networking options

Patrick Geoffray patrick at
Fri Sep 27 05:55:04 PDT 2002

Hi Philippe,

On Wed, 2002-09-25 at 05:11, Philippe Blaise - GRENOBLE wrote:
> -> You seem to forget the Cray T3E experience. SCI network solution is
> very closed, isn't it ?

I don't forget it at all.
Many people liked the T3E much more because its software (parallel
execution environment, compiler, tools, etc) than its hardware. The
exception in software was its MPI implementation, performing very poorly
compared to its SHMEM interface. 
It was also a big success because there was not a lot of good
alternative solutions at that time and Cray was doing a good job
packaging the whole machine.

With disappointing MPI performance and slow clock CPUs compared to
today's processors, asserting from the T3E the effectiveness of a
Torus-based network for a large machine today is a risky shortcut.

The link bandwidth needs to be much more that the I/O bandwidth to be
comfortable in a Torus (shared links). It will certainly happen in the
future, but as the PCI bandwidth is increasing quickly (500 MB/s PCI, 1
GB/s PCI-X, 2-4 GB/s PCI Express), it will be technologically hard (and
expensive) to increase the link bandwidth X times as fast, IMHO.

|   Patrick Geoffray, Ph.D.      patrick at 
|   Myricom, Inc.      
|   Cell:  865-389-8852          685 Emory Valley Rd (B)
|   Phone: 626-821-5555          Oak Ridge, TN 37830

More information about the Beowulf mailing list