NFSv3 and v4 slow with newnfs code

[Adam Vande More]

On Mon, Aug 23, 2010 at 12:53 AM, Nikolai Schupbach <nikolai at net24.co.nz>wrote:

> We are doing some performance testing on a new system. We have a
> OpenSolaris NFS server sharing a folder on a ZFS filesystem and a FreeBSD
> 8.1 NFS client. The machines are directly connected using 10GbE (no switch
> in-between).
>
> Below are the performance figures we attained when doing simple 10GB dd
> write (dd if=/dev/zero of=/mnt/file.tmp bs=1M count=10240) and read (dd
> of=/dev/null if=/mnt/file.tmp) tests over NFS from the FreeBSD client using
> various mount options.
>
> We performed these tests numerous times and all results are roughly the
> same for each test. We have tuned kern.ipc.maxsockbuf,
> net.inet.tcp.recvspace and net.inet.tcp.sendspace. This didn't result in any
> significant differences in the test results. Both NFS client and server NICs
> have MTU set to 9000; this improves performance noticeable.
>
> Currently it appears sticking with the stable NFSv3 code yields the best
> results. Both NFSv3 and NFSv4 with the newnfs code has disappointing
> performance. We installed Linux on the client machine as a test and
> unfortunately Linux has the best performance by far.
>
> Are there any other options we can use to improve the performance of NFSv4
> for large sequential writes and reads?
>

You can use a UDP mount, should help a little.  Have you measured your NIC's
performance to see if that is the bottleneck.  Perhaps the driver is subpar.

-- 
Adam Vande More