Modeling Data Center Networks with Message Passing Neural Network and Multi-task Learning

Abstract

Network modeling is a pivotal component to operate network efficiently in future Software Defined Network (SDN) based Data Center Networks. However, obtaining a general network model to produce accurate predictions of key performance metrics such as delay, jitter or packet loss jointly at minimal cost is difficult. To this end, we propose a novel network model based on message passing neural network (MPNN) and multi-task learning, which could unveil the potential connections between network topology, routing and traffic characteristics to produce accurate estimates of per-source/destination mean delay, jitter and packet drop ratio with only one model. Specifically, an extended multi-output architecture is proposed and an elaborate loss function is introduced to facilitate the learning task. In addition, we present the modules of our simulation environment for generating the training samples, which is generic and easy to deploy. Experimental results show that our approach can get better performance compared to the state of the art.