Parallelization Strategies for GPU-Based Ant Colony Optimization Solving the Traveling Salesman Problem

摘要

Ant Colony Optimization (ACO) is a well known population-based algorithm used for solving combinatorial optimization problems, such as the Traveling Salesman Problem (TSP). The parallelization of ACO becomes necessary when tackling bigger instances of the TSP due to the high number of calculations performed. Many parallel approaches have been already proposed for ACO, in particular for contemporary high-performance hardware, such as GPUs. Typically, the ants are treated in parallel, since they are largely independent. In the case of the TSP, this concerns in particular the tour construction phase. Furthermore, strategies for parallelizing the pheromone deposit and evaporation phase were proposed. The achieved overall speedup hence depends on a combination of different parallelization strategies, where the impact of each strategy depends on the characteristics of the considered application problem and the hardware used. In the present paper, we aim to compare and analyze the performance of ACO implementations using distinct parallelization strategies when solving TSP instances of different magnitudes. At first, a comparison is made between a coarse-grain and a fine-grain parallel ACO. Furthermore the impact of the parallelization of the pheromone deposit process is also analyzed. The results show that there is no overall best parallelization strategy. Also, they highlight the importance of key-points that lead to a reduction of the execution time, such as the occupancy of the GPU and the work load shared among threads.