A (Somewhat Dated) Comparative Study of Betweenness Centrality Algorithms on GPU

TL;DR

This paper compares three GPU algorithms for calculating Betweenness Centrality in large, directed networks, analyzing their design trade-offs and performance on NVIDIA hardware.

Contribution

It provides a detailed analysis of different GPU algorithm strategies for BC computation, highlighting their trade-offs and performance implications.

Findings

Trade-offs in thread scheduling, atomic operations, and synchronization affect performance.

All algorithms tested on NVIDIA Tesla M2050 GPU.

Insights into optimizing BC algorithms for GPU architectures.

Abstract

The problem of computing the Betweenness Centrality (BC) is important in analyzing graphs in many practical applications like social networks, biological networks, transportation networks, electrical circuits, etc. Since this problem is computation intensive, researchers have been developing algorithms using high performance computing resources like supercomputers, clusters, and Graphics Processing Units (GPUs). Current GPU algorithms for computing BC employ Brandes' sequential algorithm with different trade-offs for thread scheduling, data structures, and atomic operations. In this paper, we study three GPU algorithms for computing BC of unweighted, directed, scale-free networks. We discuss and measure the trade-offs of their design choices about balanced thread scheduling, atomic operations, synchronizations and latency hiding. Our program is written in NVIDIA CUDA C and was tested on an NVIDIA Tesla M2050 GPU.