Data-Driven Management of CDN Performance

Abstract

Content Distribution Networks (CDNs) carry most of the web content, with the goals of offering good performance to users at a low cost. In this thesis, we introduce measurement and analysis techniques to help CDNs balance these goals.

First, we allocate resources efficiently across the distributed edge servers by jointly minimizing network latency and cache misses. We propose a unified framework for CDNs to jointly solve the problems of placement, mapping, and disk allocation, while including the impact of cache misses. We evaluate our methods using request logs from a commercial CDN. We show that including the impact of cache misses in the post-mapping disk allocation enhances the performance significantly for a small increase in the cost. Still, there are other sources of performance problems on the end-to-end (e2e) path of the servers to clients, which takes us to the next part.

Second, to detect and diagnose performance problems that cause poor experience for users, we propose a fine-grained instrumentation of the e2e path. In particular, we focus on the video delivery path because video is now the dominant application of the Internet. We deploy our instrumentation and diagnosis methods in a commercial content provider, enabling us to join the server-side, TCP statistics, and client-side measurements for the first time, and characterize the performance problems in a large set of videos. We uncover a wide range of problems, some of which were unknown before, and can only be discovered by an e2e instrumentation. While capable of diagnosing these problems, our tool is limited in how frequently it measurers network. To remedy this, we propose our final solution.

Finally, we dive deeper into diagnosing network problems by monitoring TCP connections directly in the network devices. Our tool can pinpoint if the performance of a TCP connection is hindered by the sender, receiver, or network. 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. We infer fine-grained TCP metrics from the edge device (e.g, NIC), without imposing storage or monitoring overhead on the servers.