[wplug] RAID performance
Patrick Wagstrom
pwagstro at andrew.cmu.edu
Fri Mar 2 21:48:19 EST 2007
Michael Skowvron wrote:
> Patrick,
>
> On software vs. hardware RAID:
> Software RAID benchmarks higher than hardware RAID because you tend to
> have all CPU resources available for the RAID calculations during the
> test. On a heavily loaded system, software RAID could end up having a
> big impact on system performance because the software RAID calculations
> compete with other tasks for CPU resources. So, hardware RAID may be
> slower, but it will be consistent performance regardless of system load.
>
> For comparison, my 3Ware 7500 using 4x250GB PATA drives in a RAID5:
>
> Write:
> bfs:~/lmbench2/bin/i686-pc-linux-gnu # ./lmdd of=/safe/lmddtest move=1g
> sync=1
> 1073.7418 MB in 26.7281 secs, 40.1728 MB/sec
>
> Read:
> bfs:~/lmbench2/bin/i686-pc-linux-gnu # ./lmdd if=/safe/lmddtest
> 1073.7418 MB in 12.0338 secs, 89.2270 MB/sec
Maybe it was somewhat the artificial nature of the tests where I was
seeing some of the issues with the hardware RAID getting pummeled. In
particular, after running lmbench I saw that the hardware raid was much
faster for reads:
Hardware Write: ./lmdd of=dummy move=1g sync=1
1073.7418 MB in 25.6633 secs, 41.8396 MB/sec
Hardware Read: ./lmdd if=dummy
1073.7418 MB in 4.1905 secs, 256.2301 MB/sec
Software Write: ./lmdd of=/mnt/store/dummy move=1g sync=1
1073.7418 MB in 18.0682 secs, 59.4273 MB/sec
Software Read: ./lmdd if=/mnt/store/dummy
1073.7418 MB in 6.3530 secs, 169.0131 MB/sec
Of course, these are completely synthetic benchmarks but they're making
me feel a little better about the performance of the RAID 5 on the
system. Some of the write difference is most likely because of the
change from SATA I to SATA II. I'm not positive, but I don't believe
that the SATA I drives even have NCQ -- the hardware RAID machine is
about two years old.
Anyway, it's been a fun little adventure, I guess there probably isn't
anything seriously wrong with the hardware raid on the machine. Just
slow disks for writes. For reads I'm happy with the performance. A
PCI-X 133MHz card maxes out around 400MB/s, so getting 256MB/s over four
disks is pretty darn nice.
Incidentally, the stripe size for both sets of devices is identical at 256k.
--Patrick
More information about the wplug
mailing list