An Efficient Parallelization Strategy for Dynamic Programming on GPU

Optimization methods generally do not fall into the most suitable algorithms for parallelization on a GPU. However, a relatively good efficiency still can be obtained if the method is properly adapted to the GPU programming model, which is the case for dynamic programming. In this article, we propose a parallelization strategy for thread grouping for dynamic programming in CUDA. We show that parametrizing the solver parallelism according to the hardware allows better performance. The strategy provides good acceleration compared to a standard GPU parallel strategy on a dynamic programming-based implementation of the knapsack problem. We show this strategy is helpful in the case of the multi-dimensional knapsack problem, where computing multi-dimensional indices is a costly operation.