AGL: a Scalable System for Industrial-purpose Graph Machine Learning

TL;DR

AGL is a scalable, fault-tolerant system designed for industrial graph machine learning, enabling efficient training and inference of GNNs on large-scale graphs using MapReduce infrastructure.

Contribution

The paper introduces AGL, a novel system that integrates training and inference for GNNs at scale, leveraging message passing and MapReduce for improved efficiency and scalability.

Findings

Trains a 2-layer GAT on billion-node graphs in 14 hours.

Completes inference on large graphs in 1.2 hours.

Supports large-scale industrial graph machine learning tasks.

Abstract

Machine learning over graphs have been emerging as powerful learning tools for graph data. However, it is challenging for industrial communities to leverage the techniques, such as graph neural networks (GNNs), and solve real-world problems at scale because of inherent data dependency in the graphs. As such, we cannot simply train a GNN with classic learning systems, for instance parameter server that assumes data parallel. Existing systems store the graph data in-memory for fast accesses either in a single machine or graph stores from remote. The major drawbacks are in three-fold. First, they cannot scale because of the limitations on the volume of the memory, or the bandwidth between graph stores and workers. Second, they require extra development of graph stores without well exploiting mature infrastructures such as MapReduce that guarantee good system properties. Third, they focus on training but ignore the optimization of inference over graphs, thus makes them an unintegrated system. In this paper, we design AGL, a scalable, fault-tolerance and integrated system, with fully-functional training and inference for GNNs. Our system design follows the message passing scheme underlying the computations of GNNs. We design to generate the $k$-hop neighborhood, an information-complete subgraph for each node, as well as do the inference simply by merging values from in-edge neighbors and propagating values to out-edge neighbors via MapReduce. In addition, the $k$-hop neighborhood contains information-complete subgraphs for each node, thus we simply do the training on parameter servers due to data independency. Our system AGL, implemented on mature infrastructures, can finish the training of a 2-layer graph attention network on a graph with billions of nodes and hundred billions of edges in 14 hours, and complete the inference in 1.2 hour.