A New Measure for Partitioning of Block-Centric Graph Processing Systems

Block-centric graph processing systems have received significant attention in recent years. To produce the required partitions, most of these systems use general-purpose partitioning methods. As a result, the performance of them has been limited. To face this problem, special partitioning algorithms have been proposed by researchers. However, these methods focused on traditional partitioning measures like the number of cutting-edges and the load-balance. In return, the power of block-centric graph processing systems is due to unique characteristics that are focused on the design of them. According to basic and important characteristics of these systems, in this paper two new measures are proposed as partitioning goals. To the best of our knowledge, the proposed method is the first work that considers the diameter and size of the high-level graph as optimization factors for partitioning purposes. The evaluation of the proposed method over real graphs showed that we could significantly reduce the diameter of the high-level graph. Moreover, the number of cutting-edges of the proposed method are very close to Metis, one of most popular centralized partitioning methods. Since the number of required supersteps in block-centric graph processing systems mainly depends on the diameter of the high-level graph, the proposed method can significantly improve the performance of these systems.