On Fri, 19 Nov 2010, Mike Snitzer wrote: > On Sun, Oct 24 2010 at 9:03pm -0400, > Mikulas Patocka <mpatocka@xxxxxxxxxx> wrote: > > > dm-kcopyd: Delayed unplug of the queues > > > > This patch improves performance aby about 20% when writing to the snapshot > > origin. We keep track of two block devices to unplug (one for read and the > > other for write) and unplug them when exiting "do_work" thread. If there are > > more devices used (in theory it could happen, in practice it is rare), we > > unplug immediatelly. > > Again, more specifics on how you realized this 20% performance would be > appreciated. I'd imagine this patch is really where the increased > latency comes back (due to removing immediate unplug)? > > But does the fact that >2 devices will unplug immediately mitigate the > worst of the higher latency (e.g. for DM mirrors with >2 legs)? > > Mike Hi The test setup is the same as in previous email. The previous patch (removing SYNC) is already applied. cfq is used as the scheduler (as in previous tests). Sequential write throughput in MB/s (4k, 32k, 512k chunksize) without this patch: non-shared snapshots: 15.3, 18.5, 17.3 shared snapshots: 10.9, 24.0, 25.9 with this patch: non-shared snapshots: 14.4, 22.6, 23.0 shared snapshots: 13.2, 24.7, 26.1 Mikulas -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
