Resource Management for Data Analytics in Edge Computing Networks

Abstract

The emergence of Multi-access Edge Computing (MEC) aspires to divert the aggregation of data and their computation away from the cloud, and closer to the end users. At the same time IoT and mobile devices create a plethora of data, that in many cases are ephemeral and can be consumed at the network's edge. Hence, they should not burden the core network with data transfers, and the cloud with computations. We study the case of executing data analytic services on such networks. Data analytics can be very demanding in terms of computation and energy requirements, which can limit the effectiveness of executing on small edge devices. On the other hand, they are also delay sensitive, and solely relying on the cloud for their execution is inadequate. More importantly however, they are usually non-deterministic tasks, and the selection of execution algorithm or Machine Learning (ML) model can further impact their overall performance.

We study such trade-offs that arise in edge computing networks. We formulate and solve problems that decide the efficient allocation of communication and computation resources, towards optimizing key performance metrics for the underlying services. Moreover we provide the necessary incentives for the cooperation of nodes. Next, we propose a scheduling algorithm that makes no assumptions about the generally unknown system parameters, and thus is able to adapt to the varying system dynamics. We then study mobile edge computing systems, where a central edge node (cloudlet) serves multiple users based on resource availability. Finally, we delve into the specifics of a typical data analytic task, i.e. real time object recognition, and study the impact of application and network specific parameters to highlight various system trade-offs. We then proceed to optimally tune the service in an online fashion, paving the way towards automatic deployment of machine learning services over edge computing networks.