How can I make my HW RAID reads faster? (request for FAQ)

[Sitsofe Wheeler]

Matthias Saou wrote:

(Matthias: Thanks for freshrpms - it's been very handy over the years)

> Looking for information on XFS filesystem options, I found this very
> instructive thread :
> 
> http://oss.sgi.com/archives/xfs/2007-09/msg00322.html
> 
> I read about this simple change in one of the posts, which improved
> performance a lot for the person making the benchmark, so I ran it :
> # blockdev --setra 16384 /dev/sdb
> 
> I now get 500-520MB/s read speed with hardware RAID5 where I used to
> get 30-35MB/s... pretty amazing, I must admit.
> 
> Looking at the man page, this just sets a read-ahead value... I'm
> puzzled yet very happy!
> 
> Would any of the others who reported slow read speeds with their PERC5s
> and MD1000s care to also try it and report back to the list?

This small piece of wisdom/open secret for 2.6 kernels keeps having to be
dug up over and over again (along with noatime
(http://lwn.net/Articles/244829/ ) and stripe alignment
(http://insights.oetiker.ch/linux/raidoptimization.html )). You can see me
finding the similar results earlier last year on
http://lists.us.dell.com/pipermail/linux-poweredge/2007-August/032471.html
(the pity is that it doesn't seem to have helped anyone else). It would be
good if this sort of thing made it into a FAQ because it is such a common
tip (it should be the among the first thing that people try when running
all the standard hardware RAID cards).

Obviously one needs to be aware of drawbacks of doing this (might slow down
random I/O of lots of small files, focuses on one benchmark) but it is
still handy to know and compared to other tips it is relatively easy to
try.

-- 
Sitsofe | http://sucs.org/~sits/