Measuring latency on my Apple G3

Measuring latency on my Apple G3

[Ismael Valladolid Torres]

Hi,

I am trying latencytest-0.42-png.tar.gz for measuring latency on my 
Apple G3 (Debian/Sid PowerPC). Of course I am enabling USE_GENERIC_TIMER 
and tuning my hard disk using hdparam before tests. But, results show a 
very poor performance. Indeed, any X11 stressing involves a raise of the 
latency and lots of dropouts.

I can't blame hardware. Apple hardware is top-class, and this same 
machine features under 1,5 ms of latency using CoreAudio on Mac OS X. 
So, whether the PowerPC kernel or X11 are not fully optimized, or the 
latencytest application is not reliable on non-Intel platforms.

Any ideas?

Sorry for crossposting... ;)

Regards, Ismael

Re: Measuring latency on my Apple G3

[Michel Dänzer]

On Mon, 2003-07-21 at 11:58, Ismael Valladolid Torres wrote:
> 
> I am trying latencytest-0.42-png.tar.gz for measuring latency on my 
> Apple G3 (Debian/Sid PowerPC). Of course I am enabling USE_GENERIC_TIMER 
> and tuning my hard disk using hdparam before tests. But, results show a 
> very poor performance. Indeed, any X11 stressing involves a raise of the 
> latency and lots of dropouts.
> 
> I can't blame hardware. Apple hardware is top-class, and this same 
> machine features under 1,5 ms of latency using CoreAudio on Mac OS X. 
> So, whether the PowerPC kernel or X11 are not fully optimized, or the 
> latencytest application is not reliable on non-Intel platforms.

First of all, you mentioned in another post that you use x11perf to
create X11 stress. Are there also problems with real world apps?

Also keep in mind that neither the vanilla 2.4 kernel nor the X server
were designed for low latency. Have you tried the low latency and/or
O(1) scheduler kernel patches, and not running the X server with
negative nice values if you are?

Last but possibly not least of all, the dmasound driver has been less
prone to dropouts in my experience than the ALSA driver.


-- 
Earthling Michel Dänzer   \  Debian (powerpc), XFree86 and DRI developer
Software libre enthusiast  \     http://svcs.affero.net/rm.php?r=daenzer

Re: Measuring latency on my Apple G3

[Ismael Valladolid Torres]

Michel Dänzer escribio el 21/07/03 13:09:
> First of all, you mentioned in another post that you use x11perf to
> create X11 stress. Are there also problems with real world apps?

"Real world" apps work properly (except for the Gnome theme manager 
which displays garbage). I only find video performance using Linux a lot 
lower than using any of the Mac OS's.

> Also keep in mind that neither the vanilla 2.4 kernel nor the X server
> were designed for low latency. Have you tried the low latency and/or
> O(1) scheduler kernel patches, and not running the X server with
> negative nice values if you are?

Both of those patches (A. Morton and R. Love's ones) were applied to my 
kernel. I don't know about running the X server with different nice 
values, which advantage would I get? (I am using gdm as my X session 
manager).

> Last but possibly not least of all, the dmasound driver has been less
> prone to dropouts in my experience than the ALSA driver.

I should try with pure OSS. Let's not forget that the latency test suite 
is using the OSS compatibility layer of ALSA.

Thanks a lot. Any feedback is still very useful.

Regards, Ismael

Re: Measuring latency on my Apple G3

[Michel Dänzer]

On Mon, 2003-07-21 at 13:46, Ismael Valladolid Torres wrote: 
> Michel Dänzer escribio el 21/07/03 13:09:
> > First of all, you mentioned in another post that you use x11perf to
> > create X11 stress. Are there also problems with real world apps?
> 
> "Real world" apps work properly (except for the Gnome theme manager 
> which displays garbage). 

Probably a GTK bug, not related to this thread. The point is real world
as opposed to synthetic throughput benchmarks like x11perf.

> I only find video performance using Linux a lot lower than using 
> any of the Mac OS's.

Which isn't very surprising, as they can use parts of the graphics chip
that we don't have specs for, for one.


> > Also keep in mind that neither the vanilla 2.4 kernel nor the X server
> > were designed for low latency. Have you tried the low latency and/or
> > O(1) scheduler kernel patches, and not running the X server with
> > negative nice values if you are?
> 
> Both of those patches (A. Morton and R. Love's ones) were applied to my 
> kernel. 

Beware that at least the low latency patch needs fiddling with
arch/ppc/config.in to actually be enabled (check with grep LOLAT
.config), and that the preempt patch (which I assume you mean by R.
Love's) actually made things worse for me when I tried it on PPC a while
ago. This may have been fixed in the meantime though.

> I don't know about running the X server with different nice values, 

sudo dpkg-reconfigure xserver-common

> which advantage would I get?

The X server might take less CPU away from other processes. Or maybe the
problem is the other way around. :) YMMV.


-- 
Earthling Michel Dänzer   \  Debian (powerpc), XFree86 and DRI developer
Software libre enthusiast  \     http://svcs.affero.net/rm.php?r=daenzer

Re: Measuring latency on my Apple G3

[Ismael Valladolid Torres]

Michel Dänzer escribio el 22/07/03 02:22:
> Probably a GTK bug, not related to this thread.

I know, I know... ;)

> Which isn't very surprising, as they can use parts of the graphics chip
> that we don't have specs for, for one.

Sure. Only I remember having installed Mandrake 9.1, and I think video 
performance was somewhat better. Maybe I am wrong and memory cheats...

> Beware that at least the low latency patch needs fiddling with
> arch/ppc/config.in to actually be enabled (check with grep LOLAT
> .config), and that the preempt patch (which I assume you mean by R.
> Love's) actually made things worse for me when I tried it on PPC a while
> ago. This may have been fixed in the meantime though.

You're right! First line in the patch file says:

arch/i386/config.in         |    3 +

So the patch to the config.in file is only applied automatically for the 
i386 architecture. Indeed CONFIG_LOLAT_SYSCTL was not in my .config, so 
should I paste this manually into arch/ppc/config.in? Would this be enough?

And about the preempt, patch, well, it seems I still have enough tests 
to perform so I can spend my holydays safely.

 > YMMV

IJLUW (indeed, just like using Windows!) ;)

Regards, Ismael