Synchronization issues of the parallel A Star Search

Abstract

An experimental study was carried out on the performance of synchronous and asynchronous implementations of the A* Search on a multiprocessor network. Master-Slave parallelism was used to distribute the Search among the 8 processor nodes of a transputer network. The test programs were run on 4 different types of maps. Measurements were taken in the form of percentages of time spent in computation and communication in each cycle of the search as an artificial delay in the computation phase was increased. The results from the map tests showed that the asynchronous implementation spent a larger percentage of each cycle performing calculations rather than communicating or waiting for communications as the artificial delay was increased. This means that the efficiency of the asynchronous approach increases more rapidly than the efficiency of the synchronous approach as the computational complexity of a parallel program is increased. This was found to be true for all artificial delays on all test maps for the Master-Slave A* Search. The results might vary with different implementations and search methodologies.