| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <9df22ebbc2e6d426099bf380477a0ed885068896.camel@gmx.de>
Date: Thu, 17 Oct 2024 14:36:52 +0200
From: Mike Galbraith <efault@....de>
To: Peter Zijlstra <peterz@...radead.org>, bigeasy@...utronix.de,
tglx@...utronix.de, mingo@...nel.org
Cc: linux-kernel@...r.kernel.org, juri.lelli@...hat.com,
vincent.guittot@...aro.org, dietmar.eggemann@....com, rostedt@...dmis.org,
bsegall@...gle.com, mgorman@...e.de, vschneid@...hat.com,
ankur.a.arora@...cle.com
Subject: Re: [PATCH 0/5] sched: Lazy preemption muck
On Mon, 2024-10-07 at 09:46 +0200, Peter Zijlstra wrote:
> Hi!
>
> During LPC Thomas reminded me that the lazy preemption stuff was not there yet.
>
> So here goes, robot says it builds, and I checked both a regular and PREEMPT_RT
> build boots and can change the mode.
>
> Please have a poke.
Lazy seems to be the least sleeper friendly, and voluntary the most
switch happy of the three models tested (a bit counter-intuitive). Not
sure I should post these results, nobody will ever notice these deltas
in their desktop, but they are carefully gathered cold hard numbers, so
what the heck, deleting them takes only a moment.
All data collected against virgin master with only this patch set then
applied, ie there are zero deltas other than the lazy set and model
selection. The load profiled is my favorite completely hands off mixed
compute vs chrome desktop load, 3 runs of lazy and full compared to 2
runs of voluntary, I called that good enough after looking, not being
willing to do any more of what I have done oodles of in the past days.
I found sum delay deltas interesting.. and sensible after pondering,
nothing being free, delaying preemption of wider bodied compute tasks
has to impact somewhere. Should desktop components become very busy,
they'll be doing the impacting.
Full per run summaries attached, high speed scroll version below.
desktop util 18.1% static voluntary - virgin source
desktop util 18.3% static voluntary - +lazy patches
desktop util 17.5% lazy - ditto...
desktop util 17.0% lazy
desktop util 16.8% lazy
desktop util 17.8% full
desktop util 17.8% full
desktop util 17.8% full
Profile summary bottom lines, same order (ignore bottom line delay max, includes nice tasks)
----------------------------------------------------------------------------------------------------------
Task | Runtime ms | Switches | Avg delay ms | Max delay ms | Sum delay ms |
----------------------------------------------------------------------------------------------------------
voluntary
TOTAL: |2281560.794 ms | 18514617 | | 60.836 ms | 1101651.105 ms |
TOTAL: |2280493.511 ms | 18679820 | | 49.278 ms | 1117078.210 ms |
lazy avg 1109364.657 1.000
TOTAL: |2302856.995 ms | 15141600 | | 68.303 ms | 1316229.527 ms |
TOTAL: |2299911.697 ms | 15979830 | | 29.361 ms | 1276128.457 ms |
TOTAL: |2298180.079 ms | 16665916 | | 34.918 ms | 1271137.849 ms |
full avg 1287831.944 1.160
TOTAL: |2279751.696 ms | 16628663 | | 36.929 ms | 977433.245 ms |
TOTAL: |2278306.187 ms | 15846728 | | 79.954 ms | 974772.318 ms |
TOTAL: |2282262.311 ms | 17152068 | | 28.004 ms | 990068.590 ms |
avg 980758.051 .884
Detail extract for the 3 top players:
voluntary
massive_intr:(9) |1869421.285 ms | 5752848 | avg: 0.056 ms | max: 23.502 ms | sum:324608.328 ms |
massive_intr:(9) |1864362.822 ms | 5606194 | avg: 0.059 ms | max: 20.343 ms | sum:328920.012 ms |
avg 326764.170 1.000
dav1d-worker:(8) | 155700.590 ms | 1091524 | avg: 0.300 ms | max: 25.215 ms | sum:327735.383 ms |
dav1d-worker:(8) | 160730.867 ms | 1123134 | avg: 0.299 ms | max: 26.783 ms | sum:335553.750 ms |
avg 331644.566 1.000
X:2534 | 51690.650 ms | 223470 | avg: 0.122 ms | max: 10.367 ms | sum:27371.213 ms |
X:2522 | 51355.479 ms | 232615 | avg: 0.121 ms | max: 14.223 ms | sum:28243.127 ms |
avg 27807.170 1.000
lazy
massive_intr:(9) |1900362.729 ms | 4443987 | avg: 0.066 ms | max: 23.997 ms | sum:294806.423 ms |
massive_intr:(9) |1909494.321 ms | 4624111 | avg: 0.061 ms | max: 20.215 ms | sum:282640.873 ms |
massive_intr:(9) |1913482.381 ms | 4896729 | avg: 0.057 ms | max: 19.473 ms | sum:277651.582 ms |
avg 285032.959 .872
dav1d-worker:(8) | 164649.511 ms | 849691 | avg: 0.531 ms | max: 21.537 ms | sum:451592.116 ms |
dav1d-worker:(8) | 151579.745 ms | 850584 | avg: 0.505 ms | max: 29.119 ms | sum:429540.789 ms |
dav1d-worker:(8) | 145472.762 ms | 860714 | avg: 0.489 ms | max: 22.005 ms | sum:420927.393 ms |
avg 434020.099 1.308
X:2508 | 53119.214 ms | 317117 | avg: 0.135 ms | max: 11.995 ms | sum:42917.248 ms |
X:2508 | 53047.311 ms | 311231 | avg: 0.137 ms | max: 17.974 ms | sum:42601.534 ms |
X:2508 | 48692.053 ms | 333137 | avg: 0.129 ms | max: 13.263 ms | sum:43140.962 ms |
avg 42886.581 1.542
full
massive_intr:(9) |1875426.524 ms | 5559171 | avg: 0.057 ms | max: 19.817 ms | sum:315854.813 ms |
massive_intr:(9) |1873997.628 ms | 4945112 | avg: 0.064 ms | max: 18.035 ms | sum:315257.596 ms |
massive_intr:(9) |1876339.600 ms | 5569489 | avg: 0.057 ms | max: 20.473 ms | sum:317771.140 ms |
avg 316294.516 .967
dav1d-worker:(8) | 157496.413 ms | 914670 | avg: 0.333 ms | max: 24.006 ms | sum:304944.125 ms |
dav1d-worker:(8) | 161723.573 ms | 879090 | avg: 0.354 ms | max: 26.070 ms | sum:310806.145 ms |
dav1d-worker:(8) | 157653.936 ms | 929556 | avg: 0.331 ms | max: 21.664 ms | sum:307315.109 ms |
avg 307688.459 .927
X:2502 | 52845.691 ms | 131220 | avg: 0.135 ms | max: 11.479 ms | sum:17656.654 ms |
X:2502 | 47884.248 ms | 134000 | avg: 0.148 ms | max: 12.068 ms | sum:19768.297 ms |
X:2502 | 51731.371 ms | 145167 | avg: 0.132 ms | max: 15.995 ms | sum:19180.803 ms |
avg 18868.584 .678
View attachment "summaries.txt" of type "text/plain" (19365 bytes)
Powered by blists - more mailing lists