Title: System support for fine-grained resource management in mobile edge computing

Ke-Jou Hsu

School of Computer Science

College of Computing

Georgia Institute of Technology

https://sites.cc.gatech.edu/~khsu38/

Date: Tuesday August 15th 2023

Time: 11:00 AM - 1:00 PM (EST)

Location: Klaus 3100 and https://gatech.zoom.us/j/4645200533

Committee:

Dr. Ada Gavrilovska (Advisor, School of Computer Science, Georgia Institute of Technology)

Dr. Ahmed Saeed    (School of Computer Science, Georgia Institute of Technology) 

Dr. Ketan Bhardwaj  (School of Computer Science, Georgia Institute of Technology)

Dr. Mostafa Ammar  (School of Computer Science, Georgia Institute of Technology)

Dr. Umakishore Ramachandra   (School of Computer Science, Georgia Institute of Technology)

Abstract: 

Multi-access (or Mobile) Edge Computing (MEC) encompasses computing infrastructure at the edges of the networks, such as at access gateways and base stations. Like the cloud, MEC presents a distributed multi-tenant platform for hosting edge applications, but potentially with faster response times and reduced backhaul network loads compared to cloud-based solutions. However, the limited resources available at a given edge location, the distribution scale and heterogeneity of the edge, and the latency-centric QoS requirements of edge workloads, make current cloud-native technologies not suitable for managing a multi-tenant edge. 

This thesis shows that to deliver on the promise of performance and efficiency benefits, the MEC infrastructure requires new system services that prioritize fine-grained operations as a core design principle. Our research reveals that by operating at fine granularity, new monitoring and network management service architectures, Colibri and ShapeShifter, can effectively mitigate up to 90% of performance violations compared to using existing cloud solutions. Furthermore, by decomposing applications into fine grained components, we provide greater deployment flexibility and control. For visual analytics workloads commonly used in edge computing, our approach, Couper, offers DNN model partitioning methods based on edge and cloud resources, resulting in reduced processing latency and frame drop rate. Similarly, fine grained control over the placement and configuration of individual components enables shorter cache query latency for MEC-CDN, by strategically placing DNS servers at edge locations.

We propose further applying these techniques to modern microservice-based application architectures and providing systems support for managing their edge cloud deployment and configuration.