2007-02-26 13:08:56

by Ananiev, Leonid I

Subject: Cfq first request patch changes IO performance

The patch "cfq-iosched: defer slice activation to first request being
active"
http://www.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=c
ommitdiff;h=44f7c16065c83060cbb9dd9b367141682a6e2b8e;hp=99f9628aba4d8fb3
b8d955c9efded0d0a1995fad

Sysbench fileio (random read/write mix) -15%
Tiobench sequential write -3%
Tiobench random write -5%
Iozone record re-write -21%
Iozone mmap sequential read -20%
Iozone mmap random write +45%
Aio-stress direct sequential read +29%
For other 3 tiobench, 7 aiostress and 13 iozone tests the result changes
are less than variation.

Considered patch does not impact on performance if all threads have
equal IO activity (tiobench).
If single thread has IO activity the performance becomes unstable and is
decreased mainly.

Leonid

2007-02-27 20:40:17

by Jens Axboe

Subject: Re: Cfq first request patch changes IO performance

On Mon, Feb 26 2007, Ananiev, Leonid I wrote:
> The patch "cfq-iosched: defer slice activation to first request being
> active"
> http://www.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=c
> ommitdiff;h=44f7c16065c83060cbb9dd9b367141682a6e2b8e;hp=99f9628aba4d8fb3
> b8d955c9efded0d0a1995fad
>
> Sysbench fileio (random read/write mix) -15%
> Tiobench sequential write -3%
> Tiobench random write -5%
> Iozone record re-write -21%
> Iozone mmap sequential read -20%
> Iozone mmap random write +45%
> Aio-stress direct sequential read +29%
> For other 3 tiobench, 7 aiostress and 13 iozone tests the result changes
> are less than variation.
>
> Considered patch does not impact on performance if all threads have
> equal IO activity (tiobench).
> If single thread has IO activity the performance becomes unstable and is
> decreased mainly.

Yep there are a few unfortunate side effects, I have later fixed those.
I just need to split that patchset up and commit it.

--
Jens Axboe