Using the MPPA Architecture for UCT Parallelization
Résumé
We present here a study of the use of a MultiPurpose Processor Array (MPPA) architecture for the parallelization of the UCT algorithm applied to the field of General Game Playing. We evaluate the constraints imposed by this architecture and show that the only parallelization of UCT proposed in the literature that is feasible on MPPA is a leaf parallelization. We show that the MPPA provides good scalability when increasing the size of the communications, which is useful when using synchronous communications to send large sets of game initial positions to be processed. We consider two approaches for the calculation of the playouts: the distributed computing of a playout on each cluster and the calculation of several playouts per cluster; we show that the second approach gives better results. Finally, we describe experiments concerning the thread management and present a surprising result: it is more efficient to revive threads than keep them alive and try to communicate with them.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...