Lessons for massively parallel applications on message passing computers

摘要

A decade's work on message passing MIMD (multiple instruction-multiple data) parallel computers in the areas of hardware, software, and applications is reviewed. The author concludes that distributed memory parallel computing works, and he describes the implications of this for future portable software systems. He claims that the message passing model was and will continue to be very successful. The vendors will build better hardware with lower communication latency and reasonable t/sub comm//t/sub calc/or approximately=10. The author views the message passing software model as 'assembly-language', which in many cases one can and should hide from the user with a software model 'nearer' that of the problem. Optimizing compilers will translate from a problem-oriented software model convenient for users to the message passing level supported by the machine. This latter level will continue to be used directly for applications for difficult cases which are not efficiently supported in the high-level software.