[OpenAFS] cpu load

[Jenkins, Steven]

These details are very helpful, and I see a few items for concern:

- 1 server to 90 clients *may* (or may not) be an issue.  Do you have a
specific volume that is frequently accessed, but isn't replicated?
Alternatively, are all (or many) of the read-write volumes on this one
server?  AFS can certainly scale beyond 1 server to 90 clients, but a
lot depends on access patterns and hardware.

- Your disk configuration is not balanced:

/dev/md4               97G   33G   60G  35% /vicepg
/dev/md3              2.9G   33M  2.7G   2% /viceph
/dev/md2              2.9G   33M  2.7G   2% /vicepi

and note the following (shortened from your raid config):

md0 (your root partition) is built on /dev/hda3 and /dev/hdc3
md4 is built on /dev/hda8 and /dev/hdc8
md3 is built on /dev/hda7 and /dev/hda7
md2 is built on /dev/hda5 and /dev/hda5

This configuration means that all of your disk accesses are competing
with each other.  You'd be *much* better off to put all of your AFS
accesses onto dedicated disks that don't compete with your OS.  I'd
suspect that's the biggest culprit.

Also, I don't see any benefit to having md3 & md2 -- you'd be better off
simply rolling them 3 all into md4 and having one large partition (at
least in terms of disk performance -- there are other issues to consider
with large partitions, such as backups and restores.  I personally would
probably have 3 32G partitions, but that's a personal preference, not
really backed up by any hard data.). =20

Note: you may be wanting to do some tricky things with startup and
volume accessibility -- if so, there are some problems with how you've
set this up, but if you're not trying to play startup tricks, then I
don't see how a single large volume would be worse (for your disk access
patterns) than what you have now.

- As for your question: "Is this mean even if I remove software raid.
Users can not=20
      do lots of compiles of files in AFS?" =20
  Distributed filesystems aren't always good idea if you're doing lots
of frequent file accesses, especially lots of writes.  Most people won't
have problems, but if you have users that are very sensitive to compile
times, you may see some issues.  I don't think that's an issue here,
unless you have several developers trying to build software in the same
volumes. =20

If you have an environment where several developers are building
software simultaneously (or lots of users doing lots of writes), you may
want to look at how you're distributing your server load.  But the gain
(which is in compile time speeds) you do by doing that will probably be
less than you'll see simply by splitting off your AFS devices from your
OS devices (which will help the problem you came to openafs-info on: cpu
load caused by your software raid).

Steven

 =20