A dynamic cooperative hybrid MPSO+GA on hybrid CPU+GPU fused multicore

Todays multi-core architectures with accelerators provide tremendous compute power. Population-based metaheuristic algorithms have proven particularly amenable to single instruction multiple data (SIMD)-style parallelization due to the fine-grained parallelism provided by these algorithms. While SIMD hardware allows one to run large scale simulations, obtaining better solution quality often requires a more thoughtful reorganization of the search technique itself. In this paper, we design a hybrid heuristic algorithm that dynamically alternates between Multi-Swarm Particle Swarm Optimization (MPSO) and Genetic Algorithm (GA) to improve solution quality. We parallelize the hybrid algorithm on a hybrid multicore computer, accelerated processing unit (APU) to improve performance. We take advantage of the close coupling the APU provides between CPU and GPU devices. Our hybrid algorithm results indicate an improvement in average solution quality over Multi-Swarm PSO across a set of standard mathematical optimization functions. We study the effect and performance of switching between CPU and GPU devices.