Newsgroups: comp.parallel.mpi
From: pvb1@Ra.MsState.Edu (Purushotham V. Bangalore)
Subject: Re: MPI support of host nodes
Organization: Mississippi State University
Date: 10 Nov 1994 08:46:30 -0600
Message-ID: <pvb1.784477615@Isis.MsState.Edu>

jeff@is.s.u-tokyo.ac.jp (Jeff McAffer) writes:


>I've noticed a bit of a problem with the MPI topology strategy.  On
>some of the multicomputers I use there is the notion of a "front-end"
>or host computer.  Typically a regular Sun with some special interface
>cards etc which sits between the node array and the rest of the world.
>It would seem to be difficult to describe this host machine as a node
>in an MPI topology.  Well, in a Cartesian topology.  

>I suppose in principle a Graph topology is sufficient but it is sort
>of frustrating that Cartesian provides everything you want except the
>ability to have this one extra node.  

>Am I overlooking something?  Is this really not that
>important/widespread?  Is there a simple solution (other than manually
>constructing a Cartesian out of a Graph)?

>Comments?

>JEff

I feel this is a good situation where you can use intercommunicators.
The standard does not support collective communications between 
intercommunicators right now, but you can find more information about this 
in the paper "Inter-communicator Extensions to MPI in the MPIX (MPI eXtension) 
Library" by Anthony Skjellum and Nathan Doss and Kishore Viswanathan.

ftp://aurora.cs.msstate.edu/pub/reports/Message-Passing/mpix-icae.ps.Z

----------------------------------------------------------------------------
Puri                                puri@cs.msstate.edu

