Parallelizing the Crossword Generation Game in Orca

The Crossword Generation Game (kece) is a problem from the Cowichan Set Wil94], a programming benchmark designed to compare the usability of parallel systems. This paper describes the implementation of-search on kece games in the parallel language Orca BKT92], running on the Amoeba distributed operating system MvRT + 90]. A benchmark serial kece implementation was rst written in ANSI C. Some computational analysis was performed on the problem. Kece game trees turned out to have a quite variable branching factor, but showed very little variation in node values. Serial and parallel versions were then developed in Orca. This took little eeort, but the serial Orca program proved 20 times slower than the serial C version. Prooling was used to determine which Orca constructs were bottlenecks. Performance improved sharply to a factor of 1.2 slower, but only at the expense of avoiding the use of Orca objects and graphs. Measurements were made on an Amoeba system consisting of 80 SPARC processors connected by Ethernet. Linear speedup with 20% eeciency turned out to be feasible, when search parameters were carefully chosen.