Limitations and Assumptions in using Scalapack.

[Suhasrm]

Hi All,

I was wondering if I could get to know the limitations in the use of Scalapack on windows machine. I want to use Scalapack to generate the SVD over the network of windows machine (32 bit, Intel dual core).

It would be of great help if anyone could share assumptions and limitations they know.

Thanks,
Suhas M.

[Julie]

Dear Suhas,
we have recently benchmarked ScaLAPACK SVD on a Windows cluster compiled with Intel Fortran Compiler (VS Studio Solution) and MS MPI. We did not observe anything unusual. Results are correct, performance and scalability as well.
let me know if you need anything.
Best wishes, Julie.

[Suhasrm]

Hi Julie,

Thanks for the information.
We are tying to implement concept search by calculating SVD. The documents will be in millions.
Hence the matrix for which we will be calculating SVD will be say 1 million X 1 million.
Can you please tell me if that is feasible using scalapack?
Also I would like to know for what dimension of the matrix have you benchmarked the Scalapack.

Regards, Suhas M.

[Julie]

1,000,000-by-1,000,000 ...
Your matrix is dense or sparse?
If the matrix is sparse, I would highly advise you to check on what people from the sparse community are doing. If the matrix is dense, this means several TeraBytes of RAM. You sure you have the computer to put your matrix in. For information, the largest system that have been reportedly solved as of today is of size: 2,456,063. This was for the TOP500 list in 11/2007.
see: http://www.top500.org/system/performance/8968

[Danesh_D]

Suhas,

Theoretically ScaLAPACK will be able to perform any matrix with any size only if enough memory is available. The matrix with the dimensions that you mentioned is very unusual and you may need a cluster with more than 100 nodes where each node has 10 GB memory. However if can also depend on what data type you will use within your matrix e.g. complex::double needa 16 bytes for each value.

D.

[Suhasrm]

Hi Julie,

As seen from the examples for ScaLapack, my understanding is we can store the input matrix in a file/secondary memory.
Now when we create the process grid, does all the data get loaded in the memory of the initiator machine and then get's distributed to other processors/memory?
Can you just brief out some points as how and where the matrix gets loaded?

Thanks, Suhas M.