Summary

在QUIC中，速度快的video stream被比較慢的audio stream給queue住(sharing a single UDP socket buffer)，造成audio request response latency拉長。其上層的DASH計算throughput不考慮response latency，造成實際的throughput跟預期的DASH不同。

Problem statement

  1. 2018 (Not so QUIC): focused on buffer-based ABR

these recent studies have indicated that buffer-based techniques are aggressive towards video-bitrate maximization whereas suffers in terms of playback smoothness and rebuffering

  1. 2020 (Does QUIC Suit): focused on modern ABR

We investigate further to understand the protocol-level behavior of QUIC, which impacts the QoE performance

ABR’s important metrics: predicted throughput during video streaming.

Research objectives

  • Explore the performance of advanced ABR techniques(MPC,Pensieve)

    MPC-Fast/MPC-Robust -> predictive control Pensieve -> Deep Learning

Methods (EXPERIMENTAL SETUP)

computed 3 QoE metrics

  1. average playback bitrate
  2. total rebuffering duration
  3. playback smoothness

image-20210511143913279

a linear representation q(Rn)= Rn,similar to [4], indicating that the playback quality increases linearly with the increase of playback bitrate.

server: lsquic,Go-QUIC. client: google chrome (we modify dash.js to add support for advance ABR algorithms like Pensieve and MPC)

To emulate realistic traffic behavior:

a Web-based javascript DASH player provided by DASH Industry Foundations (DASHIF) to stream the videos at the client side.

  1. Mahimahi: benchmark traffic shaper(Accurate Record-and-Replay for HTTP)，用于記錄來自基于HTTP應用程序的流量，並針對模擬的網路環境進行模擬。
  2. Public FCC dataset(模擬網路中dataset，compatible from Mahimahi): a broadband trace from FCC.

Evaluation

比較QOE

image-20210511144058340

  1. Average Video Bitrate: buffer-based ABR mechanisms aggressively use the highest quality levels.
  2. Playback Smoothness:
    1. A fluctuation in the quality level indicates less smoothness in the video playback, and, therefore, reduces the QoE.
    2. Advanced ABR techniques, such as MPC-Fast and Pensieve, provide better playback smoothness with DASH/QUIC, although the supported playback quality is lower compared to DASH/TCP.
  3. Rebuffering Time: based on which ABR technique is adopted.

比較應用場景

if the response latency is high, this segment may take longer time to reach the client, resulting in a rebuffering

?

The Throughput: the download time = the first and the last bytes received 兩者時間差. The Response Latency: initiation of the HTTP request and the time when the first byte of the response is received 兩者時間差 ( 收到的第一個ACK的時間 - 發起HTTP請求的時間 )

why high response latency observed during the video streaming using QUIC?

在create two parallel HTTP requests情況下，generate two HTTP 相互依賴的stream，結果與上述差不多。 key:socket buffers between HTTP streams.在QUIC中，速度快的stream被比較慢的stream給queue住了。

  • TCP creates two separate sockets for the two HTTP streams, each of the sockets maintains its own socket buffer. (Independent)
  • QUIC multiplexes both the streams and uses a single UDP socket having a single socket buffer, the HTTP responses from both the streams interfere, and higher response rate at one stream affects the queuing delay for the response at the other stream (dependent)

在傳統DASH中，video的data rate比audio更高(video data > audio data)，TCP separate socket中的queue depending on their data generation rate.而對於QUIC來說，每個需要播放的segment都需要client發一個http request for video及一個http request for audio，由於video segemnt request先送，UDP socket buffer被video segement塞滿，audio segment必須等video segement在socket buffer中被釋放。

image-20210512155755783

(a) The audio data has to wait in the queue (the red timeline) before it gets served. (b) The audio streams at QUIC experiences a much higher latency compared to TCP. Audio Stream request時間較久 = response latency is higher

Conclusion

  • The QUIC multiplexing of audio and video streams over a single UDP socket results in additional response latency for the audio segments, which are not captured during the calculation of channel throughput. As a consequence, the ABR algorithms take incorrect decisions during selecting the bitrates based on the calculated throughput over a QUIC connection

Note

​ #Anything additional that is not included in this outline

References

​ #Relevant articles that might be useful to look at and research on