Quality-aware Media Streaming Optimization (based on Pensieve)[slide]

Table of Contents

• Quality-aware Media Streaming Optimization (based on Pensieve)[slide]
  • Table of Contents
  • Objectives
  • Setup
  • Apache Server
    • Linux
      • Installation & Basic Setup
      • Change Port Number
    • Windows
      • Installation & Basic Setup
    • Mac
      • Installation & Basic Setup
      • Change Port Number
  • Chromedriver & Selenium
  • Addition of 9-level quality video
  • Version Changes compared to original pensieve
  • Additional changes compared to original pensieve (especially on frontend)
  • Real-world experiments
  • Training Results
  • Future Work
  • Dependency
• Reference

Objectives

• Recreate the real-world experiments
• Customize the adaptation algorithms for different needs (smooth, balanced, high-quality)
• Optimize the current ABR algorithm (based on RL) so that it could be applied in HLS (HTTP Live Streaming) (desired ABR algorithm could enable the video to have smooth playback while retaining a much smaller buffer level than playback environment)
• Gather a much larger dataset (FCC, Belgium, Oboe, Waterloo Streaming QoE Dataset)

Setup

Apache Server

Linux

Installation & Basic Setup

• update package local index

sudo apt update

• install apache2

sudo apt install apache2

• start server

sudo service start apache2

• test that the server is working properly by entering localhost:80 in the browser and it should show the apache default webpage.

Change Port Number

• use vim to open ports.conf

vim /etc/apache2/ports.conf

• Change Listen 80 to Listen [port number]
• restart the server

sudo service restart apache2

• test that the server is working properly by entering localhost:[port number] in the browser and it should show the apache default webpage.

Windows

Installation & Basic Setup

• download Apache Lounge
• download Apache Haus and unzip the file under specific disk location
• more details can be viewed in windows-pensieve/

Mac

Installation & Basic Setup

• stop running (older) Apache server and remove it

sudo apachectl stop
sudo launchctl unload -w /System/Library/LaunchDaemons/org.apache.httpd.plist

• install httpd via Homebrew

brew install httpd

• start server

sudo brew services start httpd

• test that the server is working properly by entering localhost:8080 in the browser and it should show the apache default webpage.

Change Port Number

• use vim to open httpd.conf

vim /etc/apache2/httpd.conf

• change Listen 8080 to Listen [port number]
• test that the server is working properly by entering localhost:[port number] in the browser and it should show the apache default webpage.

Chromedriver & Selenium

• Detailed steps are provided in setup.py
• make sure that Google Chrome is already installed before doing the following installation and setup.
• check Google Chrome version
• download corresponding chromedriver and put it under abr_browser_dir/

Addition of 9-level quality video

In addition to 6-level quality video already provided under video_server, we have added the other 9-level quality video. The manifest can be downloaded here

https://dash.akamaized.net/akamai/bbb_30fps/bbb_30fps.mpd

and video segments can be downloaded here

https://dash.akamaized.net/akamai/bbb_30fps/bbb_30fps.zip

The following table shows corresponding changes.

Item Description	6-level Quality Video File Name	9-level Quality Video File Name
get video sizes	sim/get_video_sizes.py	sim/get_video_sizes_bbb_30fps.py
video sizes	sim/video_size_[index] (index ranges from 0 to 5)	sim/video_size_bbb_30fps_[index] (index ranges from 0 to 9)
simulation folder	sim	sim_bbb_30fps
server folder	rl_server/	rl_server_bbb_30fps/

Version Changes compared to original pensieve

• As pensieve is originally written in Python 2.7, the following changes are mainly migrations to Python 3.7.
  • add parantheses around print statements
  • all the open commands are changed from wb (or rb) to w (or r)
• Under rl_server/ and rl_server_bbb_30fps/, programs used to build backend server for transfer of state information include

File Name	Supported Algorithm
dash_server_original.py
mpc_server.py	MPC
robust_mpc_server.py	robustMPC
rl_server_no_training.py	A3C (RL)
simple_server.py

• The following changes should be applied in all files above

Location	Original	Current
first two lines	from BaseHTTPServer import BaseHTTPRequestHandler, HTTPServer import SocketServer	from http.server import BaseHTTPRequestHandler, HTTPServer import socketserver
server_address	('localhost', port)	('', port)
self.wfile.write	[string]	[srting].encode()

Additional changes compared to original pensieve (especially on frontend)

• In order to adapt to various needs of users (smooth, high-quality, balanced), we add a vertical slider that can change 3 different reward schemes in-time.

• we refurbished the frontend for comparison between RL algorithms and traditional algorithms. Below are the screenshots.

  

  

• As we need to simultaneously play two videos, we need to start two servers for file transfer. Therefore, we need to change port number. In this case, we used BOLA for comparison, and hence we change simple_server to port 8334 and change the port number in dash.all.min.js correspondingly.

• Caution: it is preferred (and a better practice) that JavaScript changes should be first made in original source files and then use grunt dist to update dash.all.min.js. But for simplicity and convenience, we have directly updated the changes in dash.all.min.js.

Real-world experiments

• set up the server
• copy web application files to Apache Server Document Root folder

cp -r dash.js/ /usr/local/var/www/

• restart server

sudo brew services start httpd

• run RL server

python rl_server_bbb_30fps/rl_server_no_training.py

• run simple server

python rl_server_bbb_30fps/simple_sever.py

• launch reference client by localhost:[port number]/dash.js/samples/dash-if-reference-player/deecamp.html

Training Results

• The following are training results from original pensieve.

All of 3 reward schemes show very high instability, and this is partially reflected in the frequent changes of bitrates by the video client.

Future Work

• As described in Training Results, the training curves oscillate significantly. Therefore, this issue suggests variance reduction, detailed in Variance Reduction for Reinforcement Learning in Input-Driven Environments.
• Create a model ensemble consisting of other RL models, including PPO, VPG, DQN and DQN-based framework D-DASH.

Dependency

• Python 3.7
• Tensorflow 1.9

Reference

• pensieve: https://github.com/hongzimao/pensieve/
• Ubuntu Apache Setup: https://www.digitalocean.com/community/tutorials/how-to-install-the-apache-web-server-on-ubuntu-18-04-quickstart
• Mac Apache Setup: https://tecadmin.net/install-apache-macos-homebrew/
• Windows Apache Setup (zh-CN)
  • Installation & Setup of VC14 and Apache Lounge: https://blog.csdn.net/liyang4534/article/details/78036591
  • Installation & Setup of Apache Haus: https://blog.csdn.net/weixin_43738701/article/details/86607148
• Installation & Setup of Chomedriver and Selenium on Windows: https://blog.csdn.net/Booboochen/article/details/80531155