Re: [mm, page_alloc] e6cbd7f2ef: pixz.throughput -5.1% regression

[Mel Gorman <mgorman@techsingularity.net>]

[-- Attachment #1: Type: text/plain, Size: 1426 bytes --]

On Sun, Aug 28, 2016 at 11:51:20AM +0200, Thorsten Leemhuis wrote:
> Lo! On 08.08.2016 10:29, kernel test robot wrote:
> > 
> > FYI, we noticed a -5.1% regression of pixz.throughput due to commit:
> > 
> > commit e6cbd7f2efb433d717af72aa8510a9db6f7a7e05 ("mm, page_alloc: remove fair zone allocation policy")
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> > 
> > in testcase: pixz
> > on test machine: 48 threads Ivytown Ivy Bridge-EP with 64G memory
> > with following parameters:
> > 
> > 	nr_threads: 100%
> > 	cpufreq_governor: performance
> 
> Mel, this report made it to the regression list for 4.8, but it seems
> nothing happened after the initial report. Was it discussed (and maybe
> even fixed?) elsewhere? Or was it deemed not important enough? Should I
> drop it for the regression list?
> 

Drop it for the moment.

My expectation is that it's a relatively minor hazard. The removal of the
fair zone allocation policy is a shorter path which benefits a number
of workloads but also potentially changes the color of pages used in
microbenchmarks which can have a cache effect.  It's on my TODO list to
reproduce this exactly as LKP does but my own preliminary experiments using
pbzip (yes, it's different) on 4 machines showed gains on all 4 machines
so something relatively subtle is going on or it's machine specific.

-- 
Mel Gorman
SUSE Labs