linux-kernel - Re: niced tasks on SMT system

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

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Thu, 28 Dec 2017 19:53:54 +0100
From:   Pavel Machek <pavel@....cz>
To:     Mike Galbraith <efault@....de>
Cc:     Thomas Gleixner <tglx@...utronix.de>,
        kernel list <linux-kernel@...r.kernel.org>, mingo@...hat.com,
        hpa@...or.com, x86@...nel.org, peterz@...radead.org
Subject: Re: niced tasks on SMT system

Hi!

> > Ok, so I'm compiling, and I'd like to run a flight simulator.
> > 
> > Flightgear normally does 20fps on my system... kinda low but playable.
> > 
> > I have reniced make -j 5 fo kernel running. Scheduler gives 100% of
> > one of CPUs to Flightgear (good), but the smt sibling is used by the
> > compilation, and I'm down to 9 fps. Not good.
> > 
> > Even with single-threaded make, I have 10-13fps.
> > 
> > Is there way to learn which CPUs are SMT siblings?
> 
> cat /sys/devices/system/cpu/cpu{N}/topology/thread_siblings

Thanks for a hint.

Well, something is definitely wrong there:

pavel@duo:/data/l/linux-n900$ cat
/sys/devices/system/cpu/cpu0/topology/thread_siblings
03

I believe that means that cpu0 & cpu1 are sharing physical cpu. Yet,
when I run flightgear and "nice while1". flightgear goes to CPU#0 and
while1 to CPU#1. Would not it be nice if while1 went preferably to
CPUs #2 and #3? Ok. after a while while1 moved to cpu#2, good.

> > Is there way to disable SMT during runtime?
> 
> You could offline them, but wouldn't it be better to tell each which
> CPUs they can use, or perhaps partition your box with cpusets?

Let me try offlining first... and yes, SMT seems to be problem here.

| CPU0              | CPU1   | CPU2   | CPU3 |        |
| flightgear        | off    | off    | off  | 21 fps |
| flightgear        | while1 | off    | off  | 17 fps |
| flightgear+while1 | off    | off    | off  | 10 fps |
| flightgear        | off    | while1 | off  | 21 fps |

> > Can I do something to improve Flightgear performance and still do
> > compilation?
> 
> Sure, run everything associated with your game as RT, and everything
> not game gets the leftover cycles.  If there are none, box will
>  throttle RT to save itself from it's psycho :) driver and you'll know
> that you need a bigger box. (it's likely your phone)

> Oh yeah, echo NO_RT_RUNTIME_SHARE > /sys/kernel/debug/sched_features
> before you try that, otherwise the throttle won't help.

Sure, but that is not a problem I have.

Scheduler already (correctly) decides that the game is important, and
gives game all the time it wants (100% of one CPU). That is not a problem.

Problem seems to be that it schedules other tasks to SMT sibling, and
that slows down the game quite significantly.

Best regards,

									Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Download attachment "signature.asc" of type "application/pgp-signature" (182 bytes)