[openib-general] ip over ib throughtput

[Grant Grundler]

On Wed, Jan 05, 2005 at 09:23:29AM -0800, Peter Buckingham wrote:
> stupid question: why are we limited to a 2K MTU for IPoIB?

Reminder: The MTU size is not as important for "server" type
applications if the NIC supports "Large Send Offload" (LSO).
Reducing the number of interrupts to one per send instead
of one per MTU is a big win (along with TX Cksum offloading).

We just happen to be using a benchmark that doesn't support that.

ia64 uses 16K pages by default and many folks would like
to bump that up the 64K. So a 4K MTU is less than optimal
but not a disaster if we can do something like LSO.

grant