Re: [Hampshire] [hardware] RAID5 - hardware or software, bas…

Top Page

Reply to this message
Author: Adrian Bridgett
Date:  
To: Hampshire LUG Discussion List
Subject: Re: [Hampshire] [hardware] RAID5 - hardware or software, based?
On Mon, May 18, 2009 at 17:06:54 +0000 (+0000), Andy Smith wrote:
> Indeed, unless under very heavy write load I expect software RAID
> will be fine.


To explain this further, if you do a write on RAID-5, you often have
to _read_ the disks first - this can add lots of latency.

Personally, I do RAID-1 (software or hardware just _not_ fakeraid
rubbish) on anything "small" (e.g. upto 4 disks). Anything "big"
(read "expensive"), get a decent _battery backed_ hardware controller
or an external RAID array with the same.

If you do lots of small writes and wait for them to hit disk, this can
make a massive difference (I've personally seen a box go from 100% IO
bound to 1% IO when we discovered that the server had been shipped
with battery backed RAID controller, but in "write though" not "write
back" mode). Oddly enough the database went just a tad quicker.

Regarding "enterprise" disks, I think it's mostly a completely fallacy
reliability wise. Performance wise, a 15Krpm drive will out perform a
5400rpm drive. However there is another issue - how hard a disk will
try and recover data.

Western Digital "RE" (RAID Edition) drives give up quite early - they
assume you are running in a RAID (not RAID-0) environment, so if they
have trouble reading something, they give up quickly and carry on.

Most desktop drives try _really_ hard (and don't give up) - thus they
effectively "lock up" for a few mins upon encountering an error.

Adrian
--
Email: adrian@??? -*- GPG key available on public key servers
Debian GNU/Linux - the maintainable distribution -*- www.debian.org