This report contains the results of the graduation project for the Computer Science course at the University of Amsterdam. The work for the project was done within the Information & Communication Service department at Shell Research, Amsterdam (Koninklijke/ShellLaboratorium, Amsterdam), where I have been employed since 1985. Professor Dr. L.O. Hertzberger, Dr. P.M.A. Sloot, Dr. P. van Emde Boas (University of Amsterdam) and Dr. P.A.J. van Deurzen (Shell Research, Amsterdam) were the supervisors. The initial goal of the project, defined in the spring of 1992, was to explore to what extent distributed systems could contribute to a more efficient use of the networked workstations installed at KSLA (Koninklijke/Shell-Laboratorium, Amsterdam). Follows some initial problems in defining the scope of the project, the idea was born to use PVM and Condor - two available packages supporting distributed computing - to run an existing scientific application. This would pave the way for a test with the DCE (Distributed Computing Environment) products from OSF (Open Software Foundation). However, partially due to personnel movement the test with the scientific application had to be postponed. The results of this study are: - a rudimentary form of DYNAMICPVM which combines PVM and Condor, - a better understanding of the impact the introduction of distributed systems will have on the KSLA organisation. I would like to thank Professor Hertzberger, who put me on the right track, Peter Sloot, who kept me on the track and Paul van Deurzen who patiently guided me past many obstacles during the project and had his interrupt bit turned off for me. A. de Beer, Th. Breet, R.P. Bosma and G. Noordenbos have provided me royally, on behalf of Shell, with all the facilities I required for this course...

This paper discusses methods which can be used to achieve improved computing performance from the available hardware. We illustrate general principles with a case study of the development and tuning of a seismic migration program on an IBM RISC System/6000 computer. We improved the uniprocessor performance from 8 to 26 Mflops on a 20 MHertz machine and restructured the code to run on a network of workstations. The principles can be applied to many scientific codes and are relevant to attaining high performance on a wide class of machines, including parallel and distributed systems. 1 INTRODUCTION Progress in numerically intensive computation (NIC) has relied largely on the ability of application experts to devise better algorithms and on the ability of engineers to build more powerful computing machinery and to connect more machines together in more powerful configurations. Unfortunately, there have not been comparable advances in the art and science of efficiently using these machine...

We show how task level parallelism can be exploited within the framework of DSO. We explain why this kind of parallelism is natural, and show how to implement a message passing program using PVM on a network of workstations. We present experimental results showing the success of our initial implementation, and indicate directions for further work. 1 Introduction It is becoming more and more fashionable to exploit clusters of workstations as if they were a single computer, and it has been already demonstrated that it is possible to obtain performances approaching, or better than that of supercomputers. PVM, the Parallel Virtual Machine is rapidly becoming the tool of choice for heterogeneous computing, as attested by the success of the First PVM User's Group [2]. This paper shows seismic inversion can benefit from this technique, by describing a message passing implementation of DSO, using PVM 3, that runs on a network of workstations. We use coarse grain parallelism, of SPMD type in w...