[Beowulf] Rackable / SGI

Eugen Leitl eugen at leitl.org
Sun Apr 5 11:49:22 PDT 2009

On Sat, Apr 04, 2009 at 12:19:02PM -0700, Greg Lindahl wrote:

> Exactly like shared-bus multiprocessors.
> The incremental method of solving this is what Opteron/Nehalem does.

None of these had to deal with hundreds or thousands of cores in a
single socket yet (arguably GPGPU is pushing the envelope here,
and I wonder how long they'll manage to avoid going embedded
memory). SMP (shared global memory) in general would fall flat on its face
well before kilocore, not even talking about mega or gigacore.

> The more radical method is what Origin/Altix did. It all comes down to

As long as it's shared-nothing message-passing which doesn't pretend
to create an illusion of shared global memory it's ok.

> how many pins you're willing to commit to memory and how few pins you
> can squeeze a memory bus down to; once you bump up against that limit,
> then you need an improvement in packaging, which is no more radical
> than what Opteron/Nehalem or Origin/Altix did. And, of course, all

I would call http://www.cc.gatech.edu/~loh/Papers/isca2008-3Ddram.pdf
not your run-of-the-mill approach. Not radical, no.

> this was invented before Opteron or Origin.
> The sky is not falling.

Of course it isn't. Though I would like to see how Linux handles a megacore 
or a gigacore, each with a MByte or a couple of embedded memory, on a 
high-dimensional signalling mesh.

Eugen* Leitl <a href="http://leitl.org">leitl</a> http://leitl.org
ICBM: 48.07100, 11.36820 http://www.ativel.com http://postbiota.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE

More information about the Beowulf mailing list