Summary

QUIC不及TCP在DASH的表現，由於大多數ABR algorighm都經過優化與TCP配合使用，因此沒有充分利用QUIC提供的功能。

Problem statement

  • What is the impact of QUIC on QoE?
  • How should adaptive bitrate streaming be built to leverage the benefits provided by QUIC?

Research objectives

  • A Performance Study and future work

Methods

  • we studied QoE performance of different DASH quality adaptation algorithms that are either solely based on the playout buffer filling, or on the video segment download rate, or on both.
  • DASH quality adaptation algorithms that are mainly categorized as throughputand/or playout buffer-based techniques.

BB2-2: based on buffer. map current buffer occupancy to a quality bitrate.

the client tends to download a quality that is higher than the measured rate in order to provide the user with maximum average quality.

SQUAD: based on rate that is specifically optimized for variations in TCP download rates. BOLA: based on both.

Evaluation

How does the author appraise their methodology?

  1. Testbed (Parallel Server / Single Server) bottleneck link of 10Mbps / Python-based DASH player / Caddy server
  2. Internet (wifi on campus / wifi on residential / wired campus Network ) with 17 bitrates ranging from 100Kbps on Aamazon EC2

How is the experiment set up?

  1. Average Quality Bitrate (AQB) = Chosen Rate

  2. Number of Quality Switches (#QS)

  3. Spectrum (H): A lower H indicates a better QoE. (bitrate變化的集中度量)

  4. Rebuffering Ratio:

    $$ \frac {Total;length - video;length } {video;length } $$

  • We investigated different QoE performance metrics such as the average quality bitrate, a measure of quality variations (denoted spectrum) (影片質量差異-頻譜) and the average video stalling duration.

Testbed

image-20210504142314652

image-20210504141515242

BBA-2 seems to benefit from using QUIC to download segments because of the slow-start or initial phase of the algorithm.

the client tends to download a quality that is higher than the measured rate in order to provide the user with maximum average quality. 只有較大的RTT時，會造成buffer不足，間接使得BBA-2 client切換到lower quality.

(rule-of-thumb)bandwidth-delay product(BDP): B = C * RTT (receive buffer收到的大小，已傳送但尚未確認的資料)

Internet

image-20210504152404870

  1. Wifi significantly worse quality bitrate: QUIC is likely to be competing with more TCP streams than other QUIC streams, it takes longer than TCP to fetch the same segments

Future Work

  1. HOL blocking (Multiplexing,Multi-Path)

  2. Pacing

    1. If a client requests multiple qualities for a single segment, the server could pace the streams to deliver segments at regular intervals instead of using AIMD. 如果單個網段提供多種質量，定期時間下，以stream的速度下去交付segments而不是AIMD

    2. If a client requests multiple segments over the same connection via multiple streams the server could implement a decreasing pacing rate for segments requested depending on how soon they are required for playback. client如果在多個stream中請求多個segment，Server可以根據「請求的segment」需要多長時間來降低pacing rate

      AIMD -> TCP擁塞控制的方法

    3. Eliminate Congestion Control Redundancy two congestion on client; in trans/ in DASH algo

      ABR clients can disable congestion control in the QUIC transport layer and continue to use their existing congestion control

      QUIC provides the benefits of pacing and assistance for loss recovery in the form of NACKs

      RTP does that.

Conclusion

  • we find through testbed and Internet measurements that QUIC does not provide a boost to current DASH algorithms but instead a degradation in the chosen quality bitrates.
  • Although we observe a lower magnitude of quality variations (較低的質量變化幅度), the degradation of streamed quality bitrate with the use of QUIC is detrimental to overall QoE

References

ABR work/ congestion control / congestion avoidance

[14]: Tcp hollywood: An unordered, time-lined, tcp for networked multimedia applications (2016)

TCP Hollywood,a TCP variant, which implements out-of-order delivery and inconsistent retransmissions in order to improve good-put of video streaming applications

[21]: Media qoe enhancement with quic (2016)

a new congestion control mechanism using QUIC that aggressively varies download rate according to a buffer-based priority level assigned by the ABR streaming client

[4]: ABR protocol over UDP (2003)

the authors employ a form of congestion avoidance where the sending rate at the server is increased by a single packet for every RTT measurement.This design is different from the AIMD congestion control employed by TCP and QUIC since it eliminates the effect of slow start and attempts to provide an accurate estimate of the available bandwidth in the network. Some drawbacks of this approach are the requirement of two UDP sockets for every connection and the use of Berkeley Packet Filters to collect timestamps at the server and client for every video stream, thus, reducing both performance and scalability of the system