Accelerating continuous GRASP with a GPU

This work proposes a GPU-based parallelization for the Continuous GRASP (C-GRASP), a local search metaheuristic for finding cost-efficient solutions to continuous global optimization problems subject to box constraints. C-GRASP has demonstrated competitive performance on several well-known multimodal test functions as well as on difficult real-world problems, hence a GPU parallelization might increase even further the applicability of this metaheuristic. Although GPU parallelizations have been proposed for many metaheuristics, little has been done for C-GRASP. We conduct an extensive set of experiments and compare our proposal with state-of-the-art GPU parallelizations of other metaheuristics, such as Scatter Search and Differential Evolution. We also compare our GPU approach with two other C-GRASP implementations: a sequential version and a multi-core version. Experimental results show our GPU C-GRASP outperforms other GPU-based metaheuristics and the multi-core C-GRASP. Besides, we observed speedups of up to \(154{\times }\) over the sequential version.