Beowulf with Gigabit Ethernet

[Joachim Worringen]

> On Fri, 11 Jan 2002 alex at compusys.co.uk wrote:
> 
> > I think that in general point to point performance information has a
> > limited value, whatever vendors might quote on their web-page. SCI might
> > be performing pretty well if it comes down to just latency and bandwidth
> > between two machines, but it is a ring topology. If you have
> > more machines on a ring they will share that same bandwidth.

You might want to get a little bit more informed on SCI topologies and
their scalability characteristics. One good paper I can recommend is
from the SCI Europe 98 conference, available at
http://www.scali.com/whitepaper/scieurope98/scale_paper.pdf . It gives
some general calculations of all-to-all communication scalability with
SCI-torus-topologies. The numbers are somewhat outdated, but the
principle is still correct.

Your statement is not well founded because SCI is not a ring-topology,
but a point-to-point topology. It's similar to somebody saying
"(Centralized) Switches do not scale because all the traffic needs to to
through one box". It depends on the switch, I'd rather say. And on some
other things, like packet format, routing method etc.

Regarding the limited value of point-to-point performance, you are
right, but this applies to all networks.

  Joachim

-- 
|  _  RWTH|  Joachim Worringen
|_|_`_    |  Lehrstuhl fuer Betriebssysteme, RWTH Aachen
  | |_)(_`|  http://www.lfbs.rwth-aachen.de/~joachim
    |_)._)|  fon: ++49-241-80.27609 fax: ++49-241-80.22339