Re: [ck] Re: Best nice level for X with SD

Re: [ck] Re: Best nice level for X with SD

[Con Kolivas]

On Friday 27 April 2007 06:11, Martin Steigerwald wrote:
> Am Dienstag 24 April 2007 schrieb Martin Steigerwald:
> > Hello once again,
> >
> > I now tested cfs-v5 on my ThinkPad T23 with
> >
> > martin@deepdance:~> cat /proc/version
> > Linux version 2.6.20.7-tp23-sws2-2.2.9.9-sd-0.46 (martin@deepdance)
> > (gcc-Version 4.1.2 20061115 (prerelease) (Debian 4.1.1-21)) #1 PREEMPT
> > Sun Apr 22 21:40:30 CEST 2007
> >
> > I had the KDE desktop + Amarok running and nothing else. But music
> > playbacks stopped for about a second every now and then. Quite
> > annoyingly often. I think thats quite similar to the mouse pointer
> > freezes I had on the T42. I did not look whether the mousepointer
> > really froze, but I can check that again.
> >
> > I then did the same naive user test than I did with sd-0.44 and
> > sd-0.46, starting building debian kernel-source package, opening lots
> > of applications - mostly KDE ones, anything I could find on my toolbar
> > from Konsole, KWrite  to Konqueror, Iceweasel ... - and then moved
> > around Amaroks window. Well I managed to completely stop music playback
> > for more than 5 seconds. Music playback completely paused. This again
> > was with nice -10 configured in CFSv5 source.
> >
> > I repeated the test with X.org reniced to 0. It got better but I still
> > managed to pause music playback for a second or so. But maybe that was
> > one of that music stops that also happen when the machine is idling
> > around. They also happen with X.org reniced to 0.
>
> Hi!
>
> Okay, this got way better with cfs-v6. I tested my way through this with
> different cfs versions and settings as guided by Ingo. We did this via
> private mail as Ingo didn't want to (ab)use the ck-patch mailinglist for
> this. But now after having ranted publically about cfs, I at least also
> write publically that things have gotten better ;-).
>
> After having a drastic sound playback regression cfs-v6-rc7 which I could
> not reproduce today, I have way better results with cfs-v6 than with
> cfs-v5. This is with 2.6.20.7, sws2 2.2.9.9 and cfs-v6.
>
> But still there were some pauses in music playback with default setting.
>
> For smooth music playback with Amarok on my T23 I had to go as low as:
>
> ---------------------------------------------------------------------
> deepdance:/proc/sys/kernel> grep ".*" sched*gran*
> sched_granularity_ns:2000000
> sched_wakeup_granularity_ns:0
> deepdance:/proc/sys/kernel> echo "1500000" >sched_granularity_ns
> deepdance:/proc/sys/kernel> grep ".*" sched*gran*
> sched_granularity_ns:1500000
> sched_wakeup_granularity_ns:0
> deepdance:/proc/sys/kernel> echo "1000000" >sched_granularity_ns
> deepdance:/proc/sys/kernel> grep ".*" sched*gran*
> sched_granularity_ns:1000000
> sched_wakeup_granularity_ns:0
> deepdance:/proc/sys/kernel> echo "500000" >sched_granularity_ns
> deepdance:/proc/sys/kernel> grep ".*" sched*gran*
> sched_granularity_ns:500000
> sched_wakeup_granularity_ns:0
> deepdance:/proc/sys/kernel> echo "250000" >sched_granularity_ns
> deepdance:/proc/sys/kernel> grep ".*" sched*gran*
> sched_granularity_ns:250000
> sched_wakeup_granularity_ns:0
> ---------------------------------------------------------------------
>
> Then music playback was working nicely when I did not interfere ;-)
>
> Even then I was able to stop audio playback for longer than a second by my
> naive user test... compiling a kernel, opening lots of apps and moving
> the Amarok window around like mad.
>
> With
>
> ---------------------------------------------------------------------
> deepdance:/proc/sys/kernel> echo "0" >sched_granularity_ns
> deepdance:/proc/sys/kernel> grep ".*" sched*gran*
> sched_granularity_ns:0
> sched_wakeup_granularity_ns:0
> ---------------------------------------------------------------------
>
> it was working nice. Subjectively on par with sd-0.46.
>
> buddabrod can you test these suggestions? I am using cfs-v6 with X11
> renicing.

Clearly there are some serious regressions for audio playback with CFS.
This is incredible effort to go to with CFS. 

> Still sd-0.46 is giving me as default what I have to configure with
> cfs-v6 ;-). And as a user I want this behavior as default, instead of
> having to fiddle with the schedular settings. Smooth music playback is a
> must as Ingo agreed already ;).

Nice to hear that SD does everything CFS strives to achieve. I'm glad I 
continued development on it so that it remains the reference for CFS to 
compare to.

> No need to continue that thread here from my side... I just wanted to
> mention that it got better. Its still Cons ck-patch list. A general
> scheduler hacking and testing list would come in handy I guess.

lkml is perfectly suited for that discussion provided everyone follows the 
convention of cc'ing everyone on their replies, and I have taken the liberty 
of cc'ing it on this thread too.

-- 
-ck

Re: [ck] Re: Best nice level for X with SD

[Martin Steigerwald]

Am Freitag 27 April 2007 schrieb Con Kolivas:
> Clearly there are some serious regressions for audio playback with CFS.
> This is incredible effort to go to with CFS.

Hi Con!

Well, at least on my two ThinkPads T42 and T23.

I perceive sd-0.46 as more mature as CFSv6. And if asked to include one of 
them in the kernel *right now* or *soonish*, I would probably choose  
sd-0.46 and give Ingo a bit more time. 

Or I would suggest a plugging interface. For me it seems difficult to make 
large scale testing efforts when I have to build two kernels all the 
time. So at least a boot time configuration parameter, even better a 
parameter to switch schedulers on the fly would be good. This way it 
would be way easier to compare different schedulers. 

Just like with I/O schedulers.  I switched from "anticipatory" to Ingo's 
great "cfq" before it became the new default. I think that a few weeks of 
testing will never be enough. How long did it take till cfq became the 
new default? How much testing did people, including distributors such as 
SUSE / Novell - who switched to cfq eariler - do? Would that torough 
testing have been done when it was a patch outside of mainline all the 
time and there was no easy possibility to switch I/O schedulers?

I think that on process schedulers this would not be any different and any 
quick prefer this or this one will miss needed long time test results. 
But I have no clue whether plugsched, especially with online switching, 
wouldn't be too much overhead.

Well just my subjective opinion. I didn't do any professional benchmarks. 
But actually what I care about most is my usually workloads and whether I 
can click together corner cases to bring schedulers out of the bounds I 
like them to operate it. Cause thats what relevant for me when using a 
Linux machine as a desktop ;-). I didn't do any testing on a server 
tough, my virtual server just uses bog standard Debian etch kernel and I 
see no need to change that.

> > Still sd-0.46 is giving me as default what I have to configure with
> > cfs-v6 ;-). And as a user I want this behavior as default, instead of
> > having to fiddle with the schedular settings. Smooth music playback
> > is a must as Ingo agreed already ;).
>
> Nice to hear that SD does everything CFS strives to achieve. I'm glad I
> continued development on it so that it remains the reference for CFS to
> compare to.

I am happy, too. And Ingo is working really hard on it. I already have the 
next patch to test here. And it becomes better.

> lkml is perfectly suited for that discussion provided everyone follows
> the convention of cc'ing everyone on their replies, and I have taken
> the liberty of cc'ing it on this thread too.

Okay, I am CCing this to kernel list, too!

But my original concern was, whether you want to have these CFSv6 related 
tests on ck-patch-ml at all? Strictly spoken its about ck-patches and 
your stuff, and not other schedulers, but well, no one adhered that much 
to its topic in the last weeks.

Ok, but now to my daily tasks. These still (should) have priority ;-).

Regards,
Martin


Am Donnerstag 26 April 2007 schrieb Martin Steigerwald:

> it was working nice. Subjectively on par with sd-0.46.

Hi!

Well, there are still some audio glitches here and then. Even with

deepdance:/proc/sys/kernel> grep ".*" sched*gran*
sched_granularity_ns:0
sched_wakeup_granularity_ns:0

cfs-v6 has quite high context switch rates when driven with this 
configuration. This is context switch rates while Kaffeine plays music 
and a kernel compiles:

procs -----------memory---------- ---swap-- -----io---- -system-- ----cpu----
 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id 
wa
 3  0    136  22864    276 484680    0    0   149   102  379 1102 77  7 13  
3
 2  0    136  22864    276 484680    0    0     0    11  357 1849 99  1  0  
0
 1  0    136  22744    276 484808    0    0   128     0  413 2653 97  3  0  
0
 2  0    136  22744    276 484812    0    0     0     0  374 2382 96  4  0  
0
 1  0    136  22564    276 484940    0    0   128     0  411 2819 94  6  0  
0
 1  0    136  22564    276 484940    0    0     0     0  370 1938 98  2  0  
0
 1  0    136  22444    276 485068    0    0   128     0  374 2235 98  2  0  
0
 2  0    136  22444    276 485072    0    0     0     0  393 2329 97  3  0  
0
 1  0    136  22324    276 485200    0    0   128     0  351 1561 100  0  
0  0
 1  0    136  22324    276 485200    0    0     0     0  376 1790 99  1  0  
0

Maybe on desktop a schedular needs such high rates to provide excellent 
desktop responsiveness?

Ingo, I send you strace and vmstat stuff in a seperate mail. Ok, enough of 
it, we can continue this in private mail and I test your new cfs-v7-rc ,)


Back to my original question:

What nice value would do you recommend for X in sd-0.46? Really -10? I 
didn't renice at all... well what have others tried? Hmmm, Mike Mattie 
used -1. Maybe  I should try -5 or so? Well maybe I just try some values. 
I just would like to hear whether there an official recommendation? Like 
the -10 that cfs-v6 sets as default if X renicing is enabled.

And would it make sense to renice kernel threads as well with sd-0.46? I 
have seen these reniced to -10, but maybe thats default?

Regards,
-- 
Martin 'Helios' Steigerwald - http://www.Lichtvoll.de
GPG: 03B0 0D6C 0040 0710 4AFA  B82F 991B EAAC A599 84C7

Re: [ck] Re: Best nice level for X with SD

[Martin Steigerwald]

Am Freitag 27 April 2007 schrieb Martin Steigerwald:
> Am Freitag 27 April 2007 schrieb Con Kolivas:
> > Clearly there are some serious regressions for audio playback with
> > CFS. This is incredible effort to go to with CFS.
>
> Hi Con!
>
> Well, at least on my two ThinkPads T42 and T23.
>
> I perceive sd-0.46 as more mature as CFSv6. And if asked to include one
> of them in the kernel *right now* or *soonish*, I would probably choose
> sd-0.46 and give Ingo a bit more time.

Hi!

I tried cfs-v7 where Ingo fixed a 'big bug' he found while looking at the 
debug data I provided to him.

Well, its rocking now on both my Amarok machine, the T23, and the T42. 
Absolutely smooth music playback, no matter what. No freezes anymore 
whatsoever.

Even when stress testing with my usual naive user tests, having make-kpkg 
linux_image running, starting tons of apps and solid moving Amarok window 
like mad - that was on my T23.  On my T42 I repeated that Kaffeine test, 
playing a video with it, while moving the window like mad, no complete 
freezes anymore...

Congrats, Ingo! cfs-v7 is rocking now ;-).


Tested on 2.6.21.1 with sws2 2.2.9.13, dynticks and CONFIG_HZ=300 as the 
older kernels I tested with.

That is without X renicing just like I run sd-0.46 as well.

Now it seems that I have *two* schedulers to choose from. ;-) I am 
satisfied with both sd-0.46 and cfs-v7 now.

Actually cfs-v7 feels a bit smoother for me now. Thats nothing concrete at 
the moment, nothing I can get a hold of it.  cfs-v7 just feels a bit 
smoother to me.

Will do some longer term testing with cfs-v7 now.

Regards,
-- 
Martin 'Helios' Steigerwald - http://www.Lichtvoll.de
GPG: 03B0 0D6C 0040 0710 4AFA  B82F 991B EAAC A599 84C7