[Beowulf] raw ethernet

[Robert G. Brown]

On Wed, 21 Jul 2004 050675 at student.unife.it wrote:

> Now I'll try to find a stable situation wich lets me have a good
> throughput with very little frame-loss (so interrupt coalescence, NAPI,
> e1000 params tuning and so on)
> I don't know if someone else is making my studies but I would like to
> share ideas (in internet I didn't found many news about raw ethernet, so I
> think that is a quite new mechanism, interesting to develop)

Not so new, as I think several other responses pointed out, but still
interesting and an excellent student project to learn about
link-through-transport-layer networking.  I think there may be room for
some fully GPL drivers for various parallel libraries, as well as most
of the GPL stuff I'm familiar with tends to run on top of IP (either UDP
or TCP).

The challenge will be to significantly reduce latency as measured by
e.g. netpipe in whatever environment you choose.  This won't necessarily
be easy, and you'll need to learn a LOT about network programming as
you'll probably have to develop a library that in many ways emulates the
standard socket libraries and transport layer, leeching off of arp
tables somehow to achieve a semi-portable mapping between hostname and
ethernet number on the flat network.  As was suggested, look at the open
source drivers that already exist for this where you can find them, and
then proceed with your own ideas.

   rgb

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu