lkml.org 
[lkml]   [2014]   [Jun]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[PATCH 0/5] Improve sequential read throughput v3
Date
Changelog since V2
o Simply fair zone policy cost reduction
o Drop CFQ patch

Changelog since v1
o Rebase to v3.16-rc2
o Move CFQ patch to end of series where it can be rejected easier if necessary
o Introduce page-reclaim related patch related to kswapd/fairzone interactions
o Rework fast zone policy patch

IO performance since 3.0 has been a mixed bag. In many respects we are
better and in some we are worse and one of those places is sequential
read throughput. This is visible in a number of benchmarks but I looked
at tiobench the closest. This is using ext3 on a mid-range desktop and
the series applied.

3.16.0-rc2 3.16.0-rc2
vanilla lessdirty
Min SeqRead-MB/sec-1 120.92 ( 0.00%) 140.73 ( 16.38%)
Min SeqRead-MB/sec-2 100.25 ( 0.00%) 117.43 ( 17.14%)
Min SeqRead-MB/sec-4 96.27 ( 0.00%) 109.01 ( 13.23%)
Min SeqRead-MB/sec-8 83.55 ( 0.00%) 90.86 ( 8.75%)
Min SeqRead-MB/sec-16 66.77 ( 0.00%) 74.12 ( 11.01%)

Overall system CPU usage is reduced

3.16.0-rc2 3.16.0-rc2
vanilla lessdirty-v3
User 390.13 390.20
System 404.41 379.08
Elapsed 5412.45 5123.74

This series does not fully restore throughput performance to 3.0 levels
but it brings it close for lower thread counts. Higher thread counts are
known to be worse than 3.0 due to CFQ changes but there is no appetite
for changing the defaults there.

include/linux/mmzone.h | 210 ++++++++++++++++++++++-------------------
include/linux/writeback.h | 1 +
include/trace/events/pagemap.h | 16 ++--
mm/internal.h | 1 +
mm/mm_init.c | 4 +-
mm/page-writeback.c | 23 +++--
mm/page_alloc.c | 173 ++++++++++++++++++++-------------
mm/swap.c | 4 +-
mm/vmscan.c | 16 ++--
mm/vmstat.c | 4 +-
10 files changed, 258 insertions(+), 194 deletions(-)

--
1.8.4.5



\
 
 \ /
  Last update: 2014-06-27 10:41    [W:0.091 / U:3.760 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site