[Beowulf] K Computer built for speed, not use

Vincent Diepeveen diep at xs4all.nl
Wed Oct 10 08:00:59 PDT 2012


Linpack is not an interesting benchmark.

If you look at a tad older systems, total identical systems, you will  
notice that built in gigabit ethernet versus QDR infiniband,
for the same size cluster with thousands of cores, that it's only a  
factor 2 difference for linpack in speed.

That's not realistic. Nowadays 1 node can have major crunching power,  
so the bandwidth node to node is important
as well specialistic software that can use you manycore hardware.

If we speak about supercomputers of this size, you don't build  
supercomputers to run existing software, you should optimize
your software to take advantage of the hardware!

On Oct 10, 2012, at 4:49 PM, Ivan M wrote:

> http://www.sciencemag.org/content/338/6103/26.full?rss=1
>
> "Japan's K computer made headlines in June 2011 as the world's fastest
> supercomputer and again last November when it became the first
> computer to top 10 petaflops—or 10 quadrillion calculations per
> second—solving a benchmark mathematical problem. (...) And now, after
> a year of testing and software development, as the $1.4 billion K
> computer is put to work on real-world problems, some scientific users
> say it was too narrowly built for speed."
>
> Interesting claim. What kind of architecture structure would benefit
> Linpack and would hinder real-world applications?
>
> Cheers
> _________________________________________
> Ivan S.P. Marin, PhD
>
> Postdoctoral Associate
> Département de géologie et de génie géologique
> Pavillon Adrien-Pouliot, local 3744
> 1065. ave de la Médecine
> Université Laval
> Québec (Québec) Canada
> G1V 0A6
> 418-656-2131 poste 7246
>  ivan.silvestre-paganini-marin.1 at ulaval.ca
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin  
> Computing
> To change your subscription (digest mode or unsubscribe) visit  
> http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list