A Cache-Aware Parallel Implementation of the Push-Relabel Network Flow Algorithm and Experimental Evaluation of the Gap Relabeling Heuristic

Abstract

The maximum flow problem is a combinatorial problem of significant importance in a wide variety of research and commercial applications. It has been extensively studied and implemented over the past 40 years. The push-relabel method has been shown to be superior to other methods, both in theoretical bounds and in experimental implementations. Our study discusses the implementation of the push-relabel network flow algorithm on present-day symmetric multiprocessors (SMP's) with large shared memories. The maximum flow problem is an irregular graph problem and requires frequent fine-grained locking of edges and vertices. Over a decade ago, Anderson and Setubal implemented Goldberg's push-relabel algorithm for shared memory parallel computers; however, modern systems differ significantly from those targeted by their implementation in that SMP's today have deep memory hierarchies and different performance costs for synchronization and fine-grained locking. Besides our new cache-aware implementation of Goldberg's parallel algorithm for modern shared-memory parallel computers, our main new contribution is the first parallel implementation and analysis of the gap relabeling heuristic that runs from 2.1 to 4.3 times faster for sparse graphs.