Re: [Openipmi-developer] Improving IPMI performance under load

[Matt Domsch <Matt_Domsch@dell.com>]

On Tue, Mar 24, 2009 at 08:08:36AM -0500, Corey Minyard wrote:
> Martin Wilck wrote:
> >Hi Corey,
> >
> >yesterday I posted some results about the IPMI performance under CPU 
> >load, which can be up to 25 times slower than in an idle system. I think 
> >it might be worthwhile to try to improve that behavior as well.
> >  
> Yes, that would be expected, as kipmid would never be scheduled in a 
> busy system, and it would just be the timer driving things.
> 
> >I made a variation of my patch which introduces a second parameter 
> >(kipmid_min_busy) that causes kipmid not to call schedule() for a 
> >certain amount of time. Thus if there's IPMI traffic pending, kipmid 
> >will busy-loop for kipmid_min_busy seconds, then starting to schedule() 
> >in each loop as it does now, and finally go to sleep when 
> >kipmid_max_busy is reached. At the same time, I changed the nice value 
> >of kipmid from 19 to 0.
> >  
> I would guess that changing the nice value is the main thing that caused 
> the difference.  The other changes probably didn't make as big a difference.
> 
> >With this patch and e.g. min_busy=100 and max_busy=200, there is no 
> >noticeable difference any more between IPMI performance with and without 
> >CPU load.
> >
> >The patch + results still need cleanup, therefore I am not sending it 
> >right now. Just wanted to hear what you think.
> >  
> I'm ok with tuning like this, but most users are probably not going to 
> want this type of behavior.

I still get complaints from users who see their CPU utilization spike
attributed to kipmi0 when userspace throws a lot of requests down to
the controller.  I've seen them want to limit kipmi0 even further, not
speed it up.

-- 
Matt Domsch
Linux Technology Strategist, Dell Office of the CTO
linux.dell.com & www.dell.com/linux