If I understand what you're asking -- each process is essentially a
standalone entity and they periodically contact each otherr -- then,
yes. The MPI-2 dynamic functions can be used for this purpose (e.g.,
MPI_COMM_CONNECT and MPI_COMM_ACCEPT) -- you might want to have a look
through the "Process creation and management" chapter in the MPI-2 spec
for more details.
On Nov 24, 2004, at 9:05 AM, Tomek wrote:
> I am rather new to MPI - so probably my problem will
> appear to be rather basic or simply imposible to be solved:
>
> I am writting a library that will have an mpi version.
> The finall aplication will use several such libraries and I would
> like to avoid any run schemas that would depend on sequence
> of runs of workers for different libraries. Actually I would like
> all the workers work like 'deamons' on cluster nodes and
> respond somehow for the request from master application
> - when separate libraries are initialized and create communicator
> objects for their workers. So I would like to have workers running
> permanently and only "checking in" to the communicator
> when asked (they could fork to take advantage of smp nodes).
> Is it possible to be done with MPI ?
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
