On Sun, 31 Aug 2014 16:05:50 +0000 Markus Stockhausen <stockhausen@xxxxxxxxxxx> wrote: > Hello, > > I had some time for additional benchmarks. My initial numbers > were too low because I didn't push the system to its limits with > high parallelism. > > Test program http://www.corpit.ru/mjt/iot.c > settings: 180 seconds random write, 3 runs, 64 threads, direct I/O > command line: iot -d -W64 -b <blocksize> -t 180 /dev/md0 > > bsize w/o patch w/o patch w/o patch > min avg max > 4K 317KB/s 318KB/s 319KB/s > 8K 611KB/s 614KB/s 617KB/s > 16K 1153KB/s 1153KB/s 1154KB/s > 32K 2239KB/s 2253KB/s 2263KB/s > 64K 3944KB/s 3959KB/s 3978KB/s > 128K 7792KB/s 7891KB/s 8025KB/s > 256K 15456KB/s 15574KB/s 15805KB/s > 512K 32406KB/s 32564KB/s 32799KB/s > 1024K 57378KB/s 57965KB/s 58550KB/s > > bsize with patch with patch with patch real math gain > min avg max gain in theory > 4K 417KB/s 420KB/s 423KB/s +32,1% +66% > 8K 806KB/s 818KB/s 826KB/s +33,2% +66% > 16K 1504KB/s 1518KB/s 1531KB/s +31,6% +66% > 32K 2993KB/s 3001KB/s 3016KB/s +33,2% +66% > 64K 5396KB/s 5477KB/s 5543KB/s +38,3% +66% > 128K 8971KB/s 9082KB/s 9169KB/s +15,1% +25% > 256K 15406KB/s 15545KB/s 15751KB/s -0,2% 0% > 512K 32361KB/s 32619KB/s 32981KB/s +0,1% 0% > 1024K 58134KB/s 58239KB/s 58419KB/s +0,4% 0% > > As I got no feedback from Neil until now: > > - Shall I resend the patches with the new numbers? > - What else can I do to push the patches further? Sorry I've been silent - there has been lots of other stuff happening. Thanks for the extra testing - it definitely looks good. I'll try to move this along soon, no need to resend. Thanks, NeilBrown
Attachment:
signature.asc
Description: PGP signature
