Frenchess: A Cray T3D at the 8th World Computer Chess Championship (1995)
| Venue: | In 1st European Cray-T3D Workshop |
| Citations: | 1 - 0 self |
BibTeX
@INPROCEEDINGS{Baudot95frenchess:a,
author = {Marc-françois Baudot and Jean-christophe Weill and Jean-luc Seret and Michel Gondran},
title = {Frenchess: A Cray T3D at the 8th World Computer Chess Championship},
booktitle = {In 1st European Cray-T3D Workshop},
year = {1995}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
This paper presents our work on parallel search algorithms at the University of Paris 8 and the Direction des Etudes et Recherches (E.D.F.). It focuses on our parallel chess program "Frenchess' ' which, running on the Commissariat à l'Energie Atomique's 128 processors T3D, just finished third of a field of 24 at the 8th World Computer Chess Championship in Hong Kong (May 25- May 30 1995). Frenchess relies on the new ABDADA (Alpha Beta Distribué Avec Droit d'Aînesse) algorithm, an efficient and easy to implement parallel version of the alpha-beta algorithm (which was formerly known not to parallelize well). ABDADA had first been tested on a TMC CM5 and a Cray CS6400, and the adaptation to the CRAY T3D, using SHMEM, has proved very satisfying. Even though single PE performance was disappointing, the T3D gave us excellent parallel performance, proving that our ABDADA algorithm scales very well up to at least 128 processors (with a speedup of over 70 in tournament conditions), and that it takes full advantage of the T3D's fast parallel architecture: it enabled us to build a world class chess program in just a few months ' work. The ABDADA algorithm can certainly be useful for the parallelization of related problems, such as Branch and Bound problems, on this architecture.
Keyphrases
world computer chess championship cray t3d energie atomique parallel chess program quot alpha-beta algorithm related problem cray cs6400 single pe performance full advantage abdada algorithm scale processor t3d parallel version bound problem parallel search algorithm world class chess program abdada algorithm excellent parallel performance new abdada tmc cm5 alpha beta distribu avec droit tournament condition parallel architecture hong kong direction de etude
