Research And Implementation Of GPU-based Swarm Intelligence Algorithm

With the continuous development of high-performance platforms such as GPU in recent years,the implicit parallelism of swarm intelligence algorithms provides a favorable platform foundation for the implementation of the algorithm on the GPU,and the swarm intelligence algorithm based on GPU platform is becoming the first choice to solve the high-dimensional complex optimization problem.Taking the particle swarm optimization algorithm as an example,this paper studies how to further utilize the computing performance of the CUDA(Compute Unified Device Architecture)platform on the GPU and give full play to the advantages of the parallelism of the swarm intelligence algorithm.The main tasks are as follows:In order to make full use of the computing resources of GPU and improve the convergence precision of the algorithm,a multi-swarm PSO algorithm based on CUDA streams is proposed.The parallel multi-swarm mechanism based on island model is combined with the stream concurrency mechanism of CUDA platform to improve the original algorithm.The degree of parallelism makes the parallel PSO algorithm achieve higher level of grid-level parallelism on the basis of the original thread-level parallelism,making full use of the computing resources of the GPU.When using 4 or 8 streams,the proposed algorithm reduces the runtime by about 30%compared to the single-population m,ode.In this paper,the convergence effect of pseudo-random number on the swarm intelligence algorithm on the CPU and the GPU is discussed,and the results show that using the pseudo-random number on GPU finally makes the probability of the algorithm converging to the threshold value increase by about 10%,and the parallelization degree of the swarm intelligence algorithm is improved,since the initialization of the algorithm is implemented in parallel.Finally,in order to solve the problem that PSO algorithm is easy to fall into local optimization,a multi-start local search algorithm framework based on CUDA dynamic parallel mechanism is proposed.Because of using dynamic parallel,when multiple initial solutions are searched in parallel,the neighbor solution of each solution will also be searched in parallel,which makes the parallel level of the algorithm is higher.By integrating the framework into the multi-swarm PSO algorithm based on CUDA streams,the algorithm obtains at least two orders of magnitude elevation compared with the average optimal solution of the algorithm before the improvement,which enhances the optimization ability of the algorithm and improves the robustness of the algorithm.