Limitations to parallel processing
Walter H. Burkhardt · 1990
The reasons why massively parallel processor systems show such dismal performance are analyzed. The degradation by seriality and communication has been investigated theoretically for linear, rectangular, and square configurations and by experiment on a Transputer system of modifiable dimension from 1 to 4, and up to 16 processors. Eight different application programs run can be grouped into academic and practical problems. They are written in a functional programming language for maximal parallelism. The measured results are analyzed by regression. The quadratic polynomial from the theoretical section shows the best fit. Optimal number of processors, maximal systems performance, and communications factors range from 5 to 14, from 2 to 7, and from 1% to 18%, respectively. Analysis of two experiments from literature on a tightly and a loosely coupled system show similar severe performance degradation, depending on configuration and problem type, with excellent applicability of the theoretical models.>