[MPICH] Re: Dynamic processes with MPI-2
William Gropp
gropp at mcs.anl.gov
Mon Sep 12 13:23:40 CDT 2005
At 07:19 AM 9/12/2005, Alok G Singh wrote:
>I have some questions that probably sound rather silly but I have spent
>about a week and a half reading and searching online for some documents on
>this subject. My starting point was
>http://www.mpi-forum.org/archives/mail/mpi-dynamic/0001.html
You should refer to the MPI standard which is available at
www.mpi-forum.org/docs; in particular, see
http://www.mpi-forum.org/docs/mpi-20-html/node88.htm#Node88 .
>However, when I searched the MPI-2 standard I could not find the functions
>described in that chapter. I have ordered the book 'Using MPI-2' but it
>will take a couple of months to trickle down to me :( Is the dynamic
>process model described in that chapter not the way it was finally
>implemented ? I tried writing a test program to check the behaviour of
>MPI_Comm_spawn() but it just seems to add more processes to the ring
>already established. Is there someway to add more nodes to the ring and
>spawn processes onto them ?
The MPI Forum decided not to specify how an MPI program controls the pool
of available machines (e.g., adding or deleting machines from a process
manager). An MPI implementation may use the MPI_Info parameter to pass
information to a process manager through the MPI_Comm_spawn command, for
example, but there is nothing like an "MPI_Add_host" command.
The Forum struggled with this. The problem was that there was no consensus
on what the interface for something like this should look like, so in the
end, we couldn't make it part of the standard.
Bill
>The 'preliminary chapter' makes perfect sense to me. I think the resource
>manager component could be implemented using the Globus Toolkit. Has
>anyone worked on it from this perspective ?
>
>Any help would be most welcome. I would be doubly grateful to any current
>online resources.
>
>TIA,
>Alok
>
>---
>Linux, the way to get rid of boot viruses
> -- MaDsen Wikholm, mwikholm at at8.abo.fi
William Gropp
http://www.mcs.anl.gov/~gropp
More information about the mpich-discuss
mailing list