Effective Parallel Backtracking Methods for Operations Research Applications

We present two generic schemes for heuristic backtracking search on massively parallel MIMD systems. The rst one employs a dynamic work-load balancing strategy based on recursive stack splitting. Due to high communication overheads it can only be run on moderately parallel systems with up to 256 processors. Our second scheme employs a work-load balancing strategy based on pre-computed, xed work packets. In its iterativedeepening version, the global work-load balance is automatically improved during the search, resulting in an overall e ciency of 80% on a 1024-processor MIMD system. Analytical results indicate that similar performance can be expected on even larger MIMD systems, provided that the input problem size increases quadratically with the system size. Both schemes are easily portable. They use commonMIMD message passing primitives and can be applied to a wide range of application domains in operations research and arti cial intelligence.