On 04/03/2013 21:50, Nick Evans wrote: > iostat (slow case): > > tty mfid0 mfid1 cd0 cpu > tin tout KB/t tps MB/s KB/t tps MB/s KB/t tps MB/s us ni sy in id > 0 479 0.00 0 0.00 80.00 2 0.16 0.00 0 0.00 11 0 1 0 88
> 0 600 0.00 0 0.00 32.00 3 0.09 0.00 0 0.00 13 0 1 0 86 > 0 553 0.00 0 0.00 0.00 0 0.00 0.00 0 0.00 13 0 1 0 85 > 0 1631 0.00 0 0.00 38.40 5 0.19 0.00 0 0.00 14 0 1 0 85 > 0 665 0.00 0 0.00 32.00 4 0.12 0.00 0 0.00 14 0 1 0 85 > 0 737 0.00 0 0.00 32.00 1 0.03 0.00 0 0.00 15 0 1 0 84 > 0 485 32.00 2 0.06 127.05 770 95.56 0.00 0 0.00 13 0 1 0 85 > 0 624 0.00 0 0.00 126.95 61 7.55 0.00 0 0.00 13 0 1 0 85 > 0 1586 0.00 0 0.00 32.00 1 0.03 0.00 0 0.00 14 0 1 0 85 > 0 565 0.00 0 0.00 32.00 1 0.03 0.00 0 0.00 14 0 1 0 85 > 0 488 0.00 0 0.00 32.00 2 0.06 0.00 0 0.00 14 0 1 0 85 > 0 515 0.00 0 0.00 32.00 1 0.03 0.00 0 0.00 13 0 1 0 85 > iostat (fast case): > > tty mfid0 mfid1 cd0 cpu > tin tout KB/t tps MB/s KB/t tps MB/s KB/t tps MB/s us ni sy in id > 0 542 0.00 0 0.00 32.00 300 9.37 0.00 0 0.00 14 0 6 1 79 > 0 540 32.00 10 0.31 32.00 318 9.93 0.00 0 0.00 14 0 5 1 79 > 0 688 0.00 0 0.00 32.00 331 10.33 0.00 0 0.00 14 0 5 1 79 > 0 1863 8.00 2 0.02 32.00 300 9.37 0.00 0 0.00 14 0 6 1 79 > 0 720 0.00 0 0.00 32.00 306 9.55 0.00 0 0.00 14 0 5 1 80 > 0 565 0.00 0 0.00 32.00 299 9.33 0.00 0 0.00 14 0 6 0 79 > 0 1470 0.00 0 0.00 116.51 3413 388.27 0.00 0 0.00 10 0 5 1 > 83 > 0 553 0.00 0 0.00 32.00 303 9.46 0.00 0 0.00 13 0 7 1 79 > 0 1730 0.00 0 0.00 32.00 337 10.52 0.00 0 0.00 14 0 6 0 80 > 0 649 0.00 0 0.00 32.00 342 10.68 0.00 0 0.00 14 0 6 1 79 > 0 901 0.00 0 0.00 32.00 317 9.90 0.00 0 0.00 14 0 6 1 80 > 0 1654 0.00 0 0.00 32.00 328 10.24 0.00 0 0.00 15 0 5 1 79 I don't know if you've noticed, but: 1) In your slow case, the load is almost entirely in userspace (USR), i.e. mysql, while in the fast case, the kernel load (SYS) is 5x-6x greater. Unless the first case actually excercises your disks more (unlikely since you have almost no IO load at all) it means that it's mysql that is slow, not the kernel. 2) Ordinary iostat doesn't show the disk busy-time estimation. You need to use "-x" (e.g. "iostat -x 1") to see the service time and busy percentage stats. 3) you seem to have bursts of traffic, probably writing, which may be caused by the file system IFF mysql is not doing anything to fsync the data sooner (i.e. you have everything in a single transaction), but the burst in the "fast" case is 388 MB while the burst in the "slow" case is 100 MB 4) your benchmark seems to be single-threaded 5) You have a lot of 32 KB transactions, which probably means you have created the file system in 9.x with the default options, leading to 32 KB blocks. This is probably not the source of your current problems, but in the future you might want to match the block size to innodb's block size (16 KiB). You should also do that on the RAID array level and match the block alignments. 6) Are you using innodb_file_per_table? You probably should.
signature.asc
Description: OpenPGP digital signature
