[Beowulf] Harpertown Numbers
deadline at eadline.org
Fri Nov 9 15:40:05 PST 2007
> -------------- Original message --------------
> From: "Douglas Eadline" <deadline at eadline.org>
>> As many of you know Intel is releasing the Harpertown
>> next week at SC07. I will have a white paper with my
>> requisite NAS parallel MPI results. Appro commission the
>> paper so I will be hanging around their booth next week
>> handing out the paper. (book geek that I am).
>> I cannot mention any numbers until next week
>> and the white paper will eventually show up on
>> Appro's website, but I can say it worth taking a
>> look at the numbers :)
> Cool Doug!
> So are your numbers consistent with:
> A 2 socket, 8 core Linpack number of about ~77 GFLOPs
> Stream Triad for the same 2 socket board of ~8.6 GBytes/sec.
Well mostly I look at MPI scalability. I'm not two concerned
with Linpack as I am sure it will run well. I like to
beat on the NAS suite (no dead horse were harmed in these
> Are the numbers you are reporting from Appro Dual Socket Blades? I did
> not think they had announced Harpertown Blades yet. As I recall the
> Tri-Labs win was for Dual Socket Barcelona's and there Intel blades
> solution was Clovertown. Can you clarify without violating an embargo
The hardware was two - 2 socket 1U servers, connected with IB
(16 cores total) with which I tested the Clovertown last year.
So it makes for a good comparison (pop in new processors)
I also used a newer kernel (2.6.22) and gcc/gfortran 4.1.2
(last year I used gcc/g77 3.4.5 and kernel 2.6.9) I can't
talk about Appro product launches because
I don't really know about them. I can say last year I was
less than enthused about Clovertown. This year, I think the
numbers are well worth looking at. :-)
> ... ;-) ...
> "Making predictions is hard, especially about the future."
> Niels Bohr
> Richard Walsh
> Thrashing River Consulting--
> 5605 Alameda St.
> Shoreview, MN 55126
> Phone #: 612-382-4620
More information about the Beowulf