20.1 Lessons

The CP program, from its very initial proposal and project implementation, was designed to directly answer such questions as:

• Does parallel computing work?
• What are the needed software and hardware?

The contents of this book illustrate our answers to these questions with such results as:

• Large synchronous and loosely synchronous problems parallelize in a scalable fashion.
• Domain decomposition is a universal methodology for massive parallelism.
• C, Fortran-plus message-passing is a powerful but low-level software model of general effectiveness.
• Data-parallel languages are probably more attractive to most users. These languages are much harder to implement (requiring sophisticated compilers) and not quite as general as message passing. The latter handles functional parallelism and asynchronous computations outside the scope of data parallelism.

As in all research projects, we made many unexpected discoveries. One of the most interesting was Computational Science.  Namely, much of the work described in this book is clearly interdisciplinary. It mixes physics, chemistry,  engineering and other applications with mathematics and computer science. The national high performance computing initiative has stressed interdisciplinary teams in both its planning documents [FHPCP:89a] and implementations in federal proposal (Commerce Business Daily) solicitations. This idea was indeed part of the initial makeup and proposals of CP. However, this is not actually what happened in many cases. Probably the most important work in CP was not from teams of individuals-each with their own specialized skills. Rather, CP relied on the research of individuals and each individual possessed a mix of skills. We can give some examples.

Otto developed the initial QCD codes (Chapter 4.3) for the Cosmic Cube and its prototype. This required intricate knowledge of both the best physics and its numerical formulations. However, Otto also participated in the design and implementation of the hardware and its support software which later became Express. Otto obtained a physics Ph.D., but is now on the computer science faculty at the Oregon Graduate Center.

As a different example, we can quote the research in Chapter 11 which uses physics methods (such as simulated annealing)  to solve a mathematics problem (optimization) for a computer science application (load balancing). Again, the design of higher level languages (Chapters 13, 15 through 17) requires deep computer science compiler  and operating system expertise, as well as application understanding to design, say, the high-performance Fortran directives or MOVIE functionality. This mix of interests, which combines the skills of a computer scientist with those of an application area such as physics, was the rule and not the exception in CP. In the following, we comment on some general implications of this.