Authors: Joao V.F. Lima, Nicolas Maillard
Addresses: Institute of Informatics, Federal University of Rio Grande do Sul, CEP 91.501-970 Porto Alegre, RS, Brazil. ' Institute of Informatics, Federal University of Rio Grande do Sul, CEP 91.501-970 Porto Alegre, RS, Brazil
Abstract: In recent years, distributed platforms became largely used on HPC, and most of these architectures have different levels of parallelism. Hence, one of the key design stages in parallel programming is task mapping which attempts to maximise processor utilisation and minimise communication cost. However, this depends on a programming environment with efficient mapping scheme. This paper presents a library to MPI-2 (libSpawn) that implements a scheme to map tasks between processes and threads in order to minimise communications and task creation costs. We evaluated the libSpawn with two dynamic MPI programs: Fibonacci and Mergesort. Our experiments demonstrate that the mapping scheme offers significant performance improvements.
Keywords: dynamic programs; task mapping; load balancing; multithreaded programming; high performance systems; parallel programming; threads.
International Journal of High Performance Systems Architecture, 2009 Vol.2 No.2, pp.81 - 89
Received: 09 Jun 2009
Accepted: 31 Aug 2009
Published online: 04 Mar 2010 *