[tech] Meito

[Nick Bannon]

On Fri, Jul 11, 2003 at 01:26:01PM +0800, Leighton Haynes wrote:
> On Fri, Jul 11, 2003 at 01:22:32PM +0800, Grahame Bowland wrote:
> > I'd like to reinstall Meito with Linux rather than Solaris. The disk
> > performance is hideously bad, which makes it fairly worthless as a
> > fileserver. I suspect it just doesn't know how to talk to the IDE
> > chipset at a reasonable speed.

Good move.

We do need better network FS performance at the UCC:

 * Clean up morwong - 9GB SCA discs or better, probably some mirroring.
   I'll probably buy these:
http://cgi.ebay.com/ws/eBayISAPI.dll?ViewItem&category=11160&item=2739673102

 * Reinstall meito. I've got nothing against Solaris, some of my best
   friends run Solaris^W^W^W^W^W^W <ahem> _I_ run Solaris, but yeah,
   meito sucks right now. Let's get a Sun with decent RAM and a disc
   array sometime.

 * When meito proves itself, we can put together a uber-PC with big(ger)
   IDE discs and a fast network.

[...]
> Can you back up everything to make sure nothing is lost during the
> reinstall, and can you provide me documentary evidence that the Linux
> box will suck less? ("I think Linux is better" does not count :P)
> 
> Leighton...

Yeah - in particular, could you do before and after, local and remote
(NFS) bonnie runs on it, Grahame? (actually, I can and will do that
too, but I still recommend you do it in controlled circumstances)

It's not as if there's any doubt whatsover that this will help, but I'd
like to know just how much Slowlaris' ass will be whoop-ed. ::-)

On Fri, Jul 11, 2003 at 03:40:38PM +1000, David Basden wrote:
> NFS file serving under linux sucks, mainly because it isn't able to do
> asyncronous writes.

No - I think the traditional "problem" was that the original user-space
Linux NFS server _only_ supported (or was it defaulted to?)
asynchronous writes, which made it seem fast, but it was playing fast
and loose with the data.

On other tacks ;

 * NFS locking between meito and morwong broke sometime around the end
   of May. Dunno why, the rpc.lockd/rpc.statd's have been restarted and
   indeed the machines have been rebooted since then, but the problem's
   still there. It means mutt refuses to write to mailboxes saved on
   meito.

 * Linux has advanced filing systems that shoot Solaris UFS (and VxFS)
   every which way to Sunday. Yes, Solaris UFS "scales" in that a 16
   CPU server can do some reasonable proportion of 16 times the I/O's
   and directory operations per second compared to a 1 CPU server, but
   when Linux can be a factor of 500 times faster from the start (I've
   measured...), the choice is clear.

> NFS3 allows for the server to write to disk without an ack from the client
> if specifically allowed on the connection. For some reason, Linux does not
> support this in the server, so writes to a Linux NFS server are mind
> bogglingly slow. The FAQ I read said this is unlikely to change.

Which FAQ? Can't see it here...
http://nfs.sourceforge.net/

> This would be a big reason to stay away from Linux for NFS serving, and the
> reason (from memory) it has been avoided in the past. If we didn't rely
> as much on NFS, it wouldn't be so much of an issue, but the whole thing
> is pretty much held together by NFS at the moment.
[...]

Held together by NFS to morwong (Tru64), that is. If meito sucked less
we might be using it by now, but so far that's been out of the
question.

Nick.

-- 
   Nick Bannon   | "I made this letter longer than usual because
nick-sig at rcpt.to | I lack the time to make it shorter." - Pascal