Messages in this thread | | | Date | Tue, 23 Feb 2010 15:48:18 +1100 | From | Dave Chinner <> | Subject | Re: disk/crypto performance regression 2.6.31 -> 2.6.32 (mmap problem?) |
| |
On Mon, Feb 22, 2010 at 01:39:37PM +0100, Mikael Abrahamsson wrote: > On Fri, 19 Feb 2010, Mikael Abrahamsson wrote: > >> I have a system with 8 gigs of ram, c2d 2.4GHz and md >> raid6->crypto->lvm/lv->xfs that I've had for quite a while. It used to >> be raid5, but with 2.6.32 I was able to restripe it into a raid6. > > This is not a 2.6.32 regression, the behaviour is similar in 2.6.31 I > have verified during the weekend. > > It seems iostat isn't really able to see exactly where the bottleneck is, > and my guess right now is that it's indeed the md raid6 layer that is > causing my performance problems, even though it doesn't look like it in > iostat.
XFS issues IO barriers by default. They were recently enabled in md for raid5/6, so that might be the cause of the slowdown/latency. You could try using the "nobarrier" mount option to see if this make the problem go away, but beware that this can result in filesystem corruption if the machine crashes.
If it is not barriers that are causing this, then the other thing you might want to look at is if XFS is configured with lazy-count=1 (xfs_info <mntpt>). If it is not enabled (0), then a significant amount of latency could be coming from the superblock buffer being locked during transaction commits. Unfortunately enabling this feature is an offline operation (via xfs_admin) so enabling may not be feasible for you.
Cheers,
Dave. -- Dave Chinner david@fromorbit.com
| |