Paper showing Linpack scalability of mainstream clusters
RICHARD,BRUNO (HP-France,ex1)
bruno_richard at hp.com
Fri Sep 28 00:00:16 PDT 2001
Hi Roger,
If you read our paper, you will see that one of our goals is to model an
enterprise network. Which typically will not have SMP machines but standard
PCs.
If we were targeting pure performance/price, we would have rather chosen SMP
machines. An experiment is ongoing at INRIA, with a 200 nodes cluster, SMP.
This will have a much better performance.
Here the goal is to use *existing* PCs, hence acquisition cost is not
considered (so infinite performance/price ratio;)
Regards, -bruno
-----Original Message-----
From: Roger L. Smith [mailto:roger at ERC.MsState.Edu]
Sent: Thursday, September 27, 2001 20:38
To: RICHARD,BRUNO (HP-France,ex1)
Cc: 'Scott Shealy'; beowulf at beowulf.org
Subject: RE: Paper showing Linpack scalability of mainstream clusters
Bruno,
Hmmm, I thought that some of the others were purely ethernet based, but
after doing some quick research, I guess I'll stand corrected.
If you intend "mainstream" to mean only single-processor desktop-type
machines, then I'll completely concede your point.
As of the last Top 500 list, our cluster was listed as 158th, and is
entirely ethernet based, using a single 100Mb/s interconnect to each node,
and GigE interconnects between switches. However, our nodes are 1U with
dual processors, so I guess maybe it doesn't fit the definition of
mainstream that you stated. It doesn't predate the I-cluster, but it does
at least tie with it. :-)
-Roger
On Thu, 27 Sep 2001, RICHARD,BRUNO (HP-France,ex1) wrote:
> Hi Roger,
>
> What we mean by "mainstream" is that these could be your grandma's
> machine, unmodified except for the Software. Actually I-Cluster *is*
> the first cluster of this type to enter the TOP500. Some PC-based
> clusters are already registered there of course, but they cannot be
> called "mainstream": Most require specific (non-mainstream at
> all!) connectivity such as Myrinet, SCI, Quadrix... Some are based on
> PCs equipped with several LAN boards (not mainstream either). If you
> restrict to off-the-shelf monoprocessor (excluding non-mainstream
> Alpha, MIPS and such) interconnected through standard Ether100, no
> cluster ever entered the TOP500 list. And I-Cluster is still the only
> one to be there. Let me know if you think otherwise.
>
> Regards, -bruno
> _____________________________________________
> Bruno RICHARD - Research Program Manager
> HP Laboratories
> 38053 Grenoble Cedex 9 - FRANCE
> Phone: +33 (4) 76 14 15 38
> bruno_richard at hp.com
>
>
> -----Original Message-----
> From: Roger L. Smith [mailto:roger at ERC.MsState.Edu]
> Sent: Thursday, September 27, 2001 15:33
> To: RICHARD,BRUNO (HP-France,ex1)
> Cc: 'Scott Shealy'; beowulf at beowulf.org
> Subject: RE: Paper showing Linpack scalability of mainstream clusters
>
>
> On Tue, 18 Sep 2001, RICHARD,BRUNO (HP-France,ex1) wrote:
>
> > Sorry Scott, I sent you a wrong reference. The actual link is
> > http://www.hpl.hp.com/techreports/2001/HPL-2001-206.html. Enjoy,
> > -bruno
>
>
> I'd be REALLY interested in hearing how you justify the following
> statement in the paper:
>
> "Being the first ones to enter the TOP500 using only mainstream
> hardware (standard PCs, standard Ethernet connectivity)...".
>
>
> _\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_
> | Roger L. Smith Phone: 662-325-3625
|
> | Systems Administrator FAX: 662-325-7692
|
> | roger at ERC.MsState.Edu http://WWW.ERC.MsState.Edu/~roger
|
> | Mississippi State University
|
> |_______________________Engineering Research
> |Center_______________________|
>
_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_\|/_
| Roger L. Smith Phone: 662-325-3625 |
| Systems Administrator FAX: 662-325-7692 |
| roger at ERC.MsState.Edu http://WWW.ERC.MsState.Edu/~roger |
| Mississippi State University |
|_______________________Engineering Research
|Center_______________________|
More information about the Beowulf
mailing list