Scheduling with Machine-Learning-Based Flow Detection for Packet-Switched Optical Data Center Networks

Abstract

A scalable, low-latency, high-speed, and energy-efficient data center network is a key element in the deployment of future large-scale data centers, and photonic switching has recently been recognized as a promising solution to fulfill these goals. In this study, we present a packet-switched optical network (PSON) architecture with centralized control for intra-data-center connectivity. For efficient PSON operation, intelligent yet low-complexity bandwidth-scheduling algorithms are critical. To align with realistic traffic flows in a data center, we consider “mice flow,” which occurs frequently but carries a small number of bytes, and “elephant flow,” which occurs occasionally but has a huge number of bytes. To classify traffic flows with different characteristics, we investigate various machine-learning (classification) techniques, such as C4.5 and Naïve Bayes Discretization, and compare their performance in terms of accuracy and classification speed. We also develop a priority-aware scheduling algorithm for packet switching, which is optimized for PSON, and is adaptive to flow classification under a dynamic traffic scenario. Numerical simulations show that our proposed scheduling algorithm assisted by flow-classification techniques can outperform a benchmark algorithm in terms of average delay and packet-loss ratio.