lkml.org 
[lkml]   [2018]   [Jan]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] mm: numa: Do not trap faults on shared data section pages.
From
Date


> On Jan 16, 2018, at 1:26 PM, Mel Gorman <mgorman@suse.de> wrote:
>
> On Tue, Jan 16, 2018 at 11:28:44AM -0800, Henry Willard wrote:
>> Workloads consisting of a large number processes running the same program
>> with a large shared data section may suffer from excessive numa balancing
>> page migration of the pages in the shared data section. This shows up as
>> high I/O wait time and degraded performance on machines with higher socket
>> or node counts.
>>
>> This patch skips shared copy-on-write pages in change_pte_range() for the
>> numa balancing case.
>>
>> Signed-off-by: Henry Willard <henry.willard@oracle.com>
>> Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
>> Reviewed-by: Steve Sistare steven.sistare@oracle.com
>
> Merge the leader and this mail together. It would have been nice to see
> data on other realistic workloads as well.
>
> My main source of discomfort is the fact that this is permanent as two
> processes perfectly isolated but with a suitably shared COW mapping
> will never migrate the data. A potential improvement to get the reported
> bandwidth up in the test program would be to skip the rest of the VMA if
> page_mapcount != 1 in a COW mapping as it would be reasonable to assume
> the remaining pages in the VMA are also affected and the scan is wasteful.
> There are counter-examples to this but I suspect that the full VMA being
> shared is the common case. Whether you do that or not;
>
> Acked-by: Mel Gorman <mgorman@suse.de>

Thanks. The real customer cases where this was observed involved large, 1TB or more, eight socket machines running very active RDBMS workloads. These customers saw high iowait times and a loss in performance when numa balancing was enabled. Previously there was no reported iowait time. The extent of the loss of performance was variable depending on the activity and never quantified. The little test program is a distillation of what was observed. In the real workload, a large part of the VMA is shared, but not all of it, so this seemed the simplest and most reliable patch.

Henry

>
> --
> Mel Gorman
> SUSE Labs

\
 
 \ /
  Last update: 2018-01-18 00:19    [W:0.074 / U:0.764 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site