On Fri, 2011-12-16 at 22:57 +0800, Asias He wrote: > On 12/15/2011 08:15 PM, Sasha Levin wrote: > > When using AIO, submit all requests which exists in the vring in a > single > > io_submit instead of one io_submit for each descriptor. > > > > Benchmarks: > > > > Short version: 15%+ increase in IOPS, small increase in BW. > > > > Read IOPS: > > Before: > > vda: ios=291792/0, merge=0/0, ticks=35229/0, in_queue=31025, > util=61.30% > > I guess you are reading the wrong IOPS number, the 'ios' is the number > of ios performed by all groups, not the IOPS result. Find the > 'iops' ;-) > > So, Here is the number without/with this patch. > > (seq-read, seq-write, rand-read, rand-write) > > Before: > read : io=98304KB, bw=63015KB/s, iops=15753, runt= 1560msec > write: io=98304KB, bw=56823KB/s, iops=14205, runt= 1730msec > read : io=98304KB, bw=62139KB/s, iops=15534, runt= 1582msec > write: io=98304KB, bw=53836KB/s, iops=13458, runt= 1826msec > > After: > read : io=98304KB, bw=63096KB/s, iops=15774, runt= 1558msec > write: io=98304KB, bw=55823KB/s, iops=13955, runt= 1761msec > read : io=98304KB, bw=59148KB/s, iops=14787, runt= 1662msec > write: io=98304KB, bw=55072KB/s, iops=13768, runt= 1785msec > > Submit more io requests in one time is not supposed to increase the > iops > or bw so dramatically. > > I even tried to submit all read/write ops in one io_submit which still > ends up with very little iops or bw improvement. Did you test it vs /dev/shm? These are the results I see from the same test: Before: read : io=1157.7MB, bw=118110KB/s, iops=29527 , runt= 10037msec write: io=1083.6MB, bw=110867KB/s, iops=27716 , runt= 10008msec After: read : io=1304.4MB, bw=133337KB/s, iops=33334 , runt= 10017msec write: io=1292.4MB, bw=132087KB/s, iops=33021 , runt= 10019msec -- Sasha. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
