Stratified sampling for even workload partitioning applied to single source shortest path algorithm

An efficient implementation of large graph processing algorithms on distributed-memory machines requires a balanced partitioning of the graph across the machines. In a previous paper we presented an algorithm, named Workload Partitioning and Scheduling (WPS), that uses domainspecific knowledge to guide a sampling procedure in large implicitly-defined graphs. WPS’s sampling procedure is used for partitioning the workload into parts of similar size which is then distributed amongst different machines. This article extends that earlier study and presents an investigation of the parallel and distributed implementation of Meyer’s ∆-Stepping algorithm for solving the Single-Source Shortest Path (SSSP) problem for directed graphs. Our implementation leverages the WPS algorithm for evenly distributing the workload involved in processing the vertices of the input graph across distributed-memory machines. In contrast with the previous study, which focussed on implicitly-defined graphs, this work demonstrates that WPS is also equally applicable on explicitly-defined graphs. Empirical evidence shows that applying WPS to Meyer’s SSSP algorithm yields significant performance benefits.