Princeton University

School of Engineering & Applied Science

Data-Driven Management of CDN Performance

Mojgan Ghasemi
Engineering Quadrangle J401
Wednesday, September 6, 2017 - 1:30pm to 3:00pm

Most of the web traffic is carried by Content Distribution Networks (CDNs), who strive to offer good performance to users at a low cost. In this thesis, we introduce measurement and analysis techniques to help CDNs balance these goals. We start by allocating resources efficiently across the distributed CDN platform. We evaluate our methods in collaboration with the largest commercial CDN (Akamai), responsible for serving between 15% and 30% of all web traffic. We show that our solution enhances the performance significantly for a small increase in the operational cost.
Next, to detect and diagnose performance problems that cause poor experience for users, such as re-buffering while watching videos, we propose a fine-grained instrumentation of the video delivery path. We deploy our instrumentation and diagnosis methods in a commercial content provider (Yahoo) and uncover a wide range of problems that can cause re-buffering and were unknown before.
Finally, we dive deeper into diagnosing network problems by monitoring TCP connections directly in the network devices. Our tool can pinpoint the faulty component causing poor performance in a TCP connection. We deploy emerging programmable edge devices to implement our monitoring and diagnosis logic directly in the data plane, which runs at line-rate, without cooperation from servers.