Deep-Q: Traffic-driven QoS Inference using Deep Generative Network

Abstract

In today's IP network, it is important to provide the Quality of Service (QoS) guarantee for network services. However, in real networks with highly dynamic traffic demands, it is difficult to build an accurate QoS model even with a high cost of human expert analysis. In this paper, we present Deep-Q, a data-driven system to learn the QoS model directly from traffic data without human analysis. This function is achieved by utilizing the power of state-of-the-art deep generative networks in the deep learning area. Deep-Q provides a novel inference structure of a variational auto-encoder (VAE) enhanced by the long short-term memory (LSTM). A specially-designed module named Cinfer-loss is further applied to improve the QoS inference accuracy. By training with real traffic data, Deep-Q can infer a variety of QoS metrics over different networks given traffic conditions in real-time. We build testbeds for both the data center network and overlay IP network. Extensive experiments with 5.7TB traffic traces demonstrate that Deep-Q can achieve on average 3x higher inference accuracy than traditional queuing-theory-based solution in real networks while keeping inference time within 100ms.