[Linux-cluster] NFS/GFS performance tuning
corey.kovacs at gmail.com
Fri Feb 20 23:04:12 UTC 2009
Figuring max xfer over a 1G network is 125MBs (100 is reality). Put
the overhead of
NFS into the mix and you go somewhat lower, around 60MB max, then are you using
tcp or udp? There are lots of small things that can be done.
can be bumped to increase the network stack buffers, which can aid in
NFS xfer greatly.
All in all, what you are seeing doesn't seem too bad. Best test is,
can you stand it. If you
don't require anything more, then maybe you'll be just fine. On the
other hand if you need
more write throuput, you can prolly get a few more megs if you really hunt.
By the way, you don't say what's on the back end? How many clients and
hence how many
threads do you plan on running etc. All things to consider.
On Fri, Feb 20, 2009 at 9:20 PM, David Merhar <merhar at arlut.utexas.edu> wrote:
> Is this NFS performance to be expected? If so, we can move forward, but I
> would like to hear one way or the other.
> On Feb 19, 2009, at 7:07 AM, David Merhar wrote:
>> We have the cluster running without problem (RHEL5.3, gfs2). The
>> improvements over the last 6 months are noticeable, and appreciated.
>> The problem:
>> Local file access (cluster member) is 160/135 r/w.
>> rsh access across the network is 130/115.
>> nfs access across the network is 45/27.
>> Our test is
>> dd if=/dev/zero of=/nfs/mount/test_file bs=1M count=10000
>> The remote system is also 5.3.
>> We've tinkered with a whole lot of settings in nfs server/client, gfs2
>> mount and tune parameters, and the network itself. We've also played with
>> dd flags. Yet we can't cross the 27 Mb/s write threshold. (But we're
>> willing to try again.)
>> Any advice?
>> Linux-cluster mailing list
>> Linux-cluster at redhat.com
> Linux-cluster mailing list
> Linux-cluster at redhat.com
More information about the Linux-cluster