New Trends in Parallel Electromagnetic Fields Computation B. Butrylo (1) , F. Musy (2) , L. Nicolas (3) , R. Scorretti (3) , C. Vollaire (3) (1) Bialystok Technical University ul. Wiejska 45A, PL 15-893 Bialystok, Poland bogb@cksr.ac.bialystok.pl (2) MAPLY, UMR CNRS 5005, Ecole Centrale de Lyon, 69134 Ecully cedex, France Francois.Musy@ec-lyon.fr (3) CEGELY, UMR CNRS 5585, Ecole Centrale de Lyon, 69134 Ecully cedex, France Laurent.Nicolas@ec-lyon.fr Abstract The new trends of parallel electromagnetic fields computation are presented. New hardware developments are given. New developments in parallel methods are described: standard iterative and direct solving methods and domain decomposition methods. Special attention is paid to parallel computation using the JAVA language. The current status and properties of two prevailing programming environments (PVM and MPI) are finally given and compared. 1. Introduction Numerical computation is more and more used in engineering sciences to develop new device or to optimize existing one. Only parallel computers provide the increase in computing performances necessary to solve today’s problems. Two reasons may be highlighted: large memory is required because of a large amount of data, or speed is required to obtain the solution [1]. We have presented previously a survey of the parallelization of numerical techniques used in computational electromagnetics [2]. The objective of this paper is to present the new tendency of parallelized computational electromagnetics. The first section deals with new hardware developments. In further sections new trends in parallel methods are described: standard solving methods and domain decomposition methods. Special attention is paid to parallel computation using the JAVA language. The last section of this paper shows a comparison between the two message passing libraries PVM and MPI. 2. New Trends in Parallel Hardware A few years ago, most of the supercomputers were massively parallel computers type. Schematically, such a computer is composed of independent subsystems. Each processor has its own memory, and the communication is achieved using message passing. An additional cost due to the communications is then unavoidable and the interconnection network is crucial for the parallel performances. Furthermore, the question of synchronization between the processors arises. No tool seems really efficient to automatically parallelize codes on such architectures. The CRAY T3E is an example of such a machine. With this kind of computer, the programming mode has to be Single Program Multi Data type [3, 4]. Recently, new architectures of super computers appeared. They are made of hyper nodes. Each one is equipped with a small number of vector or scalar processors with shared memory. These hyper nodes are connected together by a very high speed network. The exchange of data is performed by accessing the same memory address. Semaphores are used to prevent the problem of simultaneous access to the same data by several processors. Automatic vectorization and parallelization tools are available. With such an architecture, multi-programming modes can be implemented. It can be mixed with a larger granularity of programming. A code can use more than one hyper node and communications are then performed with message passing library. An example of this kind of machine is the NEC SX-5. It is equipped with 4 hyper nodes. Each one is made of 16 vector processors with 128Gb of shared memory. However access to such a computer remains expensive. It is then possible to obtain good computation performances by using a cluster of workstations or personal computers (PC). Only one needs to be a conventional PC (screen, keyboard, hard disk, …), while the others may remain diskless and without others peripheries. Two networks are required, for the administration (low speed) and for the exchange of data (high speed). During the booting phase, each diskless node ask to the frontal node for an IP address and load the Proceedings of the International Conference on Parallel Computing in Electrical Engineering (PARELEC’02) 0-7695-1730-7/02 $17.00 © 2002 IEEE