Bridging the Gap between QoE and QoS in Congestion Control: A Large-scale Mobile Web Service Perspective.

To improve the user experience of mobile web services, various congestion control algorithms (CCAs) have been proposed, yet the performance of the application is still unsatisfactory. We argue that the suboptimal performance comes from the gap between what the application needs (i.e., Quality of Experience (QoE)) and what the current CCA is optimizing (i.e., Quality of Service (QoS)). However, optimizing QoE for CCAs is extremely challenging due to the convoluted relationship and mismatched timescale between QoE and QoS. To bridge the gap between QoE and QoS for CCAs, we propose Floo, a new QoE-oriented congestion control selection mechanism, as a shim layer between CCAs and applications to address the challenges above. Floo targets request completion time as QoE, and conveys the optimization goal of QoE to CCAs by always selecting the most appropriate CCA in the runtime. Floo further adopts reinforcement learning to capture the complexity in CCA selection and supports smooth CCA switching during transmission. We implement Floo in a popular mobile web service application online. Through extensive experiments in production environments and on various locally emulated network conditions, we demonstrate that Floo improves QoE by about 14.3% to 52.7%.