HP create off-the-shelf supercomputer

Patrick Geoffray patrick at myri.com
Fri Oct 5 08:36:40 PDT 2001


Folks,

I would like to react to this serie of press releases.


> How to build your own supercomputer: Take a few off-the-shelf,
> stripped-down PCs, add some network switches, a maze of Ethernet cabling
> and some homegrown Linux software, and you'll be well on your way.

This is exactely the type of sentence that shameless companies put 
in their press releases. "Supercomputer" is the magic word that will 
attract the focus of the journalists and the political people.
What is a "Supercomputer" ? A machine that will enter in the 
Top500 ? 
Everybody with a little bit of common sense will realize that the 
Top500 is not the list of the most powerfull supercomputers in the 
world. It's the list of machines where the Linpack benchmark 
(very well implemented in HPL) produces the biggest numbers. That's 
another story and the folks at INRIA know better than that.

> Hewlett-Packard, together with a national laboratory in France, tried this
> recipe out and, to the great surprise of many scientists, it worked. What

Which great surprise ?!? Some people didn't know that Linpack is 
scaling almost linearly for this size of machine ? This is CPU 
bounded.
 
> At a technical session last summer, scientists from HP's labs in Grenoble,
> France, started talking to experts at the local INRIA Rhone-Alps (France's
> National Institute for Research in Computer Science) about the possibility
> of doing "something a little unusual." The idea was to build a
> supercomputer out of standard hardware components like those that might be
> found in the typical big business.

Or... [HP called INRIA and told them "Hey, we have 225 machines, 
we don't know what to do with them. Here, it's for you, do something 
with that and make a lot of PRs"]. No ? Ok.

> Other clusters, like ASCI Red at Sandia National Laboratories in New
> Mexico, are made up of heavily modified parts.

HP cluster at $210K versus ASCI Red. These Natl Labs are so 
stupid: why spend so much money on these ASCI boxes, when 
they can buy "mainstream" elements ?
They spend a lot (A LOT) of money, but it's not only to run 
the Linpack benchmark and have a good rank in the Top500 
(actually, it's part of the game but it's the cherry on the cake).
Try to run a sparse matrix computation on this HP cluster, or just 
a dense matrix computation with real matrices from disk and include 
the time to read these matrices, distribute them and collect the 
solution to disk.

> "You could gather the latent power from office PCs using this technique,"
> he said. "We eventually want to scale it higher, to thousands of PCs."

Do they mean "We eventually want to invent the Grid computing". 
I am afraid they will again be the first in the world to do it.


I know that PRs have to be nice, they are not always written by 
technical people and they are not written for technical people. 
But what is the message here ? 
If it's "just steal your secretaries' PCs, put them in the 
same room, add this nice software and hop, you have a parallel 
machine", that's a point of view.
If it's "Just steal your secretaries' PCs, put them in the 
same room, add this nice software and hop, you have a machine 
similar to the 499 others in the Top500", that's wrong.
There is not only dense matrix computations in the supercomputing 
world.

Feedback welcome.

Regards.

Patrick




More information about the Beowulf mailing list