NFS over GB ethernet

Bogdan Costescu bogdan.costescu at iwr.uni-heidelberg.de
Wed Nov 6 13:20:48 PST 2002


On Wed, 6 Nov 2002, Art Edwards wrote:

> My prejudice at this point is that the init scripts on each node are
> ordered correctly and that the card does come up during boot. I'm using
> a sysco catalyst 4000 switch and I believe it is the switch that has
> latency in bringing up a link. At boot time I typically get a message
> that the link is down, and then sometime later, the link is up. 

Yes, it's the switch. It probably has Spanning Tree Protocol turned on.  
Either turn it off or turn on something like "fast port restart" (I don't
have administrative access to such a switch, but all our desktop machines
are connected to one). When the NIC comes up or autonegotiation of the
link speed is restarted, the port is only partly active (I haven't found a
good definition of what partly means) - it takes 1 minute until the port
is fully activated. Some time ago Donald modified the start-up sequence of
the 3c59x driver to not restart the autonegotiation when teh driver is
initialized and that makes our desktop machines with 3Com cards being
almost ready to use the network after the driver is loaded, but the
machines equipped with tulip-driven NICs are not as lucky... The problem
with NFS mounts is that they also have a 1 minute timeout by default and
they sometimes work and sometimes don't. A quick fix is to just insert a
"sleep 15" (or 60 if you want to be on the safe side) line in the network
start-up script after the point where the interfaces are brought up and
put all other network related scripts after the NFS client one (NFS blocks
until the mount succeeds or times out).

--
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu at IWR.Uni-Heidelberg.De




More information about the Beowulf mailing list