Re: [slub shrink] 0f6934bf16: +191.9% vmstat.system.cs

[Fengguang Wu]

 On Thu, Jan 16, 2014 at 11:12:19AM -0800, Dave Hansen wrote:
> On 01/15/2014 07:07 PM, kernel test robot wrote:
> > 9a0bb2966efbf30  0f6934bf1695682e7ced973f6  
> > ---------------  -------------------------  
> >    8235933 ~ 2%     +80.6%   14872911 ~ 3%  lkp-sbx04/micro/will-it-scale/read2
> >    8235933          +80.6%   14872911       TOTAL interrupts.RES
> > 
> > 9a0bb2966efbf30  0f6934bf1695682e7ced973f6  
> > ---------------  -------------------------  
> >     161531 ~ 7%    +191.9%     471544 ~ 9%  lkp-sbx04/micro/will-it-scale/read2
> >     161531         +191.9%     471544       TOTAL vmstat.system.cs
> > 
> > 9a0bb2966efbf30  0f6934bf1695682e7ced973f6  
> > ---------------  -------------------------  
> >      32943 ~ 1%     +71.8%      56599 ~ 3%  lkp-sbx04/micro/will-it-scale/read2
> >      32943          +71.8%      56599       TOTAL vmstat.system.in
> 
> I suspect that something is wrong with that system.  My 160-cpu system
> does about 40,000 interrupts/sec and ~4300 context switches/sec when
> running 160 read2_processes.  I wonder if you're hitting swap or the
> dirty limits or something.  Are you running it with way more threads
> than it has CPUs?

lkp-sbx04 has 64 CPU threads, and I'm running will-it-scale with
thread numbers 1 16 24 32 48 64 8 

> Also, are those will-it-scale tests the threaded or process versions?

Hansen, I'm running will-it-scale with parameters

./runtest.py read2 16 1 16 24 32 48 64 8

Which runs both threaded/process tests. The runtest.py is modified to
accept a custom list of threads to run. The patch is attached.

The list of duration and thread numbers for runtest.py are computed
and is different for machines with different number of CPUs. The goal
of the computation is so that the test wall time on different machines
will be equally ~5 minutes.

In a system with 120 CPUs, the numbers will be:

./runtest.py brk1 16 1 120 15 30 45 60 90

Thanks,
Fengguang