Load balancing technique for parallel search with statistical model

A parallel performance enhancement technique using statistical model to increase parallel system efficiency is discussed in this paper. This idea is applied to solving the constraint satisfaction problem (CSP) on a multi-processor shared memory system. The combination of reducing standard deviation of the execution time in each processor and increasing the correlation coefficient between every two processors' execution times proves to be a very effective method in balancing load without resorting to dynamic approaches. Significant improvement in system efficiency over a traditional divide-and-conquer approach is shown in solving CSP in parallel. A near-optimal (100%) average load balance result is obtained. The underlying theory can also be easily applied to other problems and can be further extended to provide system performance prediction and evaluation.<<ETX>>