[ltt-dev] [RFC PATCH] block: Fix bio merge induced high I/O latency

[bgamari@gmail.com (Ben Gamari)]

I'm not sure if this will help, but I just completed another set of
benchmarks using Jens' patch and a variety of device parameters. Again,
I don't know if this will help anyone, but I figured it might help
quantify the differences between device parameters. Let me know if
there's any other benchmarking or testing that I can do. Thanks,

- Ben


			mint		maxt
==========================================================
queue_depth=1, slice_async_rq=1, quantum=1, patched
anticipatory		25 msec		4410 msec
cfq			27 msec		1466 msec
deadline		36 msec		10735 msec
noop			48 msec		37439 msec
==========================================================
queue_depth=1, slice_async_rq=1, quantum=4, patched
anticipatory		38 msec		3579 msec
cfq			35 msec		822 msec
deadline		37 msec		10072 msec
noop			32 msec		45535 msec
==========================================================
queue_depth=1, slice_async_rq=2, quantum=1, patched
anticipatory		33 msec		4480 msec
cfq			28 msec		353 msec
deadline		30 msec		6738 msec
noop			36 msec		39691 msec
==========================================================
queue_depth=1, slice_async_rq=2, quantum=4, patched
anticipatory		40 msec		4498 msec
cfq			35 msec		1395 msec
deadline		41 msec		6877 msec
noop			38 msec		46410 msec
==========================================================
queue_depth=31, slice_async_rq=1, quantum=1, patched
anticipatory		31 msec		6011 msec
cfq			36 msec		4575 msec
deadline		41 msec		18599 msec
noop			38 msec		46347 msec
==========================================================
queue_depth=31, slice_async_rq=2, quantum=1, patched
anticipatory		30 msec		9985 msec
cfq			33 msec		4200 msec
deadline		38 msec		22285 msec
noop			25 msec		40245 msec
==========================================================
queue_depth=31, slice_async_rq=2, quantum=4, patched
anticipatory		30 msec		12197 msec
cfq			30 msec		3457 msec
deadline		35 msec		18969 msec
noop			34 msec		42803 msec



On Tue, 2009-01-20 at 15:22 -0500, Ben Gamari wrote:
> On Tue, Jan 20, 2009 at 2:37 AM, Jens Axboe <jens.axboe at oracle.com> wrote:
> > On Mon, Jan 19 2009, Mathieu Desnoyers wrote:
> >> * Jens Axboe (jens.axboe at oracle.com) wrote:
> >> Yes, ideally I should re-run those directly on the disk partitions.
> >
> > At least for comparison.
> >
> 
> I just completed my own set of benchmarks using the fio job file
> Mathieu provided. This was on a 2.5 inch 7200 RPM SATA partition
> formatted as ext3. As you can see, I tested all of the available
> schedulers with both queuing enabled and disabled. I'll test the Jens'
> patch soon. Would a blktrace of the fio run help? Let me know if
> there's any other benchmarking or profiling that could be done.
> Thanks,
> 
> - Ben
> 
> 
> 			mint		maxt
> ==========================================================
> queue_depth=31:
> anticipatory		35 msec		11036 msec
> cfq			37 msec		3350 msec
> deadline		36 msec		18144 msec
> noop			39 msec		41512 msec
> 
> ==========================================================
> queue_depth=1:
> anticipatory		45 msec		9561 msec
> cfq			28 msec		3974 msec
> deadline		47 msec		16802 msec
> noop			35 msec		38173 msec