On Tue, 8 Nov 2011 10:28:57 +0100 (CET) Mikael Abrahamsson <swmike@xxxxxxxxx> wrote: > > Hello. > > I have been running mdraid->cryptsetup/LUKS->lvm->xfs on Ubuntu AMD64 with > RAID5 and now RAID6 for quite some time, dating back to 2.6.27. Around > 2.6.32 I saw quite a bit of regression in write performance (probably the > implementation of barriers), 2.6.35 was acceptable, 2.6.38 was really > really bad, and 3.0 is like a rocket. Best of them all. > > I'm talking about 10-20x in different in write performance on my workload, > in combination with the older kernels throwing me page allocation failures > when the write load gets high, and also quite often the machine would just > freeze up and had to be rebooted. > > With 2.6.38 I was down to 6-10 megabyte/s write speed, whereas 3.0 seem to > give me 100+ megabyte/s with the exact same workload, I've seen up to 150 > megabyte/s writes at good times. This is on a box with AES-NI, so the > crypto is not the limiting factor. That is an amazing improvement. I wish I know what caused it I really have no idea. You have quite a deep stack there and the change could be anywhere. Still, it is good to hear such positive reports - thanks! NeilBrown > > I have from time to time sent out an email regarding my page allocation > failures, but never really got any takers on trying to fault find it, my > tickets with ubuntu also never got any real attention. I haven't really > pushed it super hard with 3.0, but I've thrown loads at it that would make > 2.6.38 lock up. > > Just wanted to send in this success report that this finally seem to have > seen some really nice improvements! >
Attachment:
signature.asc
Description: PGP signature