Ralph Campbell wrote:
The reason a larger network MTU helps is because the Linux network stack is
more efficient when using larger MTUs.
If all your traffic goes within the IB cluster and only between Linux
nodes, the connected-mode / 64k mtu would serve you well.
If you use TCP and have to communicate with IB nodes not supporting the
connected mode or through IB/Eth gateway, you may want to use an HCA
which supports TCP stateless offloads such as LSO (Large Send Offload)
and TCP checksum offload. For example with LSO, even though the nic mtu
is 2k, the stack would let the nic to send a 64k tcp segment. As the
name suggests, "stateless offload" are inter-operable, that is there no
requirement on the other side.
Or.
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
