[lkml]   [2014]   [Jul]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 0/5] Improve sequential read throughput v4r8
On Tue, Jul 01, 2014 at 07:39:15PM +0100, Mel Gorman wrote:
> The fair zone policy itself is partially working against the lowmem
> reserve idea. The point of the lowmem reserve was to preserve the lower
> zones when an upper zone can be used and the fair zone policy breaks
> that. The fair zone policy ignores that and it was never reconciled. The
> dirty page distribution does a different interleaving again and was never
> reconciled with the fair zone policy or lowmem reserves. kswapd itself was
> not using the classzone_idx it actually woken for although in this case
> it may not matter. The end result is that the model is fairly inconsistent
> which makes comparison against it a difficult exercise at best. About all
> that was left was that from a performance perspective that the fair zone
> allocation policy is not doing the right thing for streaming workloads.

The inevitable feedback will be to reconcile those differences so I'm
redid the series and queued it for testing. Patch list currently looks

mm: pagemap: Avoid unnecessary overhead when tracepoints are deactivated
mm: Rearrange zone fields into read-only, page alloc, statistics and page reclaim lines
mm: page_alloc: Add ALLOC_DIRTY for dirty page distribution
mm: page_alloc: Only apply the fair zone allocation policy if it's eligible
mm: page_alloc: Only apply either the fair zone or dirty page distribution policy, not both
mm: page_alloc: Reduce cost of the fair zone allocation policy
mm: page_alloc: Reconcile lowmem reserves with fair zone allocation policy
mm: vmscan: Fix oddities with classzone and zone balancing
mm: vmscan: Reconcile balance gap lowmem reclaim with fair zone allocation policy
mm: vmscan: Remove classzone considerations from kswapd decisions

About 13 hours to test for ext3 on the small machine, 3 days for the larger
machine. The test could be accelerated by either reducing the iterations or
the memory size of the machine but that would distort the results too badly.

Mel Gorman

 \ /
  Last update: 2014-07-02 00:01    [W:0.047 / U:2.372 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site