Data-Parallel Programming on MIMD Computers

Data-Parallel Programming demonstrates that architecture-independent parallel programming is possible by describing in detail how programs written in a high-level SIMD programming language may be compiled and efficiently executed-on both shared-memory multiprocessors and distributed-memory multicomputers.

MIMD computers are notoriously difficult to program. Data-Parallel Programming demonstrates that architecture-independent parallel programming is possible by describing in detail how programs written in a high-level SIMD programming language may be compiled and efficiently executed-on both shared-memory multiprocessors and distributed-memory multicomputers. The authors provide enough data so that the reader can decide the feasibility of architecture-independent programming in a data-parallel language. For each benchmark program they give the source code listing, absolute execution time on both a multiprocessor and a multicomputer, and a speedup relative to a sequential program. And they often present multiple solutions to the same problem, to better illustrate the strengths and weaknesses of these compilers. The language presented is Dataparallel C, a variant of the original C* language developed by Thinking Machines Corporation for its Connection Machine processor array. Separate chapters describe the compilation of Dataparallel C programs for execution on the Sequent multiprocessor and the Intel and nCUBE hypercubes, respectively. The authors document the performance of these compilers on a variety of benchmark programs and present several case studies.

Contents Introduction * Dataparallel C Programming Language Description * Design of a Multicomputer Dataparallel C Compiler * Design of a Multiprocessor Dataparallel C Compiler * Writing Efficient Programs * Benchmarking the Compilers * Case Studies * Conclusions