Introduction

There are good reasons to believe that parallel computers will play a major role in future geophysical computing. First, the traditional serial computing is already close to its physical limits (Boghosian, 1990) and second, physical models and architectures of computers seem to develop towards similar topologies (Hillis, 1985).
Boghosian gives a good general introduction to one of the fastest, highly parallelized computers, the Connection Machine (CM). Most of the literature dealing with the Connection Machine describes special applications in geophysics, like those of Mora (1989) or Delany (1988) or in other fields (Long, 1989). Myers and Adams (1989) investigate the CM's performance in more detail.
The authors usually discuss only high-performance implementations. From the didactical point of view, comparing inefficient programs with faster versions can be more instructive. In this paper I will show the importance of an intelligent mapping of the data to the virtual processors, to reduce interprocessor communication and to achieve a high degree of parallelism.
I will demonstrate this by means of Claerbout's pixel-precise velocity analysis algorithm (Claerbout, 1989, and Claerbout, 1990) which will prove as a difficult choice.