Current trends in high performance parallel and distributed computing

Summary form only given. Parallel computing for high performance scientific applications gained widespread adoption and deployment about two decades ago. Computer systems based on shared memory and message passing parallel architectures were soon followed by clusters and loosely coupled workstations, that afforded flexibility and good performance for many applications at a fractional cost of MPP. Such platforms, referred to as parallel distributed computing systems, have evolved considerably and are currently manifested as very sophisticated metacomputing and Grid systems. This paper traces the evolution of loosely coupled systems and highlights specific functional, as well as fundamental, differences between clusters and NOW of yesteryear versus metacomputing Grids of today. In particular, semantic differences between Grids and systems such as PVM and MPICH are explored. In addition, the recent trend in Grid frameworks to move away from conventional parallel programming models to a more service-oriented architecture is discussed. Exemplified by toolkits that follow the OGSA specification, these efforts attempt to unify aspects of Web-service technologies, high performance computing, and distributed systems in order to enable large scale, cross-domain sharing of compute, data, and service resources. The paper also presents specific examples of current metacomputing and Grid systems with respect to the above characteristics, and discusses the relative merits of different approaches.