Parallelization Strategies for Variable Neighborhood Search

We analyze five parallelization strategies for the Variable Neighborhood Search (VNS) meta-heuristic. They are based on asynchronous cooperation of several search threads. We tested parallelization on various levels, from a low-level parallel neighborhood exploration, through medium-grained asynchronous execution of basic VNS steps (shaking and local search), to the coarse-grained asynchronous cooperation of various VNS algorithms. We also compared centralized and non-centralized information exchange. Parallel algorithms were implemented on multiprocessor systems containing q identical processors. We used two topologies: star configuration for centralized information exchange and processor ring for the non-centralized one. For the experimental evaluation, we applied these strategies to VNS-based procedures for Multiprocessor Scheduling Problem with Communication Delays (MSPCD). We compared the performance of the parallel searches with that of the sequential execution on benchmark problem instances with up to 500 tasks. We achieved not only the improvement of the solution quality but also reduction in execution time. The generality of the proposed strategies and their straightforward implementation make them easily applicable to various difficult combinatorial optimization problems.