MPL_Connect managing heterogeneous MPI applications interoperation and process control

Abstract

Presently, different vendors' MPI implementations cannot interoperate directly with each other. As a result, performance of distributed computing across different vendors' machines requires use of a single MPI implementation, such as MPICH. This solution may be sub-optimal since it can not utilize the vendors' own optimized MPI implementations. MPL_Connect, a software package currently under development at the University of Tennessee, provides the needed interoperability between different vendors' optimized MPI implementations. This project grew out of the PVMPI project that utilized PVM to provide inter-platform communication and process control, and was upgraded to use the new MetaComputing SNIPE system which has proven more flexible and less restrictive than PVM when operating upon certain MPPs. MPI_Connect provides two distrinct programming models to its users. The first is a single MPICOMM_WORLD model similar to that provided by the contempary PACX project. Where inter-communication is completely transparent to MPI applications thus requiring no source level modification of applications. The second is that of uniquely identified process groups that inter-communicate via MPI point-to-point calls. Both systems use the MPI profiling interface to maitain portability between MPI implementations. A unique feature of this system is its ability to allow MPI-2 dynamic process control and inter-operation between MPI implementations. Currently supported implementation include MPICH, LAM 6, IBM MPIF and SGI MPI.