Abstract: The recent state of the Internet has seen increased diversity in networked devices, traffic workloads, and TCP congestion control algorithms. For example, mobile devices have outgrown desktop devices in terms of Web traffic and is a segment slated to continue growing. In terms of traffic workloads, video traffic has become a dominant traffic segment in the Internet in terms of bytes and JSON traffic has grown significantly on the CDN. Further, new congestion controls like BBR and DCTCP, which move away from traditional TCP loss mechanisms, have become widely adopted and used throughout the Internet. In spite of all of these trends, it is unclear what effect these new inter-connected devices, traffic workloads, and congestion controls have on application performance. 

To address the above question, we perform measurement-driven analysis of these emerging devices, workloads, and congestion controls. In the first part of the dissertation, we explore the influence of the novel BBR congestion control algorithm on DASH video workloads. We find that BBR significantly underperforms compared to Cubic, the default and most popular TCP congestion. As a result, we propose changes to the mechanics of BBR's bandwidth estimation that significantly improve DASH video quality of experience. In the second part, we examine the relationship between mobile device configurations and BBR congestion control in terms of iPerf performance. Initially, we discovered that TCP packet processing suffers significantly under low-end mobile CPUs. When examining low-end mobiles and BBR, we see that BBR performs poorly under multiple connections compared to Cubic. We observe that TCP packet pacing, a new technique used by BBR, is responsible for the performance degradation. As a result, we provide optimization to TCP's packet pacing mechanism that improves iPerf performance. In the third part of our dissertation, we explore the performance of DCTCP and Cubic congestion controls when sharing router buffers in the data center. We primarily learn a generalizable ML model to capture the relationships between router configurations, network settings, and TCP performance. We find that both algorithms can saturate router buffers and starve each other if router configurations are set incorrectly, and suggest that our ML model can be used to find optimal router configurations for desired TCP performance. Our last work characterizes the JSON traffic workload on a popular CDN. We find that the JSON traffic workload has temporal and sequential patterns that can be used to improve CDN cache performance via prefetching. 

It is our thesis that large-scale measurement-driven approaches can help characterize and improve performance of applications (video, iPerf, HTTP traffic) under emerging end-devices, congestion controls, and traffic workloads for diverse networks.