[Beowulf] 1 multicore machine cluster

[Geoff Jacobs]

Залетнев Дмитрий wrote:
> 
>> is it possible to have a single multicored machine as a cluster?
>>
>> -- 
>> Jonathan Aquilina
> 
> I have a 2-core  machine with a lot of memory and GLAN NIC
PC, yes?

> and two PS3 with 8 cores each with 256 MB of RAM and slow GLAN network, thanks to PS3's "hypervisor", AIX-based native PS3 system, on witch CentOS-based YDL works under Virtual Machine.
> I'd like that my 3 machines have common address space. In '95 there was a french project Phosphorus for PVM intended for this goal. 
> Is there any common-memory module for MPICH2?
Not that I'm aware, there isn't. I believe MPICH2 has yet to be extended
for heterogeneous clusters, so if the dual core machine is not PPC and
binary compatible to the PS3s, it'll have to be used as infrastructure
only. Further, the SPEs on the Cell are managed quite a bit differently
than a standard SMP core.

For this reason, UPC and Titanium (which would be good options for
symmetric nodes) won't work too well. However, programming the SPEs as a
single vector processing unit through one of many methods, then fronting
it with UPC would work. Or, you could shell out some cash for Rapidmind
and try your luck.

So, you have some decisions to make. Good luck!

> I'd like to convert my geterogeneous GLAN with several nodes into single common-memory supercomputer, because my CFD-applications demand a lot of memory. By the way, in june PGI releases compilers with support of GPU programming. They're going to standardize this as it was with OpenMP. And I have nVidia G80 card in my PC.
You can distribute the data store efficiently if you can distribute the
compute space efficiently. You might have a problem here, as PS3s don't
have a lot of RAM.

> Dmitry Zaletnev

FYI - Universal Parallel C, Co-Array Fortran and Titanium are three
language dialects (C, Fortran, and Java-ish) which are able to abstract
memory distribution as variable indices within the language.

-- 
Geoffrey D. Jacobs