From: mbeattie@sable.ox.ac.uk (Malcolm Beattie)
Subject: Re: NFS Performance Limitations?
Date: 12 Feb 1998 12:06:47 GMT

In article <6bugk5$ir5@usenet.srv.cis.pitt.edu>,
Mark Hahn <hahn@neurocog.lrdc.pitt.edu> wrote:
>: experiences with such systems. What sort of performance can I expect?
>
>wire speed, of course.  at least at 10 Mbps; at 100 I normally see
>more like 5 MB/s.  but then I don't have switched 100bT (full duplex.)
>
>: How much (in terms of # of clients, data transfer rates, or other
>: applicable metric) can I throw at such a system before it blows up and
>: dies?
>
>huh?  if you're talking 10bT, I doubt you can kill the server.
>I expect the kernel-mode server in 2.1/2.2 is a bit faster.
>
>: - Network architecture is flexible (if you've got any particular
>:    suggestions...). 
>
>switched 100bT, if you need the bandwidth.
>
>btw, you'll never see decent Linux NFS bandwidth with 2.0 if you
>leave it using the default 1K packets.  with 4-8K packets, wire
>speed isn't hard.

I've been benchmarking a couple of Pentium II 233 boxes (plus
another Pentium 133) each connected via two Tulip 21140-based
NICs to a 3Com SSII 3000 switch (i.e. 100 Mbps full-duplex).
Channel bonding the two ethernet cards on each PC together (with
the Beolwulf patch) doesn't seem to affect bandwidth although I
might be doing something wrong. ttcp gets 11500 KBytes/sec, ftp
gets 11000 KBytes/sec for get, 10000 Kbytes/sec for put, rcp gets
6000-8000 Kbytes/sec (varying), NFS gets 7500-8500 Kbytes/sec for 
reads and 4000 Kbytes/sec on writes. Those NFS figures are for
rsize=8192,wsize=8192. With rsize=4096,wsize=4096 reading drops
slightly to 7000-8000 Kbytes/sec but writes drop to 3000 Kbytes/sec.

So you don't quite get wire speed for NFS but it's not bad.
However, the NFS server in 2.0 is single-threaded and user-mode.
With a lot of clients, you'll have to worry about CPU and context
switching. They may be fine but you'll need to consider it. The
2.1/2.2 knfsd should get rid of those (potential) problems so if
you can wait (or last out with 2.0) until 2.2 knfsd is stable
enough for you then you'll be home and dry.

--Malcolm

-- 
Malcolm Beattie <mbeattie@sable.ox.ac.uk>
Oxford University Computing Services
"I permitted that as a demonstration of futility" --Grey Roger