HTTP/3's Extensible Prioritization Scheme in the Wild

Abstract

For HTTP/2 and HTTP/3, multiple (Web page) resources are loaded by multiplexing them onto a single TCP or QUIC connection. A "prioritization system" is used to properly schedule the order in which the resources are sent. As HTTP/2's "prioritization tree" underperformed, a more straightforward setup called the Extensible Prioritization Scheme (EPS) was proposed for HTTP/3. This paper represents the first real-world measurement study into how this new scheme is supported and employed in practice by the three main browser engines and 12 different popular servers and cloud/CDN deployments. We find considerable heterogeneity in overall EPS (sub)feature support and even fundamental differences in approach/philosophy between the stacks. As incorrect prioritization can have a negative effect on (Web) performance metrics, our work not only provides essential insights for browser vendors and server deployments but also offers recommendations for future improvements.