Optimization of parallel iterated local search algorithms on graphics processing unit

Local search metaheuristics (LSMs) are efficient methods for solving hard optimization problems in science, engineering, economics and technology. By using LSMs, we could obtain satisfactory resolution (approximate optimum) in a reasonable time. However, it is still very CPU time-consuming when solving large problem instances. As graphic process units (GPUs) have been evolved to support general purpose computing, they are taken as a major accelerator in scientific and industrial computing. In this paper, we present an optimized parallel iterated local search algorithm efficiently accelerated on GPUs and test the algorithm with a typical case study of the Travelling Salesman Problem (TSP) in computational science. We introduce novel methods as follows: first, we present an efficient mapping between a neighborhood and a GPU thread. Second, we use the Roofline model to analyze the performance of existing GPU-based 2-opt kernels. Based on our analysis, we point out the limiting factor of these 2-opt kernels and provide our optimization approaches. Furthermore, we test our algorithm with standard TSP problem instances up to 4461 cities, in which our strategy leads to a speedup factor 279$$\times $$× over the sequential counterpart. We compare our approach with existing high-performance GPU-based local search algorithms, and the results demonstrate that the proposed algorithm is competitive.