RE: [pooma-dev] Some POOMA questions

[James Crotinger]

> Well - its easy to implement them, but probably a lot harder to get them
> working at acceptable speed. Also doing f.i. vector x matrix with Array<1>
> and Array<2> is not supported by any data-parallel operation I know of.

Exactly - unless the matrix is sparse, the last thing you generally want to do on massively parallel machine is a matrix-vector product.

>> 3. Is there any work being done to bring SMARTS up to speed? For
>> parallell execution I'd like to try running on a MOSIX cluster, using
>> SMARTS to do the threading, but the current SMARTS in CVS fails to
>> compile the same way it did when I tried it briefly last year (there's
>> some conflict with pthreads)

> Still the same - I have some work on OpenMP support in my queue (but need
> to finalize reduction support) that will perhaps make SMARTS unnecessary.
> I myself use cheetah for MPI parallelization on a workstation cluster.

The real advantage of SMARTS was its data-flow ability, which is not something OpenMP will recover (unless OpenMP has some magic that it didn't have when I last looked). That said, SMARTS is pretty complex and its light-weight threads are very machine-dependent -- it was still pre-beta software on anything except the SGIs when the development ended. I think it would make a great research project, but I wouldn't push it as part of production POOMA.

        Jim