Sponsored by: Fastly

Abstract

Fastly is a cloud computing services provider seeking to improve their load-testing technologies. Currently, they are unable to generate enough network traffic to accurately represent a production environment and as a result, cannot precisely measure server performance. This project aimed to improve the range of tests Fastly can run on their software stack, allowing them to detect and reproduce production issues such as memory allocation for idle connections and throughput when processing TLS handshakes and requests concurrently. To achieve this goal, our team conducted comparative analysis of existing load-testing tools and extended PTR, an internal Fastly project. Some features we integrated include support for multiple parallel tests and generating load from K6, an open-source HTTP testing framework, as well as Idle Connections, a tool we wrote that generates connected yet dormant clients. We ran PTR against Fastly's infrastructure to recreate complex scenarios such as comparing performance across machines with different processors and server builds. From our results, we were able to uncover potential bottlenecks along with other server inefficiencies. By identifying and fixing these issues, Fastly can help provide reliable service to their customers.

Download

• Final report: pdf
• Slides: pdf