Execution of Graph Algorithms on GPU Graph Frameworks

Abstract

Graphical Processing Units (GPUs) are well-suited for graph analytics problems because of their ability to parallelize applications and speed up computation. As a result, over the last decade, several high-level GPU graph frameworks have evolved [9]; these frameworks allow programmers to concentrate on expressing primitives, since they themselves take care of scaling up to parallel architectures. A few examples of these frameworks include nvGRAPH [5], CuSha [3] and Gunrock [8]. Each one of these frameworks is built around a unique programming model, which governs how it manages data movement, memory accesses, load balancing as well as how it maps irregular graph topologies to parallel hardware.

While there have been studies of individual high-level GPU graph frameworks in the past [3, 5, 8], a comprehensive study comparing their differences has not been done. A thorough investigation of these frameworks’ performance can provide useful insights about the benefits of using specific programming models for specific graph topologies, and help accelerate a variety of irregular graph applications.

In this research, we execute two graph algorithms – Single Source Shortest Paths (SSSP) and PageRank – on three high-level GPU graph frameworks – nvGRAPH, CuSha, and Gunrock – on a set of diverse graph topologies, and benchmark performance on NVIDIA Tesla P40 and P100 GPUs.

The results show that CuSha provides the best-in-class performance on traversal-based primitives such as SSSP, whereas nvGRAPH and Gunrock are the best frameworks to use for dense-based-computation primitives such as PageRank. Between nvGRAPH and Gunrock for PageRank, nvGRAPH is better suited for large low-diameter graphs, Gunrock is better suited for large high-diameter graphs, and they are both equally well-suited for small graphs of any diameter.