Performance Evaluation of Graph Partitioning on Many-core System

Abstract

Max-flow or min-cut is one of the important strategies for modelling and addressing practical problems in computer vision, image processing, and optimization theory using graph algorithms. It's been extensively used in a variety of applications, including image segmentation, flow network. As graph can be large and complex in terms of the amount of nodes/edges and the connection between them, the graph partitioning technique helps to divide the graph into sub-parts by using max-flow and min-cut method. According to the max-flow min-cut, the total weight of the edges in a flow network's minimal cut equals the maximum quantity of flow travelling from the source to the sink. This paper describes both serial and parallel implementations of the push-relabel approach for max-flow and min-cut. The parallel implementation on many-core system (GPGPU) shows better speedup than serial implementation of the prooosed algorithm.