[Beowulf] NFS Read Errors

[Mark Hahn]

>> > Fourth, with ttcp over tcp, I found that the troubled machine could
>> > send 800 MB in about 20 seconds -- the wire speed for those 32-bit PCI
>> > slots as tested by netpipe.  However, if I used ttcp over udp, I
>> > couldn't reliably send even ten 8192-byte blocks!  Successive sends

it's worth noting that ttcp with udp means that you're producing 8k 
datagrams, which means you're actually sending 6x 1500B packets back-to-back.
that's somewhat hard on the transmitter, on the receiver, and on the 
receiver's net stack, since it as to reassemble the original 8k datagram.
(there have been interesting cases where fragments get reordered for 
various reasons, and this causes real problems for the stack, since there's
often a fixed window for how many fragments will be remembered to be 
reordered...)

anyway, tcp is smart enough to send a stream of path-mtu packets,
so to some degree avoids this.

>> This sounds like it could be a flow control issue. TCP does its own flow
>> control, but UDP doesn't. Some Ethernet switches do IEEE 802.3x, which
>> might help your UDP performance.

otoh, nfs, even over udp, does its own flow-control.  it might not be very 
good, or might even not be working right here, though.