Scaling 5G Cellular Networks using Distributed Radio Access Networks
Dr Tulika Mitra, Provost'S Chair Professor, School of Computing
Abstract:
The talk focuses on a key component of cellular networks: the Radio Access Network (RAN). The RAN provides radio access and manages resource allocation among users in the cellular network. Current state-of-the-art RAN architecture involves a Cloud-based RAN (CRAN), where digital baseband processing from multiple base stations is centralized in a cloud to enable multiplexing of resources which enables operators to reduce costs and introduce new features with ease. The series of works study how NFV and SDN can help design a cellular network that not only scales efficiently with the increase in the number of devices but also supports a diverse set of applications with varying Service Level Objectives (SLOs) while reducing total costs.
The first work is a novel RAN architecture called uvRAN which uses NFV to separate RAN functionality into common and user-specific functions. The introduction of user-level virtualization also facilitates slicing of processing resources in the CRAN where users with similar SLOs can be placed together to improve performance isolation between different user types. To show the feasibility of our proposed architecture, we build a prototype implementation of uvRAN on top of an open-source software-based implementation of the LTE stack. We also perform a large-scale simulation using actual network traces to evaluate the scalability of uvRAN. Our findings suggest that using a simple load balancer uvRAN can improve performance while requiring fewer processing resources.
To support user-level virtualization in uvRAN, we also need a corresponding solution for slicing the fronthaul, which connects the base station to the CRAN. We design a Fronthaul Slicing Architecture (FSA) using programmable switches, which is the first of its kind to provide user identification for fronthaul traffic at line rate. FSA enables multipoint-to-multipoint routing and fronthaul aggregation, which are two key components required for uvRAN. Our evaluation using real network traces shows that FSA can route packets in the fronthaul and also handle network events such as reordering and packet drops. Using FSA, we also build a packet prioritization algorithm that can prioritize users belonging to critical use cases or users with extensive RAN processing.
Finally, the last work aims at reducing the power consumption of hosts in uvRAN. We develop PR3 which is a power governor that scales the processor's operating frequency and active number of cores based on the network load. The key idea here is to use the advanced network schedule generated by the cellular network to predict the expected workload. We model our problem as a Markov Decision Process and use Reinforcement Learning to find the optimal action for each state. After training the model offline on a limited number of workloads, our results show that there are significant savings in power consumption while reducing processing latency across a wide range of traffic loads.