Just as a note, RAID-5 does not always give worse performance. Most disks
have a limit to the number of operations they can perform a second, if you
are dealing with a writing a large number of small files you can see an
improved performance using RAID-5 over a normal disk area in the following
way;
(This exampleassumes that you are not reaching any CPU or IO bandwidth
limitations)
If you are writing a large number of files per second to disk (say 400 2Kb
files as can happen with large scale data collection/processing systems).
If these files are not going to be placed sequentially on the disk (doing
sequential writes to a disk improves performance if the writes are
synchronised with the head movement because there will be virtually no seek
time), and your disks are limited to x IOop/s a second (say 200 for example
), then by using a RAID-5 array split over n disks (say, 5 for example)
with the RAID data chunk larger than the largest file you are writing, you
will be able to get nx IOop/s (1000 for the example figures given) as
opposed to x IOop/s a second (200 for the example). This will mean you have
a potentially be able to write 2000Kb/s (as opposed to 400Kb/s [2Kb *
200IOop/s]), if you now deduct 20% for the RAID parity information you are
down to about 1600Kb/s, and you will probably loose more through for
various operation reasons, but you should still experience a greater total
throughput than writing to a single disk.
This problem is one I've come accross on a server before, and I think shows
a case where striping offers a speed improvment and RAID-5 gives you a good
level of data protection.
Please note, I'm not putting this example up is not to say Stephen is
wrong, it's mearly to show that under some circumstances RAID-5 can be a
better solution.
Al
"Stephen C. Tweedie" <[EMAIL PROTECTED]> on 11/11/98 18:39:06
To: Osma Ahvenlampi <[EMAIL PROTECTED]>
cc: [EMAIL PROTECTED], Stephen Tweedie <[EMAIL PROTECTED]> (bcc:
Al Sutton/CHASE)
Subject: Re: RAID and partitions?
[stuff deleted]
> Immediate thought is that number of seeks may increase, resulting in
> more time lost to seek times. On the other hand, sustained bandwidth
> is greater for all volumes. Has anyone run benchmarks to find out
> which is the more powerful factor?
For writes, you will get significantly worse performance, because each
raid-5 write has to span multiple disks (at least the data and parity
disk for that chunk). For reads you have the same tradeoffs as when
deciding whether to place different partitions on the same disk.
[stuff deleted]
--Stephen
