2009-03-31 17:39:51

by Vladislav Bolkhovitin

[permalink] [raw]
Subject: Re: [Scst-devel] Linux kernel and block I/O performance

Hello

Bart Van Assche, on 03/25/2009 09:17 PM wrote:
> Hello,
>
> For anyone who's running SCST performance measurements, please have a
> look at http://lwn.net/Articles/325307/. In this article it is
> explained why 2.6.25 and 2.6.29 perform three times faster on some I/O
> benchmarks than 2.6.26, 2.6.27 and 2.6.28. The relevant kernel commits
> are:
>
> 2.6.26: http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=18ce3751ccd488c78d3827e9f6bf54e6322676fb
> 2.6.29: http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=78f707bfc723552e8309b7c38a8d0cc51012e813

Interesting. I've just checked with 2.6.29 on the target. With iSCSI
still ext3 as a backstorage FS on the target a lot slower on writes than
xfs (25-100%, deviation between runs is very high). On my system it's
something like 50-80MB/s vs 100+MB/s. In both cases initiator formatted
the device in ext3. Test application was famous "dd if=/dev/zero
of=/mnt/q bs=512K".

Thanks,
Vlad