RE: [patch 0/2] fix perf. bug in wake-up load balancing for aim7 and db workload

["Chen, Kenneth W" <kenneth.w.chen@intel.com>]

Nick Piggin wrote on Monday, February 13, 2006 7:44 PM
> Andrew Morton wrote:
> > Well I don't see any benchmark numbers in the original patch.  Just an
> > assertion that it "should" help something.
> > 
> The regression was in a Ken's commercial database benchmark. I couldn't
> reproduce it but presumably it did fix it otherwise Ken would would have
> piped up?

I wasn't entirely happy though ;-)


> > I'm more inclined to revert it and not add the sysctl (ugh) until we have a
> > good reason to do so.
> > 
> 
> If you revert this then Ken's database benchmark gets worse. Hence the
> sysctl.

Yes, Nick is correct.  For db workload, the wake-ups type are mixed.
Some of them are random, some of them are not because we do interrupt
binding.  The break down between random/fixed is about 30/70.  Thus,
Nick's patch helps 30% of time.  With sysctl, we can regain the entire
up side for db workload while retain workload on the other end of
spectrum.

- Ken