At Tue, 04 Aug 2009 13:34:06 -0400,
Jeff Mahoney wrote:
>
> Takashi Iwai wrote:
> > At Tue, 4 Aug 2009 17:55:01 +0200,
> > Jean Delvare wrote:
> >> Le mardi 04 août 2009, Takashi Iwai a écrit :
> >>> Hi,
> >>>
> >>> just looking through SUSE kernel config files, I found that
> >>> i386/default setup is still for the server, CGROUP=y, PREEMPT=n, etc.
> >>> Can it be changed to the ones like kernel-desktop?
> >>>
> >>> According to bug #526780, kernel-desktop actually helps to improve
> >>> the audio issue (mostly PA). Since kernel-default is basically
> >>> for old machines (thus it's the only kernel with ISA support), we
> >>> don't have to keep the server optimizations there.
> >> If you install Factory/11.2 on an old system, it may be to make a
> >> server out of it and not a desktop system. Optimizing kernel-default
> >> for desktops doesn't make a lot of sense to me.
> >
> > Well, then the question is rather what kind of purpose is supposed
> > for i386 kernel-default. If it's a kernel for machines without PAE
> > support, the chance is high that it's a relatively old desktop
> > machine.
>
> kernel-default is just supposed to be kernel-pae without PAE and with
> ISA support. I'd prefer not to introduce any new differences.
Hm, OK, it's a strong opinion.
(BTW, in reality, the difference between default and pae isn't only
about PAE and ISA. The pae kernel has many options as built-in,
e.g. ATA and IPV6 while they are all modules in the default kernel.)
But, let me try to whip a dead horse again. Maybe I should have
started from the question what should be the difference between
desktop and pae (default).
As now, desktop kernel seems to have
- disablement of CGROUP and GROUP_SCHED
- HZ=250 vs HZ=1000
- PREEMPT_NONE vs PREEMPT
My guess is that the first one is crucial for responsiveness.
Unfortunately it can't be off for openSUSE-11.1 because of SLE11
requirement. This was disabled on openSUSE-11.0 due to the disastrous
results. It seems that the situation got improved, but if this is
really the key factor to worsen the responsiveness, we should
reconsider for 11.2.
About the second one, the impact by different HZ should be fairly
small as we usually NOHZ is used. So, HZ=1000 can be taken for the
server as well.
The last one, the preemption option is another question. In my very
old benchmarks, PREEMPT had some performance impact in a few %. But,
the preemption can be a win in some specific benchmarks, OTOH. So,
there are some myths in this area...
thanks,
Takashi
--
To unsubscribe, e-mail: opensuse-kernel+unsubscribe(a)opensuse.org
For additional commands, e-mail: opensuse-kernel+help(a)opensuse.org