Resumen de Performance analysis and optimization of parallel Best-First Search algorithms on multicore and cluster of multicore.
Since the emergence of clusters, multicore processors and clusters of multicore machines, researchers and developers have faced the challenge of parallelizing different kinds of applications in order to take advantage of the computing power and/or the accumulated memory that these architectures provide. In the area of Artificial Intelligence, heuristic search algorithms are used as the basis to solve combinatorial optimization problems, such as: optimal route planning, robot navigation, optimal sequence alignments, among others (Russel & Norvig, 2003). One of the most widely used heuristic search algorithms for that purpose is A* (Hart, et al., 1968), a variant of Best-First Search, which requires high computing power and a large amount of memory. Consequently, during the last decades, the development of the parallel Best-First Search algorithms has been promoted which, in particular, may benefit from the parallel architectures mentioned formerly. The contribution of the thesis is the development of two parallel Best-First Search algorithms, one that is suitable for execution on shared-memory machines (multicore), and another one that is suitable for execution on distributed memory machines (cluster). The former is based on the adaptation of the HDA* (Hash Distributed A*) algorithm for multicore machines proposed by (Burns et al., 2010), while the latter is based on the HDA* (Hash Distributed A*) algorithm proposed by (Kishimoto, et al., 2013). The implemented algorithms incorporate parameters and/or techniques that improve their performance, with respect to the original algorithms proposed by the authors mentioned above. Additionally, a comparison of the performance achieved and the memory consumed by both algorithms, when they are run on a multicore machine, is presented. These results show that a benefit would be obtained by converting HDA* into a hybrid application, that uses programming tools for shared and distributed memory, when the underlying architecture is a cluster of multicore, and therefore the bases for this hybrid algorithm are established.
