[MMTests] memcachetest and parallel IO on xfs

From: Mel Gorman
Date: Mon Jul 23 2012 - 17:19:03 EST

Configuration: global-dhp__parallelio-memcachetest-xfs
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__parallelio-memcachetest-xfs
Benchmarks: parallelio


Indications are that there was a large regression in page reclaim decisions
between 2.6.39 and 3.0 as swapping increased a lot.

Benchmark notes

This is an experimental benchmark designed to measure the impact of
background IO on a target workload.

mkfs was run on system startup.
mkfs parameters -f -d agcount=8
mount options inode64,delaylog,logbsize=262144,nobarrier for the most part.
On kernels to old to support delaylog was removed. On kernels
where it was the default, it was specified and the warning ignored.

The target workload in this case is memcached and memcachetest. This is a
benchmark of memcached and the workload is mostly anonymous. The benchmark
was chosen as it was a random client that is considered a valid benchmark
for memcache and does not consume much memory in the client. The server
was configured to use 80% of memory.

In the background, dd is used to generate IO of varying sizes. As the sizes
increase, memory pressure may push the target workload out of memory. The
benchmark is meant to measure how much the target workload is affected
and may be used as a proxy measure for page reclaim decisions.

Unlike other benchmarks, only the run with the worst throughput is displayed.
This benchmark varies quite a bit depending on the reference pattern from
the client. This hides the interesting result in the noise so we only
consider the worst case.

Machine: arnold
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__parallelio-memcachetest-xfs/arnold/comparison.html
Arch: x86
CPUs: 1 socket, 2 threads
Model: Pentium 4
Disk: Single Rotary Disk


Even for small amounts of background IO the memcached process is being
pushed into swap for 3.3 and 3.4 although earlier kernels fared better.
There are indications that there was a serious regression between 2.6.39
and 3.0 as throughput dropped for larger amounts of IO and swapping was

The "page reclaim immediate" figures started increasing from 3.2 implying
that a lot of dirty LRU pages are reaching the end of the LRU lists.

Machine: hydra
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__parallelio-memcachetest-xfs/hydra/comparison.html
Arch: x86-64
CPUs: 1 socket, 4 threads
Model: AMD Phenom II X4 940
Disk: Single Rotary Disk


Performance again dropped sharply betwen 2.6.39 and 3.0 with huge jumps
in the amount of swap IO.

As with arnold, dirty pages are reaching the end of the LRU list.

Machine: sandy
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__parallelio-memcachetest-xfs/sandy/comparison.html
Arch: x86-64
CPUs: 1 socket, 8 threads
Model: Intel Core i7-2600
Disk: Single Rotary Disk

No results available.

Mel Gorman
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/