Overview

• Video streaming dominates internet traffic

• Adaptive bitrate (ABR) is a key algorithm to optimize user's quality of experiences (QoE)

  • Decides the quality level of each video chunk to send

  • Primary goals: higher video quality, fewer stalls

    • Larger chunk: increase video quality, but can stall the video playback

  • Prior works:

    • Buffer-based: pick bitrate based on buffer occupancy

      • BBA [SIGCOMM '14]: simple buffer based ABR algorithm. Select video bitrate based only on user's playback buffer: below some buffer level, selects the lowest bitrate; above, select the highest one. In-between, linear function

    • Rate-based: pick bitrate based on predicted throughput

      • FESTIVE, PANDA, CS2P

    • Hybrid: use both throughput prediction and buffer occupancy

      • MPC [SIGCOMM '15]: model predictive control, uses harmonic mean of past throughput to predict future throughput, scale the predictive throughput linearly with the chunk size

        • Not good at throughput

        • Predicting throughput actually varying with chunk size due to congestion control and varying bandwidth

        • Network congestion occurs when a node or link carries data beyond its limit. This often leads to the queuing of packets—and in the worst case, loss of packets—as well as a decrease in the network’s Quality of Service (QoS). Congestion Control is a mechanism that controls the entry of data packets into the network, enabling a better use of a shared network infrastructure and avoiding congestive collapse.

    • Machine learning to tackle this problem

      • Pensieve [SIGCOMM '17]

        • Reinforcement learning to learn an end-to-end ABR control

          • States: past chunk throughput, download time, next chunk size, current buffer size, remaining chunks, past chunk bit rate

          • Reward: bitrate, rebuffering, smoothness

        • Requires network simulators as training environments, assume training in simulation generalizes to wild Internet

        • Requires a set of training environments that can exercise a control policy through a range of situations and actions

      • Fugu [NSDI '20]

        • Based on model predictive control

        • Transmission time predictor (TTP) (NN): how long it will take for a client to receive a given chunk

          • Input: size and transmission time, size of chunk, low-level TCP statistics (RTT, SWND)

          • Output: probability distribution over transmission time

        • Training in situ: supervised learning in situ on real data from deployment environment

      • SENSEI [NSDI 21']: incorporate variability of quality sensitivity using sensitivity weights obtained via per-video crowdsourcing (quality sensitivity varies as video content changes)

    • Other works: machine-centric video streaming

      • Maximizes for DNN inference accuracy

      • Most: source-driven approach, but suboptimal for analytics-oriented applications

      • DNN-Driven approach [Server-Driven Video Streaming for Deep Learning Inference [SIGCOMM 20']]: sends low-quality video stream to the server, the server runs the DNN to determine where to re-send with higher quality to increase the inference accuracy

    • Other works: continuous learning of video analytics models on edge compute servers

      • In each retraining window, decide which of the edge models to train; allocate edge server's GPU resources among the retraining and inference jobs; select configurations of the retraining and inference jobs

      • Maximize the inference accuracy averaged over the retrained window

      • Contributions: introduce the metric, design efficient profiler, resource scheduler

• What does it take to create a learned ABR algorithm that robustly performs well over the wild internet?