Parallelizing combinatorial optimization heuristics with GPUs

Abstract

Combinatorial optimization problems are often NP-hard and too complex to be solved within a reasonable time frame by exact methods. Heuristic methods which do not offer a convergence guarantee could obtain some satisfactory resolution for combinatorial optimization problems. However, it is not only very time consuming for Central Processing Units (CPU) but also very difficult to obtain an optimized solution when solving large problem instances. So, parallelism can be a good technique for reducing the time complexity, as well as improving the solution quality. Nowadays Graphics Processing Units (GPUs) have evolved supporting general purpose computing. GPUs have become many core processors, multithreaded, highly parallel with high bandwidth memory and tremendous computational power due to the market demand for high definition and real time 3D graphics. Our proposed work aims to design an efficient GPU framework for parallelizing optimization heuristics by focusing on the followings: distribution of data processing efficiently between GPU and CPU, efficient memory management, efficient parallelism control. Our proposed GPU accelerated parallel models can be very efficient to parallelize heuristic methods for solving large scale combinatorial optimization problems. We have made a series of experiments with our proposed GPU framework to parallelize some heuristic methods such as simulated annealing, hill climbing, and genetic algorithm for solving combinatorial optimization problems like Graph Bisection problem, Travelling Salesman Problem (TSP). For performance evaluation, we've compared our experiment results with CPU based sequential solutions and all of our experimental evaluations show that parallelizing combinatorial optimization heuristics with our GPU framework provides with higher quality solutions within a reasonable time.