On Mon, Jan 01 2007, Mark Lord wrote: > Rene Herman wrote: > >Tejun Heo wrote: > > > >>Everything seems fine in the dmesg. Performance degradation is > >>probably some other issue in -rc kernel. I'm suspecting recently > >>fixed block layer bug. If it's still the same in the next -rc, > >>please report. > > > >In fact, it's CFQ. The PATA thing was a red herring. 2.6.20-rc2 and 3 > >give me ~ 24 MB/s from "hdparm t /dev/hda" while 2.6.20-rc1 and below > >give me ~ 50 MB/s. > > > >Jens: this is due to "[PATCH] cfq-iosched: tighten allow merge > >criteria", 719d34027e1a186e46a3952e8a24bf91ecc33837: > > > >http://www2.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=719d34027e1a186e46a3952e8a24bf91ecc33837 > > > > > > > >If I revert that one, I have my 50 M/s back. config and dmesg attached > >in case they're useful. > > Wow.. same deal here -- sequential throughput drops from 40MB/sec to > 28MB/sec > with CFQ -- whereas the anticipatory scheduler maintains the 40MB/sec. > > Jens.. I wonder if the new merging test is a bit too strict? > > There are four possible combinations, and the new code > allows merging for two of them: sync+sync and async+async. > > But surely one of (not sure which) sync+async or async+sync may also be > okay? > Or would it?
Async merge to sync request should be ok. But I wonder what happens with hdparm, since it seems to trigger one of these tests. Very puzzling. I'll dive in and take a look. > This is a huge performance hit. Indeed, not acceptable of course. And not intentional :-) -- Jens Axboe - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/