Dissecting HTTP/2 and QUIC : measurement, evaluation and optimization

The Internet is evolving from the perspective of both usage and connectivity. The meteoric rise of smartphones has not only facilitated connectivity for the masses, it has also increased their appetite for more responsive applications. The widespread availability of wireless networks has caused a paradigm shift in the way we access the Internet. This shift has resulted in a new trend where traditional applications are getting migrated to the cloud, e.g., Microsoft Office 365, Google Apps etc. As a result, modern web content has become extremely complex and requires efficient web delivery protocols to maintain users’ experience regardless of the technology they use to connect to the Internet and despite variations in the quality of users’ Internet connectivity. To achieve this goal, efforts have been put into optimizing existing web and transport protocols, designing new low latency transport protocols and introducing enhance- ments in the WiFi MAC layer. In recent years, several improvements have been introduced in the HTTP protocol resulting in the HTTP/2 standard which allows more efficient use of network resources and a reduced perception of latency. QUIC transport protocol is another example of these ambitious efforts. Initially developed by Google as an experiment, the protocol has already made phenomenal strides, thanks to its support in Google’s servers and Chrome browser. However there is a lack of sufficient understanding and evaluation of these new protocols across a range of environments, which opens new opportunities for research in this direction. This thesis provides a comprehensive study on the behavior, usage and performance of HTTP/2 and QUIC, and advances them by implementing several optimizations. First, in order to understand the behavior of HTTP/1 and HTTP/2 traffic we analyze datasets of passive measurements collected in various operational networks and discover that they have very different characteristics. This calls for a reappraisal of traffic models, as well as HTTP traffic simulation and benchmarking approaches that were built on the understanding of HTTP/1 traffic only and may no longer be valid for modern web traffic. We develop a machine learning-based method compatible with existing flow monitoring systems for the classification of encrypted web traffic into appropriate HTTP versions. This will enable network administrators to identify H1 and H2 flows for network managements tasks such as traffic shaping or prioritization. We also investigate the behavior of HTTP/2 stream multiplexing in the wild. We devise a methodology for analysis of large datasets of network traffic comprising over 200 million flows to quantify the usage of H2 multiplexing in the wild and to understand its implications for network infrastructure. Next, we show with the help of emulations that HTTP/2 exhibits poor performance in adverse scenarios such as under high packet losses or network congestion. We confirm that the use of a single connection sometimes impairs application performance of HTTP/2 and implement an optimization in Chromium browser to make it more robust in such scenarios. Finally, we collect and analyze QUIC and TCP traffic in a production wireless mesh network. Our results show that while QUIC outperforms TCP in fixed networks, it exhibits significantly lower performance than TCP when there are wireless links in the end-to-end path. To see why this is the case, we carefully examine how delay variations which are common in wireless networks impact the congestion control and loss detection algorithms of QUIC. We also explore the interaction of QUIC transport with the advanced link layer features of WiFi such as frame aggregation. We fine-tune QUIC based on our findings and show notable increase in performance.