Traffic modeling and optimization in datacenters with graph neural network

Abstract

Traffic Optimization (TO) is a well-known and established topic in datacenters with the fundamental goal of operating networks efficiently. Traditional TO heuristics may suffer from performance penalty as it mismatches actual traffic, while Artificial Intelligence (AI) which has undergone a renaissance recently is gradually being applied to the network optimization and has shown excellent advantages. However, the current AI technologies (e.g., DGN, DRL, DBA, etc.) have difficulty in adapting to the dynamic and variable characteristics of the network due to their lack of generalization ability, which limits the development of intelligent networks. As Graph Neural Network (GNN) can support relational reasoning and combinatorial generalization, we research how to model and optimize traffic in datacenters with GNN in this paper. First, we proposed a GNN model for reasoning Flow Completion Time (FCT), which is able to provide accurate estimation of never-seen network states. Then we designed a GNN-based optimizer for TO, which can be used to in flow routing, flow scheduling and topology management. Finally, the experimental results verify that the GNN model has a high inference accuracy, and the GNN-based optimizer can significantly reduce the average / p10 (the 10th percentile) FCT. Therefore, GNN has a great potential in network modeling and optimization, and has a wide range of applications.