Sorry if I'm asking another question now, but I didn't found anything on
this subject either:
>does it mean to add a node in a batch-scheduled environment?
I thought it was possible NOT to give the nodes in advance (is it?)
And how should I do that (I have to admit I only played with MPI with a
fixed list of nodes which I knew in advance, and just supposed it shouldn't
be difficult to let mpich or lam/mpi "detect the PC's with mpi installed on
it by itself ???).
Thanks.
On 8/9/05, Jeff Squyres <jsquyres_at_[hidden]> wrote:
>
> On Aug 8, 2005, at 6:31 PM, Jim Lasc wrote:
>
> > And for a node, is there allso a method which isn't LAM-specific?
> > Which can be used with all MPI-2 implementations?
>
> Unfortunately, every MPI-2 implementation's version of this is
> different. Since it's not included in the MPI-2 standard, this is
> quite legal. The problem is that there are a *wide* variety of ways to
> do this, most usually having to do with what kind of run-time
> environment you're running in (e.g., what exactly does it mean to add a
> node in a batch-scheduled environment? There are several possible
> answers...).
>
> > (can't you use allso OPEN_PORT, COMM_ACCEPT/CONNECT on the "new" node
> > to connect it with a node which "allready knows what to do" ?
> > Or are there any other methods?
>
> In general, only the MPI API is standardized. MPI does not specify
> even how parallel processes are started (nor anything about resource
> allocation, etc.) -- it only specifies what the API calls do once an
> MPI process is running. This is much like how the Fortran standard
> does not specify how Fortran programs are started (in fact, the MPI
> Forum got their inspiration from Fortran ;-) ).
>
> --
> {+} Jeff Squyres
> {+} jsquyres_at_[hidden]
> {+} http://www.lam-mpi.org/
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
|
