PERC 5 and PERC 6 performance notes

[John LLOYD]

A few months ago I posted a message about RAID-0 performance not meeting
expectations with PERC5 and PERC6 controllers (we have PERC5/e on
PE6850s and PERC6/e on R900s).

Well, of course, after updating the firmware last month, the performance
is substantially improved, at least on the PERC6.  (I used bonnie++ to
perform these tests).

Environment:  R900, 2.93GHz, SLES10 SP2 64-bit, 32GB, PERC 6/e, MD1000,
filesystem XFS
 old firmware:  RAID-0, 512kB stripe, write-back, 6x400GBx15k, write:
105 Mbyte/sec, read 409 Mbyte/sec
 new firmware:  RAID-0, 64kB stripe, write-back, 6x400GBx15k,  write:
384 Mbyte/sec, read 622 Mbyte/sec

The new firmware version is 6.1.1-0047

(I have some other tests varying the stripe size with the new firmware
and for RAID-0 it makes little, less than 5%, difference.  So please do
not point out to me that the stripe size difference will affect
throughput.)

 old firmware: RAID-5, 512kB stripe, write-back, (5+1)x400GBx15k, write:
301 Mbyte/sec, read 379 Mbyte/sec
 new firmware: RAID-5, 512kB stripe, write-back, (5+1)x400GBx15k, write:
359 Mbyte/sec, read 387 Mbyte/sec

The PERC5 firmware update has a caveat that dual-connect MD1000s should
not run the latest (as of Oct 2008)  release, so we didn't update that.

FYI the linux "MD" driver, with a pair of disks configured JBOD
("raid-0" according to the PERC, but just a single disk in the
configuration rather than two or more), and striped using Linux rather
than the PERC, was very fast too.  Two disks write at 133 Mbyte/sec and
read at 156MByte/sec.  Six disks write at 306Mbyte/sec and read at 607
Mbyte/sec.  Both with md stripesize of 512k.

The conclusion is that disk performance is a complicated story: 6 disks
is not three times 2 disks.  And firmware matters.  And, (although I
don't show it here), stripe size does not matter much, when dealing with
larger files.

YMMV; these figures are not a guarantee, past performance is not a
guarantee of future returns.  IANAL.  IANAMR.


--John