[MMTests] Scheduler
From: Mel Gorman
Date: Mon Jul 23 2012 - 17:12:06 EST
Configuration: global-dhp__scheduler-performance
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__scheduler-performance
Benchmarks: hackbench-pipes hackbench-sockets pipetest starve lmbench
Summary
=======
This is a mixed bag. The results on an I7 generally look great! There are
some major improvemnets in there and I think this may be due to scheduler
developers working with the latest chips. The other machines did not far
as well. Look at pipetest on hydra for an example of a particularly bad
set of results.
Benchmark notes
===============
starve (http://www.hpl.hp.com/research/linux/kernel/o1-starve.php) was
chosen because even though it is designed to isolate a bug in the O(1)
scheduler, it is still interesting to monitor for performance regressions.
It does not take any special parameters.
hackbench was chosen because it's a general scheduler benchmark that is
sensitive to regressions in the scheduler fast-path. It is difficult
to draw conclusions from as it is somewhat sensitive to the starting
conditions of the machine but trends over time may be observed. It is
run in both pipe and sockets mode and for each number of clients, it is
run for 30 iterations.
pipetest is a scheduler ping-pong test that measures context switch latency.
It runs for 30 iterations.
lmbench is just running the lat_ctx test and is another measure of context
switch latency.
===========================================================
Machine: arnold
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__scheduler-performance/arnold/comparison.html
Arch: x86
CPUs: 1 socket, 2 threads
Model: Pentium 4
Disk: Single Rotary Disk
Status: Context switch latency is regressing.
===========================================================
starve is looking ok except for 3.0 and 3.1 where System CPU time and elapsed
time increased. This was fixed in later kernels but worth noting
for users of -stable.
lmbench showed a small regression in 3.0 where context switch latency was
increased and this has not been recovered yet. 3.3.6 was particularly
bad for low numbers of clients.
hackbench-pipes looks ok in comparison to 2.6.32. The "Time ratio" graph
shows that kernels are below the red line reflecting that most
kernels are faster. However, it also shows that 2.6.34 was the
"best" kernel and recent kernels have regressed slightly
hackbench-sockets regressed badly after 2.6.34 until 3.3 which should be
investigated. Again this is most obvious in the Time Ratio graph
pipetest is showing major regressions in latency since some time between 2.6.34
and 2.6.39.
==========================================================
Machine: hydra
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__scheduler-performance/hydra/comparison.html
Arch: x86-64
CPUs: 1 socket, 4 threads
Model: AMD Phenom II X4 940
Disk: Single Rotary Disk
Status: pipetest is particularly bad.
==========================================================
starve is generally ok although again, 3.0 and 3.1 both regressed on System
CPU time. This was improved on kernels after that but it's still
a little worse than 2.6.32 was.
lmbench shows no regression in 3.0 unlike on arnold but later kernels are
much worse with the latency of 3.4 being generally higher than it
was in 3.2
hackbench-pipes generally looks ok.
hackbench-sockets is generally bad. 3.1 was particularly bad and while
3.4 has improved the situation a bit, it is still worse than 2.6.32.
pipetest is showing major regressions. 3.2 regressed particularly badly.
==========================================================
Machine: sandy
Result: http://www.csn.ul.ie/~mel/postings/mmtests-20120424/global-dhp__scheduler-performance/sandy/comparison.html
Arch: x86-64
CPUs: 1 socket, 8 threads
Model: Intel Core i7-2600
Disk: Single Rotary Disk
Status: Generally great.
==========================================================
starve is generally ok. 3.0 regressed in terms of System CPU time but
recent kernels are very good. This might reflect that a lot
of people are testing with later Intel processors to the
detriment of older models.
lmbench is looking superb.
hackbench-pipes looks great.
hackbench-sockets does not look as great but it's still very good.
pipetest is generally looking good in comparison to 2.6.32. However,
I am concerned that 3.4 is worse than 3.3.
--
Mel Gorman
SUSE Labs
--
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/