An Implementation of State-of-the-Art Graph Bisection Algorithms

In recent decades, various graph partitioning heuristic algorithms have been developed by researchers, which are widely used in “realworld” applications. As a starting point of graph partition problem, graph bisection algorithms provide basic approaches and solutions for other graph partition instances. This work implemented part of contemporary graph bisection algorithms which produce high-quality partitions for graphs in wide-range domains. These algorithms include Kernighan-Lin heuristic[3], Linear Kernighan-Lin heuristic[4], Greedy Growing Partitioning[5], Min-Max Greedy heuristic[6], Randomized Tabu Search heuristic[6], Reactive Randomized Tabu Search heuristic[6]. Multilevel graph partition technique[5] has not been included in current implementation to achieve better performance. Instead, these heuristics are extended in grid environment so as to significantly reduce run-time by distributing mutually exclusive iterations over computing nodes with support from GXP cluster shell[9]. Further, this parallel bisection needs no communication among computing nodes, thus bearing a high fault-tolerance. Finally, experiments show that although with more computation effort comparing to mulitilevel partition, this graph bisection library is still practically useful with its state-of-the-art result.