Re: [Xen-users] Fast write performance, slow read performance in NFS

[Nathan <have.an.email@xxxxxxxxx>]

On Thu, Oct 23, 2008 at 2:54 PM, Linspeed <lee@xxxxxxxxxxxx> wrote:
>
> Hello
>
> I need to setup shared storage for several virtual machines I'm running in
> Xen.
> I've gone for NFS.
>
> Native performance on the disk that is shared is as follows:
> Read: 199 MB/second
> Write: 158 MB/second
>
> Now I've done lots of testing with NFS options on the client (the virtual
> machine).
> I've managed to get the same write performance as I get natively, but read
> performance is way below half that of native. Around 70 MB/second.

158 MB/second, counting overhead that's something like 1.5 Gbit/second
. . . What kind of network is this?

This is very off the cuff, but in your place I'd test by maximizing
your MTU and varying

 - between tcp and udp (you've tested udp)

 - nfsv3 and nfsv4 (you've tested nfsv3)

 - [r|w]sizes lower than MTU size or much bigger (you've tested much
bigger, and your default was probably 8192 with MTU of 1500 which is
no good). With default size of 8192 octets you get a boost in
performance when you push MTU to contain that size, MTU 9000 is good,
I suppose that with your network you can do that :-)

Don't forget NFS server-side cache, I suppose you can tune that, it
shouldn't change your test values though(otherwise there's something
wrong with your tests)

-- 
HTH,
Nathan

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users