Parallel Computer Architecture

Parallel computer architectures are now going to real applications! This fact is demonstrated by the large number of application areas covered in this book (see section on applications of parallel computer architectures). The applications range from image analysis to quantum mechanics and databases. Still, the use of parallel architectures poses serious problems and requires the development of new techniques and tools. This book is a collection of best papers presented at the first workshop on two major research activities at the Universitat Erlangen-N/imberg and Technische Universitat Munchen. At both universities, more than 100 resarchers are working in the field of multiprocessor systems and network configurations and methods and tools for parallel systems. Indeed, the German Science Foundation (Deutsche Forschungsgemeinschaft ) has been sponsoring the projects under grant numbers SFB 182 and SFB 342. Research grants in the form of a Sonderforschungsbereich are given to selected German Universities in portions of three years following a thoroughful reviewing process. The overall duration of such a research grant is restricted to 12 years. The initiative at Erlangen-Nurnberg was started in 1987 and has been headed since this time by Prof. Dr. H. Wedekind. Work at TU-Miinchen began in 1990, head of this initiative is Prof. Dr A. Bode. The authors of this book are grateful to the Deutsche Forschungsgemeinsehaft for its continuing support in the field of research on parallel processing. The first section of the book is devoted to hardware apects of parallel systems. Here, a number of basic problems have to be solved. Latency and bandwidths of interconnection networks are a bottleneck for parallel process communicatlon. Optoelectronic media, discussed in this section, could change this fact. The sealabillty of parallel hardware is demonstrated with the multiprocessor system MEMSY based on the concept of distributed shared memory. Scalable parallel systems need fault tolerance mechanisms to garantee reliable system behaviour even in the presence of defects in parts of the system. An approach to fault tolerance for scalable parallel systems is discussed in this section. The next section is devoted to performance aspects of parallel systems. Analytical models for performance prediction are presented as well as a new hardware monitor system together with the evaluation software. Tools for the automatic parallelization of existing applications are a dream, but not yet reality for the user of parallel systems. Different aspects for automatic treatment of parallel apphcations are covered in the next section on architectures and tools for paral]elizatlon. Dynamic lead balancing is an application transparent mechanism of the operating system to guarantee equal lead on the elements of a multiprocessor system. Randomizod shared memory is one possible implementation of a virtual shared memory based on distributed memory hardware.