very high bandwidth, low latency manner?

[Patrick Geoffray]

Steffen Persvold wrote:

> True, but if one of your Myrinet switches breaks down you loose 64 nodes
> in a 256 node system (standard "CLOS" configuration). I don't know the
> MBTF for Myrinet switches, but I would expect it to be rather high
> (redundant power supplies ?).

The calculated MTBF of the switches is +50 years. Actually, if all 6 
fans go off, it will still work, then the switch will drop more and more 
packets, then the uC will shutdown the blades one by one if they reach 
the  critical temperature limit.
If there is a failure on a blade itself, it will affect only 8 ports.
If there is a failure in a crossbar on the backplane, the mapper will 
use a redondant route (as many redondant routes as crossbars, so a 
failure in each 8 crossbars on the backplane is required to loose all 
ports).

Chuck made a very nice talk at Cluster2001 about Clos topology. It 
presents thing very clearly, I like it a lot: 
http://www.cacr.caltech.edu/cluster2001/program/talks/seitz.pdf

Regards.

Patrick

----------------------------------------------------------
|   Patrick Geoffray, Ph.D.      patrick at myri.com
|   Myricom, Inc.                http://www.myri.com
|   Cell:  865-389-8852          685 Emory Valley Rd (B)
|   Phone: 865-425-0978          Oak Ridge, TN 37830
----------------------------------------------------------