Models and Experiments for High-Frequency Mobile and Data Center Wireless Networks

Abstract

There is growing interest in the design and operation of wireless networks at higher frequencies, e.g., 10+ GHz, to relieve the mounting pressure on wireless connectivity due to spectrum scarcity and physical limitations in the frequency bands below 5 GHz. Wireless networks at higher frequencies have the potential to provide data rates of multiple gigabits-per-second (Gbps) and leverage spatial reuse in dense deployments that are infeasible in lower frequency networks. While promising, the shift to higher frequency introduces new challenges and opportunities for algorithm and protocol design to maximize the efficiency of these networks. This thesis describes several of the fundamental challenges of high frequency wireless networks with directional antennas in dense deployments scenarios for data centers and mobile indoor wireless networks. Through the study of several tasks including modulation classification and localization for beam sector selection, this thesis argues that the performance and efficiency of high frequency wireless networks can be significantly impacted by machine learning methods. For the localization task, we propose domain specific modifications to a data-driven model to address collinearity and directional antenna constraints that provide significant improvements in accuracy. We also introduce methods for data generation and augmentation, specifically, the use of generative adversarial models for generating additional training data important for wireless signal learning tasks given the expense and difficulty of measurement campaigns. This thesis also explores the use of high frequency wireless networking in data centers to provide on-demand, reconfigurable connectivity to existing wired infrastructure. We argue that dynamically reconfigurable high frequency wireless devices can match the non-uniform connectivity seen in production data centers without introducing extensive wiring costs and weight necessary to accomplish similar capability with wired infrastructure. We introduce two challenges for deploying high frequency wireless devices in data centers: interference minimization for spatial reuse of dense deployments and all-to-all device network addressing for directional wireless devices. We provide a preliminary study of interference-aware scheduling to maximize spatial reuse for MapReduce tasks and show that these choices can have a large impact on potential performance. Finally, we describe the challenges with address allocation and routing tables for high frequency wireless directional devices. We demonstrate the Nested Buddy System and show its promise for reducing the number of routing entries in these networks.