Re: Dynamic Process comment: using MPI_Init() without mpirun

David Sitsky (sits@cafe.anu.edu.au)
Tue, 30 Apr 1996 14:39:04 +1000

> > With such a function we could dispense with mpirun and friends
> > altogether, and any application could just use this function to
> > "become" the first process, and then spawn all the rest of the
> > processes.
>
> But what if your execution model is to start all of your processes
> at once, and you are in fact not even allowed to spawn new ones?
> (I'm thinking of some of the MPPs.) It's not clear that all
> architectures will be able to support the dynamic process
> chapter. At the moment, these machines are capable of supporting the
> semantics of MPI_Init() as it currently stands. This would break
> that, I think.
>
> Would some of the MPP vendors care to comment?

We have both situations on our Fujitsu AP1000 and AP1000+ machines
here.

On the AP1000, the CellOS operating system only supports the model
of starting all your processes at once.

On the AP1000+, we have Linux running on each node, which allows us
to have the spawning method in MPI_Init().

Cheers,

----------------------------------------------------------------------------

David Sitsky Email: sits@cs.anu.edu.au
CAP Research Program Phone: +61 6 249 5143
Australian National University Fax: +61 6 249 0010
Canberra ACT 0200, Australia WWW: http://cap.anu.edu.au/~sits

----------------------------------------------------------------------------