Performance of competitive OR-parallelism

Abstract

We present a very simple parallel execution model suitable for inference systems with nondeterministic choices (OR-branching points). All the parallel processors solve the same task without any communication. Their programs only differ in the initialization of the random number generator used for branch selection in depth first backtracking search. This model, called random competition, permits us to calculate analytically the parallel performance for arbitrary numbers of processors. This can be done exactly and without any experiments on a parallel machine. Finally, due to their simplicity, competition architectures are easy (and therefore low-priced) to build.

As an application of this systematic approach we compute speedup expressions for specific problem classes defined by their run-time distributions. The results vary from a speedup of 1 for linearly degenerate search trees up to clearly “superlinear” speedup for strongly imbalanced search trees. Moreover, we are able to give estimates for the potential degree of OR-parallelism inherent in the different problem classes. Such an estimate is very important for the design of particular parallel inference machines, since spedups strongly depend upon the application domain.