[Beowulf] Vector coprocessors

[Jeffrey B. Layton]

Karen Shaeffer wrote:

>On Thu, Mar 16, 2006 at 09:04:32AM +0100, Daniel Pfenniger wrote:
>  
>
>>The possibly most interesting niche for the Clearspeed cards appears to me
>>accelerating proprietary applications like Matlab, Mathematica and 
>>particularly
>>Dan
>>    
>>
>
>I just talked with Matlab. They sell an extension to the platform that
>enables it to run on a 32 bit Linux cluster -- yes they support clustered
>computing. They claim 64 bit clustered support will be in the next major
>release.
>
>http://www.mathworks.com/products/distribtb/index.html?fp
>
>Now I don't know how well this clustering feature is architected, but they
>are selling it. Has anyone used it on 32 bit nodes?
>
>I have it in my NRE budget for this spring and would appreciate any
>information about it's performance. While I really want the 64 bit
>clustering capability, I can just upgrade when it is released.
>
>Thanks,
>Karen
>  
>
It's not MPI clustering I'm sorry to say. It's embarrassingly parallel
clustering (no nodal communication).

For MPI communication you might look at interactive supercomputing.
Their product, Star-P, provides a bridge between MATLAB and clusters.
It looks pretty cool (I've seen a demo).

As for Clearspeed and MATLAB, I'm sure they are talking. One of the
bottlenecks is getting a complete BLAS/LAPACK library that is tuned
for the hardware. Then you could use the Clearspeed card for computing.
As other people have mentioned, the other bottleneck is getting data to
and from the card.

Jeff