IRIX Network Forums
Unexpected NFS performance issues - Printable Version

+- IRIX Network Forums (//forums.irixnet.org)
+-- Forum: SGI/MIPS (//forums.irixnet.org/forum-3.html)
+--- Forum: Beginners' Questions/Tips (//forums.irixnet.org/forum-6.html)
+--- Thread: Unexpected NFS performance issues (/thread-2197.html)



Unexpected NFS performance issues - elmy - 06-06-2020

Hello there!

Just to quickly introduce myself as I registered just yesterday: I've been working and coding on SGI systems during the 90s while being a CompSci student. During that time I bought my first (refurbished) workstation in 1996. It was an R4k Elan Indigo and ..err.. kind of expensive. But having it at home was mind blowing. It wasn't the only system I bought throughout the years.

Anyways, after some 15 years of keeping them in the basement I started dusting them of and checking on them. So here's my problem. I started with an Octane running IRIX 6.5.22. I booted it via tftp to reset the root-pw. The tftp-server was a raspi and it worked out perfectly.

Since I couldn't use my external SCSI CDROM I mounted CD (efs) images and exported them via nfs on Linux (ubuntu) from said raspi. That way I wanted to install Performer and Octane Demos.

I feel like I'm missing something basic, something I forgot but starting "inst" selecting packages and starting installation was blazingly fast for a minute and 30% of the CD. Then it would slow down and take several hours(!!) to complete.

The NFS server is idle and the export is available on other (linux/mac) computers as fast as expected.

On IRIX it's almost blocked. Before starting an installation, I could access the mountpoint from different terminals and it responded fast, but once it stated "shutting down" it wasn't responsive on any terminal anymore.

So, wtf? What am I missing? It's NFSv3 UDP. I'm asking, because I hope I'm not the first one having that problem.

thanks and cheers,

elmy


RE: Unexpected NFS performance issues - Raion - 06-06-2020

IRIX NFS has a bug where if it is interrupted or if there's network congestion it can hang indefinitely until the connection resumes OR require both the host and client to reboot. This was never patched.

Others will disagree, but I've never had any luck using GNU/Linux as an NFS host, there are bugs in the Linux RPCBIND implementation and other problems that I observed. I usually recommend a BSD or illumos or Solaris. It also doesn't like mixed NFSv4 and v3 IME. Basically, you have to run pure v3 across your network, make sure you're using a good OS that doesn't get saturated. And your connection has to be reliable


RE: Unexpected NFS performance issues - elmy - 06-06-2020

Thanks, that sure was a valuable hint! Back in the days i was using a setup with fixed ip addresses and an internal dns server. Now, I just plugged the Octane into the DCHP network and it keeps complaining about not receiving answers until the lease almost times out. I'll start with setting up a static subnet again.

Is there any preference of TCP over UDP on IRIX? You mention the bug kicks in if a connection is interrupted, wich couldn't occur with UDP. netstat didn't show any tcp connections.


RE: Unexpected NFS performance issues - Raion - 06-06-2020

I'm not sure if the bug has any bearing on UDP or TCP, but I'd think IRIX would prefer TCP?


RE: Unexpected NFS performance issues - dexter1 - 06-11-2020

Are you sure you have NFS v3 installed on the IRIX machines? This is a separate package for IRIX.

Mounting NFS with the 'hard' option will timeout NFS queries until the connection is re-established. This prevents possible data corruption. If however there are frequent dropouts in network connectivity you might want to try 'soft' mounting the NFS partition.

You can force TCP protocol on the IRIX side with 'proto=tcp'


RE: Unexpected NFS performance issues - jan-jaap - 06-11-2020

(06-11-2020, 08:49 AM)dexter1 Wrote:  Are you sure you have NFS v3 installed on the IRIX machines? This is a separate package for IRIX.

That's the NFSv3 server, but in this case the server is the RasPI.

I'm not sure NFS is used at all during installation, isn't it all RCP? There's also a tweak necessary on the Linux side to limit ports used to 32K if I'm not mistaken


RE: Unexpected NFS performance issues - elmy - 06-11-2020

True, NFS would not have been involved if I had installed from PROM, but instead i just ran "inst" in a terminal. Will consider using PROM, if I can't work it out, though.

I did some tests (iperf, mostly) and it seems that UDP packet loss could be an issue, so I'll definitely try the "proto=tcp" option. TCP has been made default with 6.5.24 anyways.

I did not find anything related to ports but there was an issue up to 6.5.12 where file handles had to be 32 bytes, maybe that?