Parallelization of Shortest Path Graph Kernels on Multi-Core CPUs and GPUs

In this paper, we present a study on the parallelization of the shortest path graph kernel from machine learning theory. We first present a fast sequential implementation of the graph kernel which we refer as Fast Computation of Shortest Path Kernel (FCSP). Then we explore two different parallelization schemes on the CPU and four different implementations on the GPU. After analyzing the advantages of each we propose a hybrid version which, for different pairs of graphs, dynamically chooses the best implementation from multicore execution and GPU execution. Finally, we apply our implementations to several datasets that are composed of graphs from different domains. We first evaluate our implementations on a set of synthetic datasets, then, we evaluate our implementations on a set of four real-world graph datasets. The results show that the sequential FCSP algorithm running on CPU is able to achieve a maximum 76x speedup over a naive sequential implementation of the shortest path graph kernel algorithm running on the same CPU. The results also show that our GPU implementation of FCSP offers a maximum 18x speedup over the sequential FCSP. Our GPU implementation also achieves a maximum 2x over a parallel CPU implementation of FCSP.