Slashdot Mirror

It is unfortunate that all the ones large enough to head the list are basically standard racks in rows with attractive door art.

Mare Nostrum is comparatively small; but the 'glass pod full of ebon supercomputer modules seemingly suspended in a historic Spanish chapel' effect is pretty neat...

I definitely find any claims of it being especially high performance deeply fishy, and the unveiling of it reeks of a PR stunt; but nothing about the (limited) hardware that is there looks definitely fake.

The hardware is legit, even run-of-the-mill. My first thought was that the billing system I manage (small mobile telco, 6MM users) is a lot bigger than this, except perhaps in storage. Yes, it looks halfway-cabled too, and cooling is underwhelming. That's why I wouldn't be the least bit surprised if this was a politically-motivated show.

"cluster porn" shots of the Big Serious Core Switch

I had to check out that site, just to see what kind of iron moves that much data around. An odd place to build one, BTW. Thanks!

I definitely find any claims of it being especially high performance deeply fishy, and the unveiling of it reeks of a PR stunt; but nothing about the (limited) hardware that is there looks definitely fake.

A bunch of 1-3Us, connected via rather scraggly patches from a few GigE switches to the on-motherboard GigE ports. Depending on the specs of the servers, we could be talking a few hundred cores, and a reasonable number of TB of storage...

It certainly is of abnormally low density, though. Each rack is 50% or more blanking plates, and I don't see any of the classic "cluster porn" shots of the Big Serious Core Switch, which they would definitely show if they could. No "Stylish rack opened enough to show all 72Us of servers exuding a sense of overwhelming power" shot or "View down the rack hallway that appears to go on forever, puny meatsack" shot either.

I'd totally believe that this is a (somewhat optimistic/growth prepared) conversion of some underused university basement to support a CS department curriculum and/or the compute needs of other disciplines; but it sure isn't what the PR says it is....

CBEs or successors will probably get popular in HPC.

The Barcelona Supercomuting Center, for instance, has a site on Linux on Cell BE-based Systems: http://www.bsc.es/plantillaG.php?cat_id=96

Now your problem is to parallelize the linear system solver. Normally this task takes up to 90% of total execution time so it's a good candidate for running on SPEs. For the other 10%, leave it on the PPE. And don't forget : adding "fun" stuff increases the code size, which means less space for data on SPE.

Resources are available on IBM's developerWorks site (http://www.ibm.com/developerworks/power/cell/docs_articles.html, see also the forums some interesting issues are discussed) and on Barcelona supercomputing center (http://www.bsc.es/plantillaH.php?cat_id=326).

http://www.bsc.es/plantillaA.php?cat_id=200

http://en.wikipedia.org/wiki/Image:MareNostrumReal.jpg

good luck. :D

>Are Intel processors "super computers" now or something?
No, processors ARE NOT supercomputers (actually, the are not computers at all). But if you put enough of them together in the appropriate way, they BECOME a super computer.
Super computers are no longer made from special purpose hardware. Now it makes much more economical sense to build them from general purpose hardware like those Intel or Power PC processors. Look at the Marenostrum, a super computer here is Spain.

I wonder how much faster the Intel versions will be in comparison to the G5s...

Look at #20; it's almost identical to an Intel Xserve.

If anybody's curious, here's Barcelona Supercomputing Center's brief description of their system: MareNostrum System Architecture

MareNostrum uses 2560 IBM BladeCenter JS21 blades.

It takes quite a lot of porting effort to make an application fast on Cell, not just recompiling it. See http://www.bsc.es/projects/deepcomputing/linuxonce ll/ for info on a simulator and binaries you need to get started.

If you just want to run generic code, use Fedora Core 5 for powerpc, which incidentally was also released today and happens to run on the same machines that the kernel currently supports.

Also see the Barcelona Supercomputing Centre's Linux-On-Cell project.

With the new, incompatible architecture, porting existing projects to the 360 will likely take quite a bit more effort. I understand your concerns, however IBM has opened the source to some Cell libraries. IBM has also a site devoted to the cell processor that includes kernel patches for the cell processor. We also don't know enough about the security on the 360--it may be that mods will have to be specific to each unit, making said mods more costly and possibly more dangerous (easier to screw up the installation). We also didn't know about the security of the original xbox, however look how far we were able to get with that. :)

Anyway, HP can complain as much as they want. I find that surprising. The Mare Nostrum is a result of many years of colaboration between the UPC and IBM in the form of the CIRI (CEPBA-IBM Research Institute), a center where research on parallel applications and tools is conducted. IBM had the intention to build a large supercomputer out of PowerPCs and blades and they were considering even places outside of the US. People at UPC and particularly the CIRI and the Computer Architecture department started talking with the government to bring the machine here, and along with it, start a new research institute on supercomputing (the BSC, btw my new employer). well that's at least what I've heard here at upc.

So HP would have wanted to compete??? As far as I know, all research that HP does in barcelona is dedicated to printers ...worse, it could be 4500 itaniums, can you imagine? ;-)