Better algorithms for parallel backtracking

Abstract

Many algorithms in operations research and artificial intelligence are based on the backtracking principle, i.e., depth first search in implicitly defined trees. For parallelizing these algorithms, an efficient load balancing scheme is of central importance. Previously known load balancing algorithms either require sending a message for each tree node or they only work efficiently for large search trees. This paper introduces new randomized dynamic load balancing algorithms for tree structured computations, a generalization of backtrack search. These algorithms only need to communicate when necessary and have an asymptotically optimal scalability forhypercubes, butterflies and related networks, and an improved scalability for meshes and hierarchical networks like fat trees.