Re: async vs. sync

[Bernd Schubert <bernd-schubert@web.de>]

[-- Attachment #1: Type: text/plain, Size: 2307 bytes --]

> >>My set-up has about the same performance data, however my big problem is
> >>that the server dies when I hook all my clients up... Raw throughput
> >
> >Fortunately that doesn't seem to happen in our case.
>
> Let me clarify - the server doesn't quite die, however it slows down
> drastically, having a load of ~7 while 99% CPU idle. The clients clog
> up, and for instance trying to mount anything just times out. If I
> unmount the busy share from the clients, everything settles down again.

We noticed a higher load with sync-mounts, usually we have 0-2, with 2.4. and 
sync-mounts it went until 3-4. But we also only used sync mounts for less 
than 24, so changed to async mounts as soon as the people complained.

[snip]

>
> I tried a little more performance number hunting:
>  - NFS copy and write to same device - 3Mb/s sustained (500Mb file)
>  - piozone direct write to RAID device - ~34-50Mb/s (depending on file
> size) - drbd network device - 933Mb/s
>
> Nothing really wrong here I'd say - not blazingly quick, but surely
> nothing to explain why everything just stops when I hook up my 12
> clients to this set-up?

Certainly not. We have 45 diskless booting clients.

[snip]

> >Well, as I said, when I tested our old server with sync mounts, it had the
> >same problem. Are you sure that your old server did not export
> > asynchronous?
>
> Doesn't seem so, no, the /etc/export is -extremely- simple:
>
> /export client1(rw,no_root_squash)

Which nfs-utils version is running on that server? With 1.0 the default 
changed from async to sync. I guess you only did security updates from your 
distribution and so there's <=0.3 installed.

With 2.6.7 it helped us a lot to give the no_wdelay option, maybe you should 
try it as well.

>
> Similarly, I'm not trying to be clever on the client side (/etc/fstab
> excerpt):
>
> server:/export /export nfs rw,hard,intr 0 0

You should make sure, that rsize and wsize are set to 8192 (cat /proc/mounts 
tells you the current value). Also, just try to use tcp-mounts, if you have 
an asynchronous net (e.g. the server and switch have a GBit interface, but 
the clients only 100MBit) this is necessary anyway. The same if the clients 
or server are using linux-2.6.x.


Cheers,
	Bernd

[-- Attachment #2: signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]