DiversiFi: robust multi-link interactive streaming

Abstract

Real-time, interactive streaming for applications such as audio-video conferencing (e.g., Skype) and cloud-based gaming depends critically on the network providing low latency, jitter, and packet loss, much more so than on-demand streaming (e.g., YouTube) does. However, WiFi networks pose a challenge; our analysis of data from a large VoIP provider and from our own measurements shows that the WiFi access link is a significant cause of poor streaming experience. To improve streaming quality over WiFi, we present DiversiFi, which takes advantage of the diversity of WiFi links available in the vicinity, even when the individual links are poor. Leveraging such cross-link spatial and channel diversity outperforms both traditional link selection and the temporal diversity arising from retransmissions on the same link. It also provides significant gains over and above the PHY-layer spatial diversity provided by MIMO. Our experimental evaluation shows that, for a client with two NICs, enabling replication across two WiFi links helps cut down the poor call rate (PCR) for VoIP by 2.24x. Finally, we present the design and implementation of DiversiFi, which enables it to operate with single-NIC clients, and with either minimally modified APs or unmodified APs augmented with a middlebox. Over 61 runs, where the baseline average PCR is 4.9%, DiversiFi running with a single NIC, switching between two links, helps cut the PCR down to 0%, while duplicating wastefully only 0.62% of the packets and impacting competing TCP throughput by only 2.5%. Thus, DiversiFi provides the benefit of multi-link diversity for real-time interactive streaming in a manner that is deployable and imposes little overhead, thereby ensuring coexistence with other applications.