Abstract

In this paper, we present a threads-only implementation of MPI, called TOMPI, that allows efficient development of parallel programs on a workstation. The communication and context-switching overhead is reduced significantly compared to existing MPI implementations, by the use of threads and shared memory in place of UNIX processes and often sockets. Results demonstrate the scalability of TOMPI in comparison to two MPICH devices. We provide a C/C++ preprocessor to modify the semantics of global variables to appear as if each thread has its own address space. This makes TOMPI a true MPI implementation, that is, MPI programs can be run in TOMPI without modification. Keywords: threads, MPI, parallel programming. 1 Introduction The development phase of parallel programming, including testing, debugging, and performance tuning, is the most difficult and time-consuming. Supercomputing resources are usually unavailable for development, since they are expensive and often not required. With t...

Citations