Please note: This PhD defence will take place in DC 2314 and online.
Sharon Choy, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Bernard Wong
The use of live video streaming applications over mobile, wireless networks continues to grow. In these applications, a sender (e.g., cameraperson) streams video to a receiver (e.g., video service such as YouTube or a television station), and the receiver disseminates the video to the viewers. Live video streaming over wireless (3G, LTE, 5G) links is challenging since these links often experience fluctuating latency and available bandwidth. Furthermore, as the bandwidth demands of these applications continue to increase, a single, wireless network link may be unable to continuously stream video that satisfies the viewers’ Quality of Experience (QoE) requirements. Multi-homing is a potential solution that offers applications additional bandwidth through aggregation and the ability to circumvent congested paths. Although there are a number of commercial multi-homed, adaptive live streaming solutions, their proprietary nature makes it difficult to understand their design trade-offs and evaluate their effectiveness. Furthermore, many multi-homed transport protocols are not designed for latency-sensitive data, do not adapt the video bitrate to changing network conditions, or require significant adjustments if used in a different environment from their original design (e.g., number or type of links). In this thesis, we explore solutions to the challenges of multi-homed, adaptive live video streaming.
Firstly, accurate bandwidth measurements are important because they indicate the maximum video bitrate that the available links can send while avoiding congestion. A packet train is a common measurement technique that sends a series of packets and uses the packets’ inter-arrival time to measure the available bandwidth. However, packet train measurements are inaccurate due to kernel interrupts. In this thesis, we propose our PacketBurst bandwidth measurement technique to improve the accuracy of the packet trains. This technique detects packet inter-arrival times that are affected by kernel interrupts and would yield inaccurate bandwidth measurements. Our PacketBurst technique excludes these packets from the bandwidth measurement calculation to improve the accuracy of the packet train. In addition to accurate bandwidth measurements, video streaming protocols can also benefit from future link quality predictions so that they may preemptively reduce the video bitrate or change network paths to avoid video stream disruptions. This thesis evaluates various machine learning techniques for classifying or predicting link instability, which we define as sudden increases in application-level packet loss and latency, or decreases in available bandwidth over a short, pre-determined period of time.
Secondly, using both video bitrate adaptation and multi-homing can greatly improve the live video streaming application’s user QoE by avoiding congestion that results in excessive delay or by providing high video bitrates through link aggregation. In this thesis, we present Conflux: a modular, multi-homed, adaptive video bitrate protocol for live video streaming. Conflux uses a probabilistic link quality model that is used in conjunction with a user-specific utility function to determine the video bitrate and the rate at which to send on each link. By using a simple, yet general, probability-based link quality model, Conflux can easily support different requirements and environments such as the number of links by just maximizing the expected utility. We evaluate Conflux in an emulated network environment where the available bandwidth comes from variety of wireless network traces, and the maximum available bandwidth is 25Mbps. Our evaluation shows that Conflux can obtain at least an 18% improvement when there are two available links, and 65% when there are five available links over its non-optimal, multi-homed comparison systems that have oracle-based knowledge of either each link’s available bandwidth or the total aggregate available bandwidth.
Finally, resending delayed data on alternate, non-congested paths and sending redundant data using Forward Error Correction (FEC) can enable a video frame to arrive on time in the presence of deteriorating link quality or link failures. However, determining the degree of redundancy while effectively making use of the available bandwidth so that users can have high QoE is challenging. In this thesis, we introduce extensions to Conflux that support retransmission and FEC. We present our method of calculating the probability of on-time video data arrival for a given redundancy level using Conflux’s probabilistic link quality model. This allows Conflux to select the degree of redundancy and corresponding video bitrate that maximizes the user’s expected utility. Our evaluation shows that using Conflux’s redundancy-specific user utility functions lowers the percentage of incomplete frames 11% to 0.02% in network environments that experience packet loss according to the Gilbert-Elliot loss model.
To attend this PhD defence in person, please go to DC 2314. You can also attend virtually on MS Teams.