[geeks] NFS for the last time (I hope)

[Joshua D Boyd]

I have now pinned down the cause of my NFS trouble.  It has something do to
with the packet fragments.  Specifically, if I leave it at default (4096 byte
packets sizes on both read and writes), I seem lots of "icmp: ip reassembly 
time exceeded" errors (when running tcpdump), and lots of packet fragment 
messages.  Further, I can always cause the NFS server to freeze up by doing 
heavy writing (by freeze up, all programs trying to access NFS shares are 
frozen).  Heavy reading does nothing.  The problem does not occur if all the 
test systems are on a hub (although I still see the fragment messages, but
not the icmp reassembly errors), only when they are on the lanplex.

Now, if on the client systems, I set the read and write packet sizes to 1500,
things work perfectly.  Since only writes freeze things up, I want to try
setting the packet size on writes to 1500, but leaving the reads at the default
4096.  But before I can test that, I have to wait for the frozen processes to 
come back so that I can unmount the NFS share without having to reboot.

Anyway, now the question is: what is the real cause?  Is the lanplexes buffers
just being overrun by fragments?

Second, is just setting all clients to use a write (and perhaps read) packet
size of 1500 the proper way to fix this, or what?  Until someone suggests 
something better, this is what I plan to do.

BTW, I do see that my Lanplex has simm sockets.  I wonder if beefing up the 
memory would help, assuming that the simms really are for ram.  I haven't yet
figured out exactly what they are for, but I'd assume they are covered 
somewhere in the lanplex documentation online.


-- 
Joshua D. Boyd