Re: specjbb2005 and aim7 regression with 2.6.32-rc kernels

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Fri, 2009-11-06 at 09:04 +0100, Ingo Molnar wrote:
> * Zhang, Yanmin <yanmin_zhang@linux.intel.com> wrote:
> 
> > Comparing with 2.6.31, specjbb2005 and aim7 have some regressions with 
> > 2.6.32-rc kernels on core2 machines.
> > 
> > 1) On 4*4 core tigerton: specjbb2005 has about 5% regression.
> > 2) On 2*4 stoakley: aim7 has about 5% regression.
> > 
> > On Nehalem, specjbb2005 has about 2%~8% improvement instead of 
> > regression.
> > 
> > aim7 has much dependency on schedule patameters, such like 
> > sched_latency_ns, sched_min_granularity_ns, and 
> > sched_wakeup_granularity_ns. 2.6.32-rc kernel decreases these 
> > parameter values. I restore them and retest aim7 on stoakley. aim7 
> > regression becomes about 2% and specjbb2005 regression also becomes 
> > 2%. But on Nehalem, the improvement shrinks.
> 
> Which precise 2.6.32-rc commit have you tested?
> 
> Since v2.6.32-rc6 Linus's tree has this one too:
> 
>   f685cea: sched: Strengthen buddies and mitigate buddy induced latencies
> 
> Which should improve things a bit. For 2.6.33 we have queued up these 
> two in -tip:
> 
>   a1f84a3: sched: Check for an idle shared cache in select_task_rq_fair()
>   1b9508f: sched: Rate-limit newidle
> 
> If any of them fixes a performance regression we could still merge them 
> into 2.6.32 as well.
1b9508f definitely fixes netperf UDP loopback regression.

Yanmin