Machine-Learning-Aided Dynamic Reconfiguration in Optical DC/HPC Networks (Invited)
Sandeep Kumar Singh, Che-Yu Liu, S. J. Ben Yoo, Roberto Proietti · 2022
The high bandwidth and low latency requirements of modern computing applications with their dynamic and nonuniform traffic patterns impose severe challenges to current data center (DC) and high performance computing (HPC) networks. Therefore, we present a dynamic network reconfiguration mechanism that could satisfy the time-varying applications’ demands in an optical DC/HPC network. We propose a direct and an indirect topology extraction methods based on a machine learning-aided traffic prediction approach under multi-application scenario. The traffic prediction for topology extraction and bandwidth reconfiguration (PredicTER) method could lead to frequent topology and bandwidth reconfiguration. In contrast, the indirect approach, namely traffic prediction with clustering for topology extraction and bandwidth reconfiguration (PrediCLUSTER), utilizes an unsupervised learning-based clustering model to first associate the predicted traffic to one of possible traffic clusters, and then extracts a common topology for the cluster. This restricts the reconfigured topology set to the number of traffic clusters. Our simulation results show that the time-average of mean packet latencies (and total dropped packets) over 60 seconds of timevarying traffic under the PredicTER, PrediCLUSTER and a static topology are 37.7μs,41.2μs, and 50.2μs (and 37,967, 12,305, and 36,836), respectively. Overall, the PredicTER (and PrediCLUSTER) method(s) can improve the end-to-end packet latency by 24.9% (and 17.8%), and the packet loss rate by −3.1% (and 66.6%), as compared to the static flat Hyper-X-like topology.