[Beowulf] NFSv3 client hangs - tcp v/s udp.
rmillner at fiendly.org
Mon May 8 11:19:05 PDT 2006
Were there any console messages on your NFS server?
A problem I ran into last year on a 384 node cluster was that Linux has
a hard coded limit of 20 TCP connections per nfsd. This is a bare
constant buried in the kernel server code.
The console message on the server was:
nfsd: too many open TCP sockets, consider increasing the number of nfsd
On any specific nfsd, connection #21 would cause connection #1 to drop.
The client from #1 would then make the new #21 causing someone else to
If I recall properly (and may not so check this yourself) any specific
mount resulted in a TCP connection. Size the number of nfsd processes
to be around (nodes * mounts)/20. I haven't really had problems with
UDP on a low latency, well behaved network.
Anyway, not sure if this is germane to your actual problem but its one
to be aware of.
More information about the Beowulf