lkml.org 
[lkml]   [2010]   [Nov]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 1/2] mm: page allocator: Adjust the per-cpu counter threshold when memory is low
From
Date
On Sat, 2010-10-30 at 03:40 +0800, Andrew Morton wrote:
> On Fri, 29 Oct 2010 11:12:11 +0100
> Mel Gorman <mel@csn.ul.ie> wrote:
>
> > On Thu, Oct 28, 2010 at 03:04:33PM -0700, Andrew Morton wrote:
> > > On Thu, 28 Oct 2010 16:13:35 +0100
> >
> > >
> > > I have a feeling this problem will bite us again perhaps due to those
> > > other callsites, but we haven't found the workload yet.
> > >
> > > I don't undestand why restore/reduce_pgdat_percpu_threshold() were
> > > called around that particular sleep in kswapd and nowhere else.
> > >
> > > > vanilla 11.6615%
> > > > disable-threshold 0.2584%
> > >
> > > Wow. That's 12% of all CPUs? How many CPUs and what workload?
> > >
> >
> > 112 threads CPUs 14 sockets. Workload initialisation creates NR_CPU sparse
> > files that are 10*TOTAL_MEMORY/NR_CPU in size. Workload itself is NR_CPU
> > processes just reading their own file.
> >
> > The critical thing is the number of sockets. For single-socket-8-thread
> > for example, vanilla was just 0.66% of time (although the patches did
> > bring it down to 0.11%).
>
> I'm surprised. I thought the inefficiency here was caused by CPUs
> tromping through percpu data, adding things up. But the above info
> would indicate that the problem was caused by lots of cross-socket
> traffic? If so, where did that come from?
>From my understanding, the problem is zone_nr_free_pages() will try to
read each cpu's ->vm_stat_diff, while other CPUs are changing their
vm_stat_diff. This will cause a lot of cache bounce.



\
 
 \ /
  Last update: 2010-11-02 01:55    [W:0.953 / U:0.008 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site