[Beowulf] New beowulf recommendations

[Mark Hahn]

> What is GAMMA like for stability these days? It's efficiency would be
> compelling enough for me to do some testing, at least, were I in this
> situation. Going this route would also make switch latency somewhat more
> relevant.

a lot more relevant.  I wonder whether the gamma idea could be redone with
the kernel's netpoll interface.  that's how kernel logging/dumping/debugging
happens - it's a device-independent path for UDP only, and is not intrusive
on the normal network stack.  I also wonder where the ~50 us latency actually
comes from, since I don't believe it really scales with cpu speed.  for 
instance, suppose you used UDP; could you make a faster MPI?  if UDP is a 
problem (kernel buffering perhaps), then how about using MMAP_RING (via
pcap), perhaps?