com,
        dietmar.eggemann@....com, rostedt@...dmis.org, bsegall@...gle.com,
        mgorman@...e.de, bristot@...hat.com, corbet@....net,
        qyousef@...alina.io, chris.hyser@...cle.com,
        patrick.bellasi@...bug.net, pjt@...gle.com, pavel@....cz,
        qperret@...gle.com, tim.c.chen@...ux.intel.com, joshdon@...gle.com,
        timj@....org, kprateek.nayak@....com, yu.c.chen@...el.com,
        youssefesmat@...omium.org, joel@...lfernandes.org, efault@....de,
        clm@...a.com, tj@...nel.org, djk@...a.com
Subject: Re: [PATCH 00/17] sched: EEVDF using latency-nice

On Sun, Apr 09, 2023 at 10:13:50PM -0500, David Vernet wrote:
> On Tue, Mar 28, 2023 at 11:26:22AM +0200, Peter Zijlstra wrote:
> > Hi!
> > 
> > Latest version of the EEVDF [1] patches.
> > 
> > Many changes since last time; most notably it now fully replaces CFS and uses
> > lag based placement for migrations. Smaller changes include:
> > 
> >  - uses scale_load_down() for avg_vruntime; I measured the max delta to be ~44
> >    bits on a system/cgroup based kernel build.
> >  - fixed a bunch of reweight / cgroup placement issues
> >  - adaptive placement strategy for smaller slices
> >  - rename se->lag to se->vlag
> > 
> > There's a bunch of RFC patches at the end and one DEBUG patch. Of those, the
> > PLACE_BONUS patch is a mixed bag of pain. A number of benchmarks regress
> > because EEVDF is actually fair and gives a 100% parent vs a 50% child a 67%/33%
> > split (stress-futex, stress-nanosleep, starve, etc..) instead of a 50%/50%
> > split that sleeper bonus achieves. Mostly I think these benchmarks are somewhat
> > artificial/daft but who knows.
> > 
> > The PLACE_BONUS thing horribly messes up things like hackbench and latency-nice
> > because it places things too far to the left in the tree. Basically it messes
> > with the whole 'when', by placing a task back in history you're putting a
> > burden on the now to accomodate catching up. More tinkering required.
> > 
> > But over-all the thing seems to be fairly usable and could do with more
> > extensive testing.
> 
> Hi Peter,
> 
> I used the EEVDF scheduler to run workloads on one of Meta's largest
> services (our main HHVM web server), and I wanted to share my
> observations with you.
> 
> The TL;DR is that, unfortunately, it appears as though EEVDF regresses
> these workloads rather substantially. Running with "vanilla" EEVDF (i.e.
> on this patch set up to [0], with no changes to latency nice for any
> task) compared to vanilla CFS results in the following outcomes to our
> major KPIs for servicing web requests:
> 
> - .5 - 2.5% drop in throughput
> - 1 - 5% increase in p95 latencies
> - 1.75 - 6% increase in p99 latencies
> - .5 - 4% drop in 50th percentile throughput
> 
> [0]: https://lore.kernel.org/lkml/20230328110354.562078801@infradead.org/
> 
> Decreasing latency nice for our critical web workers unfortunately did
> not help either. For example, here are the numbers for a latency nice
> value of -10:
> 
> - .8 - 2.5% drop in throughput
> - 2 - 4% increase in p95 latencies
> - 1 - 4.5% increase in p99 latencies
> - 0 - 4.5% increase in 50th percentile throughput
> 
> Other latency nice values resulted in similar metrics. Some metrics may
> get slightly better, and others slightly worse, but the end result was
> always a relatively significant regression from vanilla CFS. Throughout
> the rest of this write up, the remaining figures quoted will be from
> vanilla EEVDF runs (modulo some numbers towards the end of this writeup
> which describe the outcome of increasing the default base slice with
> sysctl_sched_base_slice).
> 
> With that out the way, let me outline some of the reasons for these
> regressions:
> 
> 1. Improved runqueue delays, but costly faults and involuntary context
>    switches
> 
> EEVDF substantially increased the number of context switches on the
> system, by 15 - 35%. On its own, this doesn't necessarily imply a
> problem. For example, we observed that EEVDF resulted in a 20 - 40%
> reduction in the time that tasks were spent waiting on the runqueue
> before being placed on a CPU.
> 
> There were, however, other metrics which were less encouraging. We
> observed a 400 - 550% increase in involuntary context switches (which
> are also presumably a reason for the improved runqueue delays), as well
> as a 10 - 70% increase in major page faults per minute. Along these
> lines, we also saw an erratic but often sigificant decrease in CPU
> utilization.
> 
> It's hard to say exactly what kinds of issues such faults / involuntary
> context context switches could introduce, but it does seem clear that in
> general, less time is being spent doing useful work, and more time is
> spent thrashing on resources between tasks.
> 
> 2. Front-end CPU pipeline bottlenecks
> 
> Like many (any?) other JIT engines / compilers, HHVM tends to be heavily
> front-end bound in the CPU pipeline, and have very poor IPC
> (Instructions Per Cycle). For HHVM, this is due to high branch resteers,
> poor icache / iTLB locality, and poor uop caching / decoding (many uops
> are being serviced through the MITE instead of the DSB). While using
> BOLT [1] to improve the layout of the HHVM binary does help to minimize
> these costs, they're still the main bottleneck for the application.
> 
> [1]: https://github.com/llvm/llvm-project/blob/main/bolt/docs/OptimizingClang.md
> 
> An implication of this is that any time a task runs on a CPU after one
> of these web worker tasks, it is essentially guaranteed to have poor
> front-end locality, and their IPCs will similarly suffer. In other
> words, more context switches generally means fewer instructions being
> run across the whole system. When profiling vanilla CFS vs. vanilla
> EEVDF (that is, with no changes to latency nice for any task), we found
> that it resulted in a 1 - 2% drop in IPC across the whole system.
> 
> Note that simply partitioning the system by cpuset won't necessarily
> work either, as CPU utilization will drop even further, and we want to
> keep the system as busy as possible. There's another internal patch set
> we have (which we're planning to try and upstream soon) where waking
> tasks are placed in a global shared "wakequeue", which is then always
> pulled from in newidle_balance(). The discrepancy in performance between
> CFS and EEVDF is even worse in this case, with throughput dropping by 2
> - 4%, p95 tail latencies increasing by 3 - 5%, and p99 tail latencies
> increasing by 6 - 11%.
> 
> 3. Low latency + long slice are not mutually exclusive for us
> 
> An interesting quality of web workloads running JIT engines is that they
> require both low-latency, and long slices on the CPU. The reason we need
> the tasks to be low latency is they're on the critical path for
> servicing web requests (for most of their runtime, at least), and the
> reasons we need them to have long slices are enumerated above -- they
> thrash the icache / DSB / iTLB, more aggressive context switching causes
> us to thrash on paging from disk, and in general, these tasks are on the
> critical path for servicing web requests and we want to encourage them
> to run to completion.
> 
> This causes EEVDF to perform poorly for workloads with these
> characteristics. If we decrease latency nice for our web workers then
> they'll have lower latency, but only because their slices are smaller.
> This in turn causes the increase in context switches, which causes the
> thrashing described above.
> 
> Worth noting -- I did try and increase the default base slice length by
> setting sysctl_sched_base_slice to 35ms, and these were the results:
> 
> With EEVDF slice 35ms and latency_nice 0
> ----------------------------------------
> - .5 - 2.25% drop in throughput
> - 2.5 - 4.5% increase in p95 latencies
> - 2.5 - 5.25% increase in p99 latencies
> - Context switch per minute increase: 9.5 - 12.4%
> - Involuntary context switch increase: ~320 - 330%
> - Major fault delta: -3.6% to 37.6%
> - IPC decrease .5 - .9%
> 
> With EEVDF slice 35ms and latency_nice -8 for web workers
> ---------------------------------------------------------
> - .5 - 2.5% drop in throughput
> - 1.7 - 4.75% increase in p95 latencies
> - 2.5 - 5% increase in p99 latencies
> - Context switch per minute increase: 10.5 - 15%
> - Involuntary context switch increase: ~327 - 350%
> - Major fault delta: -1% to 45%
> - IPC decrease .4 - 1.1%
> 
> I was expecting the increase in context switches and involuntary context
> switches to be lower what than they ended up being with the increased
> default slice length. Regardless, it still seems to tell a relatively
> consistent story with the numbers from above. The improvement in IPC is
> expected, though also less improved than I was anticipating (presumably
> due to the still-high context switch rate). There were also fewer major
> faults per minute compared to runs with a shorter default slice.

Ah, these numbers with a larger slice are inaccurate. I was being
careless and accidentally changed the wrong sysctl
(sysctl_sched_cfs_bandwidth_slice rather than sysctl_sched_base_slice)
yesterday when testing how a longer slice affects performance.
Increasing sysctl_sched_base_slice to 30ms actually does improve things
a bit, but we're still losing to CFS (note that higher is better for
throughput, and lower is better for p95 and p99 latency):

lat_nice | throughput     | p95 latency  | p99 latency   | total swtch    | invol swtch  | mjr faults | IPC
-----------------------------------------------------------------------------------------------------------------------
0        | -1.2% to 0%    | 1 to 2.5%    | 1.8 to 3.75%  | 0%             | 200 to 210%  | -3% to 14% | -0.4% to -0.25%
-----------------------------------------------------------------------------------------------------------------------
-8       | -1.8% to -1.3% | 1 to 2.3%    | .75% to 2.5%  | -1.9% to -1.3% | 185 to 193%  | -3% to 20% | -0.6% to -0.25%
-----------------------------------------------------------------------------------------------------------------------
-12      | -.9% to 0.25%  | -0.1 to 2.4% | -0.4% to 2.8% | -2% to -1.5%   | 180 to 185%  | -5% to 30% | -0.8% to -0.25%
-----------------------------------------------------------------------------------------------------------------------
-19      | -1.3% to 0%   | 0.3 to 3.5%  | -1% to 2.1%   | -2% to -1.4%   | 175 to 190%   | 4% to 27%  | -0.6% to -0.27%
-----------------------------------------------------------------------------------------------------------------------

I'm sure experimenting with various slice lengths, etc would yield
slightly different results, but the common theme seems to be that EEVDF
causes more involuntary context switches and more major faults, which
regress throughput and latency for our web workloads.

> 
> Note that even if increasing the slice length did cause fewer context
> switches and major faults, I still expect that it would hurt throughput
> and latency for HHVM given that when latency-nicer tasks are eventually
> given the CPU, the web workers will have to wait around for longer than
> we'd like for those tasks to burn through their longer slices.
> 
> In summary, I must admit that this patch set makes me a bit nervous.
> Speaking for Meta at least, the patch set in its current form exceeds
> the performance regressions (generally < .5% at the very most) that
> we're able to tolerate in production. More broadly, it will certainly
> cause us to have to carefully consider how it affects our model for
> server capacity.
> 
> Thanks,
> David

Powered by blists - more mailing lists