Ken Burner wrote:
>    We're thinking of installing RAID arrays to replace JBOD collections
>on our central AFS cell.

<...snip...>

A couple years ago, when we started averaging 250-300 user volumes per
spindle, we did the calculations about how many tapes we would have to
spin if one of those spindles ever failed, and...well...it was bad.

Our primary concern, therefore, was fault tolerance, not performance.
However, imnsho, RAID in this type of application will always take a
performance hit.  Now, before you all flame me for this, hear the
reasons why.  I did some performance benchmarks
(http://www.cs.wisc.edu/~thomas/afs/disk-data.html), which, while not
exhaustive, put the RAID boxes we purchased on a par with jbod disks
we bought about the same time.  Where I believe you take the
performance hit is that with jbod we had 7 * 4Gig disks, with each
disk seeking, reading, writing, etc., independently.  With RAID, we
have 4 * 9Gig disks (+1 parity disk), with _all_ the disks involved in
_all_ the accesses.  Now, depending on how well the fileserver can
parallelize disks accesses (not well now, but stay tuned, it's
supposed to get better), this may or may not be a problem, but in a
highly parallel environment, you can get alot more work done on jbod
than you can on RAID.  So, remember that for total throughput, you are
comparing your RAID array to a single disk, not your jbod array.

That being said, I'm glad we made the move to RAID for some of our
data.  We have given our users the choice of paying the premium for
RAID space or staying on jbod for non-home-directory space that they
'buy'.  So far, no one has bought RAID space.  Our backup system is
very good, and we haven't had any problems going this way.  You lose a
day's changes, and so you have to know that you can live with that.
This philosophy may or may not work in your environment.

Of course, given this, I don't think RAID makes sense for use with
replicated data, because with replicated data you already have fault
tolerance.

--
David Thompson  <[EMAIL PROTECTED]>

Associate Researcher                    Department of Computer Science
University of Wisconsin-Madison         http://www.cs.wisc.edu/~thomas
1210 West Dayton Street                 Phone:    (608)-262-1017
Madison, WI 53706-1685                  Fax:      (608)-262-6626
--



Reply via email to