Vertex-Centric Graph Processing: The Good, the Bad, and the Ugly

TL;DR

This paper systematically compares distributed vertex-centric graph algorithms with sequential ones, revealing inefficiencies and workload imbalances in most implementations, and discusses challenges in expressing certain algorithms within this framework.

Contribution

It provides a comprehensive analysis of vertex-centric graph algorithms' complexity and workload balance, highlighting limitations and guiding future improvements.

Findings

Most algorithms perform more work than their sequential counterparts.

Workload imbalance and many iterations are common issues.

Euler tour tree algorithm is an exception in efficiency.

Abstract

We study distributed graph algorithms that adopt an iterative vertex-centric framework for graph processing, popularized by the Google's Pregel system. Since then, there are several attempts to implement many graph algorithms in a vertex-centric framework, as well as efforts to design optimization techniques for improving the efficiency. However, to the best of our knowledge, there has not been any systematic study to compare these vertex-centric implementations with their sequential counterparts. Our paper addresses this gap in two ways. (1) We analyze the computational complexity of such implementations with the notion of time-processor product, and benchmark several vertex-centric graph algorithms whether they perform more work with respect to their best-known sequential solutions. (2) Employing the concept of balanced practical Pregel algorithms, we study if these implementations suffer from imbalanced workload and large number of iterations. Our findings illustrate that with the exception of Euler tour tree algorithm, all other algorithms either perform more work than their best-known sequential approach, or suffer from imbalanced workload/ large number of iterations, or even both. We also emphasize on graph algorithms that are fundamentally difficult to be expressed in vertex-centric frameworks, and conclude by discussing the road ahead for distributed graph processing.