Studying overheads in massively parallel MIN/MAX-tree evaluation

Studying Overheads in Massively Parallel Min/Max-lllee Evaluation (Extended Abstract) “t Rainer Feldmann and Peter Mysliwietz and Burkhard Monien Email: chess@uni-paderborn.de Department of Mathematics and Computer Science, University of Paderborn, Germany In this paper we study the overheads arising in our algorithm for distributed evaluation of Min/Max trees. The overheads are classified into search overhead, performance loss, and decrease of work load. Several mechanisms are investigated to cope with these overheads in order to achieve a high performance. We study a combination of local, medium range, and global load distribution strategies that does not only show a good behavior in terms of work load, but also has a positive influence on the search overhead. The efficient use of a virtual shared memorv. that is distributed among the processors, shows also a big ‘contribution to the overal~performance of the system. A carefully restricted application of parallelism using an improved version of the Young Brothers Wait Concept (YBWC) leads to a perfect behavior for minimal Min/Max trees and to a quite low search overhead, if well ordered trees are searched. Well ordered trees const itute the most important case in practice, since a couple of move ordering mechanisms are known that achieve a nearly optimal move ordering in many applications. The resulting combination of the methods shows an efficiency better than any previous approach. Experiments carried out using 256 DeBruijn-connected Transputers result in a speedup of 142 even applying restricted timing constraints. With a system consisting of 1024 grid connected Transputers we obtain a speedup of 344. Moreover the algorithm shows a very good scalability, especially using interconnection networks with logarithmic diameter. The experiments have been carried out using a Min/Max search program that incorporates all important state-of-theart search techniques ( ZUGZWANG, current vice world champion in computer chess) and therefore makes sure, that no artificial or simplifying assumptions on the structure of the problem are made. *This work was partly supported by the ESPRIT proJect GPMIMD and the ESPRIT Basic Research Action No. 7141 (ALCOM II) tThe paderborn Center for Parallel Computing PC2 provided Us with the parallel hardware for our experiments Permission to copy without fee all or part of this material is granted provided that the copies are not made or distributed for direct commercial advantage, the ACM copyright notice and the title of the publication and its date appear, and notice is given that copying is by permission of the Association of Computing Machinery. To copy otherwise, or to republish, requires a fee and/or specific permission. SPAA 94-6194 Cape May, N.J, USA (3 1994 ACM 0-89791-671 -9N410006..$3.5O