Re: Strange performance characteristics with ZFS

That was my fault, I misread what you had said your configuration was.  The
problem I ran into was the same as you're seeing with the system waiting on
ZFS causing the NFS client to hang and everything becoming unresponsive.
The ls script should be running on the ZFS machine and executing ls on the
ZFS directory.  If you have a smaller file to try with first I'd recommend
it.  30GB of corrupt data is just pain.  To be honest, I gave up on NFS with
ZFS and started using Samba for everything after compiling samba3 without a
couple functions (check the ZFS vs Samba debugging results thread from
earlier this month) as it actually maxes out a 100Mbit network link for the
transfer.  That's something I could never get NFS with ZFS to come close to
doing.

Just to be sure, since you're using nfsd, that means you have the sharenfs
option of the zfs pool turned off?

-- Brian

On 7/17/07, Steven Schlansker <stevenschlansker_at_berkeley.edu> wrote:
>
> That turned out to be a particularly bad idea.  As soon as I executed
> that on the server, the client's cpu pegged!  Despite the fact that
> there is about 5MB/s transfer over the network, the client is no longer
> making any progress on the copy.  The nfsd processes on the bsd server
> are still happily chugging away serving data in the same fits and bursts
> as before, but the linux machine receiving the data doesn't seem to be
> doing anything with it.  I restarted the nfsd and mountd processes,
> I'll let it run for a bit...
>
> Aha.  It corrupted data.  Now I have to start the copy over again :/
>
> This is not good!  Anything else I can try?  (Hopefully without making
> the process fail :-p )
>
> Thanks!
>
> Brian Donnell wrote:
> > When I was experimenting with ZFS over NFS I had similar experiences.
> > Do me a favor and try something that sounds a little out there.  On a
> > shell on the ZFS machine set up a looping script that executes an ls on
> > the directory you're writing the file to once every second and watch
> > your transfer rates.  I noticed a marked improvement, but I could never
> > determine if it was ZFS or the client NFS implementation.
> >
> > -- Brian
> >
> >