Re: CONFIG_PREEMPT and server workloads

[Andrea Arcangeli <andrea@suse.de>]

On Thu, Mar 18, 2004 at 10:20:23AM -0500, Tom Sightler wrote:
> Well, I reported an issue on my laptop several weeks ago where network
> activity via my aironet wireless adapter would use 60-70% of the CPU but
> only when PREEMPT was enabled.  Looking back over the list I see other

sounds like a preempt bug triggering in a softirq, or maybe an SMP bug
triggering in UP.

I certainly agree with Andrew that preempt cannot cause a 60/70%
slowdown with a network card at least (if you do nothing but spinlocking
in a loop then it's possible a 60/70% slowdown instead but the system
time that a wifi card should take is nothing compared to the idle/user
time).

> I'll try to reproduce my issue with current generation kernels (last I
> tested with PREEMPT was 2.6.1) and see if my problem is still there. 
> When I reported the issue last time no one seemed interested so I just
> learned to disable preempt.

This is fine IMHO (even if they are preempt bugs). Except if your
machine is UP, if it's UP it would be better track down the bug, since
it may still happen on a SMP box (I wouldn't be too surprised if wifi
drivers aren't smp safe). I certainly agree with Andrew when he says
preempt is a good debugging knob to debug smp bugs in UP (though I would
leave it off in production UP, like in production SMP).