[PATCH 0/3] Reduce amount of time kswapd sleeps prematurely v2
From: Mel Gorman
Date: Thu Mar 09 2017 - 02:57:30 EST
Changelog since v1
o Rebase to 4.11-rc1
o Add small clarifying comment based on review
The series is unusual in that the first patch fixes one problem and
introduces of other issues that are noted in the changelog. Patch 2 makes
a minor modification that is worth considering on its own but leaves the
kernel in a state where it behaves badly. It's not until patch 3 that
there is an improvement against baseline.
This was mostly motivated by examining Chris Mason's "simoop" benchmark
which puts the VM under similar pressure to HADOOP. It has been reported
that the benchmark has regressed severely during the last number of
releases. While I cannot reproduce all the same problems Chris experienced
due to hardware limitations, there was a number of problems on a 2-socket
machine with a single disk.
simoop latencies
4.11.0-rc1 4.11.0-rc1
vanilla keepawake-v2
Amean p50-Read 21670074.18 ( 0.00%) 22668332.52 ( -4.61%)
Amean p95-Read 25456267.64 ( 0.00%) 26738688.00 ( -5.04%)
Amean p99-Read 29369064.73 ( 0.00%) 30991404.52 ( -5.52%)
Amean p50-Write 1390.30 ( 0.00%) 924.91 ( 33.47%)
Amean p95-Write 412901.57 ( 0.00%) 1362.62 ( 99.67%)
Amean p99-Write 6668722.09 ( 0.00%) 16854.04 ( 99.75%)
Amean p50-Allocation 78714.31 ( 0.00%) 74729.74 ( 5.06%)
Amean p95-Allocation 175533.51 ( 0.00%) 101609.74 ( 42.11%)
Amean p99-Allocation 247003.02 ( 0.00%) 125765.57 ( 49.08%)
These are latencies. Read/write are threads reading fixed-size random blocks
from a simulated database. The allocation latency is mmaping and faulting
regions of memory. The p50, 95 and p99 reports the worst latencies for 50%
of the samples, 95% and 99% respectively.
For example, the report indicates that while the test was running 99% of
writes completed 99.75% faster. It's worth noting that on a UMA machine that
no difference in performance with simoop was observed so milage will vary.
It's noted that there is a slight impact to read latencies but it's mostly
due to IO scheduler decisions and offset by the large reduction in other
latencies.
mm/memory_hotplug.c | 2 +-
mm/vmscan.c | 136 ++++++++++++++++++++++++++++++----------------------
2 files changed, 79 insertions(+), 59 deletions(-)
--
2.11.0