Concurrent computing by sequential staging of tasks

Abstract

Concurrent computing by sequential staging of tasksDescribed is a new approach to parallel formulation of scientific problems on shared-memory mUltiprocessors such as the IBM ES/3090 system. The class of problems considered is characterized by repetitive operations applied over the computational domain D. In each such operation, some fields of interest are extrapolated or advanced by an amount of l::,.T. The integration variable T may be time, distance, or iteration sequence number, depending on the problem under con-sideration_ An extensively studied approach to parallel formulation of such computational problems is based on domain decomposition, which attempts to partition the domain of integration into many pieces, then construct the global solution from these local solutions. Thus, domain decomposition methods are confined to D alone at a single T level. An inquiry into the possibilities of formulating parallel tasks in T, or more significantly in the D x T domain, opens up new horizons and untapped opportunities. The aim of this paper is to detail an approach to exploit this T domain parallelism that will be referred to as sequential staging of tasks (SST). Concurrency is realized by means of ordering the tasks sequentially and executing them in a partially overlapped or pipelined manner. The SST approach can yield remarkable speedup for jobs requiring intensive paging I/O, even when a single processor is available for executing multiple tasks. Noteworthy features of the SST method are demonstrated and highlighted by using results obtained from computer experiments performed with a numerical solution method of the Poisson equation and migration of seismic reflection data.O ver the past four decades the computer industry has experienced phenomenal growth. The performance of scientific computers has increased by at least five orders of magnitude. These improvements can be attributed to advancements in technology, 646 GAZDAG AND WANG improvements in machine organization, and the developments of reliable SIMD (single-instruction multiple data) extensions, such as the pipelined vector processors.