Measuring YouTube from Dual-Stacked Hosts Appendix from - PowerPoint PPT Presentation

Measuring YouTube Google Global Caches March 2015 Jacobs University Bremen, Bremen, Germany Computer Networks and Distributed Systems Group, Aalto University, Espoo, Finland School of Electrical Engineering, Jürgen Schönwälder Jörg Ott Vaibhav Bajpai Saba Ahsan Passive and Active Measurement (PAM) Conference, New York Measuring YouTube from Dual-Stacked Hosts Appendix from Dual-Stacked Conclusion Tiroughput and Stall Events Happy Eyeballs Hosts Introduction Motivation Research Question Research Contributions Methodology Metrics Measurement Setup Measurement Setup Measurement Trials Data Analysis Speed Tests TCP Connect Times 1 / 25 v.bajpai@jacobs-university.de saba.ahsan@aalto.fi jorg.ott@aalto.fi j.schoenwaelder@jacobs-university.de Leone FP7 EU Project: leone-project.eu

Measuring YouTube Speed Tests 1 Comcast, Deutsche Telekom AG, AT&T, Verizon Wireless, T-Mobile USA from Dual-Stacked Introduction | Motivation Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs TCP Connect Times 2 / 25 Data Analysis Hosts Measurement Setup Measurement Setup Metrics Methodology Research Contributions Research Question Motivation Measurement Trials Introduction ▶ Large IPv6 broadband rollouts 1 since World IPv6 Launch Day in 2012. ▶ Increased global adoption of IPv6 to 6% (as seen by Google as of March 2015). ▶ Studies show how YouTube contributes heavily to volumes of IPv6 traffjc [1]: 100 100 other Percent of volume 80 unknown edu Percent of volume 80 gov well − known ftp − data 60 Facebook 60 ssh open source 40 Google ssl 40 YouTube dns 20 http 20 0 0 − 6 − 5 − 4 − 3 − 2 − 1 W6D + 1 + 2 + 9 +10 +11 +12 +16 +17 +18 +19 +20 +21 +22 − 6 − 5 − 4 − 3 − 2 − 1 W6D + 1 + 2 + 9 +10 +11 +12 +16 +17 +18 +19 +20 +21 +22 Fig. 7. Application mix per day for all IPv6 traffic (campus). Fig. 8. Daily HTTP mix (campus).

Measuring YouTube Speed Tests streaming YouTube videos over IPv6? Do users experience benefjt (or an added penalty) when Introduction | Research Question Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs TCP Connect Times Data Analysis from Dual-Stacked Measurement Trials Measurement Setup Measurement Setup Metrics Methodology Research Contributions Research Question Motivation Introduction Hosts 3 / 25

Measuring YouTube TCP Connect Times over IPv4 and IPv6 from dual-stacked networks. To the best of our knowledge, this is the fjrst study to compare YouTube performance 3. YouTube content caches over IPv6 are largely absent. 2. Lower throughput is achieved when streaming YouTube over IPv6. Introduction | Research Contributions Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs Speed Tests from Dual-Stacked Data Analysis Measurement Trials Measurement Setup Measurement Setup Metrics Methodology Research Contributions Research Question Motivation Introduction Hosts 4 / 25 1. TCP connect times to YouTube makes Happy Eyeballs [2] prefer IPv6.

Measuring YouTube Data Analysis Methodology Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs TCP Connect Times Speed Tests Measurement Trials from Dual-Stacked Measurement Setup Measurement Setup Metrics Methodology Research Contributions Research Question Motivation Introduction Hosts 5 / 25

Measuring YouTube Happy Eyeballs 2 Tie YouTube test supports three container formats: MP4, WebM and FLV 6. Extracts frame timestamps from container to mimic a playout. 5. Fetches audio and video streams over concurrent HTTP sessions. 4. Establishes concurrent TCP connections for audio and video streams. 3. Locally resolves DNS names of media server locations. 2. Extracts container formats 2 , available resolutions and media server locations. 1. Takes a YouTube URL as input and scrapes the downloaded HTML page. Methodology | YouTube Performance Test Appendix from Dual-Stacked Google Global Caches Tiroughput and Stall Events Conclusion TCP Connect Times Research Contributions Hosts Introduction Motivation Speed Tests Research Question 6 / 25 Methodology Metrics Measurement Setup Measurement Setup Measurement Trials Data Analysis ▶ Measures TCP connect times by recording connect(…) call completion time. ▶ DNS resolution time is not accounted. ▶ Ensures temporal synchronization between audio and video streams. ▶ Measures throughput achieved over the single TCP connection for each stream. ▶ A 2 second prebufgering is applied before starting playout timer. ▶ Measures stall duration whenever a frame fails to arrive before its playout time. ▶ A stall triggers 1 second of rebufgering before resuming playout timer.

Measuring YouTube Data Analysis Methodology | Speed Test Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs from Dual-Stacked Speed Tests TCP Connect Times Measurement Trials Research Question Hosts Measurement Setup Motivation Introduction Research Contributions Methodology Metrics Measurement Setup 7 / 25 ▶ Measures achievable throughput over the line. ▶ Uses 3 simultaneous TCP connections to fetch 1 GB, non-zero, binary fjle. ▶ HTTP GET request is made to the nearest (based on latency) M-Lab server. ▶ Detailed in the SamKnows test suite [3] description. ▶ We modifjed the test to also enable measurements over IPv6. ▶ We use these line rates to baseline our YouTube throughput measurements.

Measuring YouTube from Dual-Stacked 3 https://developers.google.com/youtube/v3/docs/videos/list 3. Video has no regional restrictions. 2. Video is available in Full HD format. 1. Video duration > 60s. Methodology | Selection of YouTube Videos Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs TCP Connect Times Speed Tests Data Analysis Measurement Trials Measurement Setup Measurement Setup Metrics Methodology Research Contributions Research Question Motivation Introduction Hosts 8 / 25 ▶ We use the YouTube v3 API 3 to prepare a list of globally popular videos where: ▶ List is refreshed every 12h on the SamKnows backend. ▶ Each probe pulls this list on a daily basis.

Measuring YouTube Speed Tests 2. Step-down mode. 1. Non-adaptive mode. Methodology | Selection of Video Bitrate Appendix Conclusion Google Global Caches from Dual-Stacked Happy Eyeballs TCP Connect Times Tiroughput and Stall Events Data Analysis Research Question Hosts Introduction Measurement Trials Motivation 9 / 25 Research Contributions Methodology Metrics Measurement Setup Measurement Setup ▶ YouTube provides a list of available resolutions (and their required bitrates). ▶ Tie YouTube test currently does not support DASH [4]. ▶ We use speed test results to limit maximum bitrate. ▶ We also support 2 operation modes: ▶ Test downloads the same video resolution despite stalls. ▶ Although, does not mimic the behavior of YouTube players. ▶ However, still useful to compare IPv4 vs IPv6 performance in identical conditions. ▶ Test steps down to a lower video resolution on a stall. ▶ Portrays a more user-oriented behaviour.

Measuring YouTube Data Analysis from Dual-Stacked Methodology | Measurement Setup Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs TCP Connect Times Speed Tests 10 / 25 Measurement Trials Hosts Measurement Setup Metrics Methodology Research Contributions Research Question Motivation Introduction Measurement Setup 1.HTTP(S) GET YouTube Servers Probe 2.Media URLs SamKnows YouTube client DSL/Cable Tests Modem results 3.Audio Stream YouTube Media 3.Video Stream Servers 4.HTTPS POST SamKnows Backend ▶ YouTube test runs every hour (once for IPv4 and subsequently for IPv6). ▶ Speed test runs every 6 hours (once for IPv4 and subsequently for IPv6).

Measuring YouTube Data Analysis from Dual-Stacked Methodology | Measurement Trials Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs TCP Connect Times Speed Tests 11 / 25 Measurement Trials Research Question Measurement Setup Introduction Motivation Hosts Research Contributions Methodology Measurement Setup Metrics # LOCATION PROVIDER TYPE #01 BREMEN KABELDEUTSCHLAND RESIDENTIAL #02 BREMEN DEUTSCHE TELEKOM RESIDENTIAL #03 STOCKHOLM SITAB RESIDENTIAL #04 FUKUOKA ASAHI NET RESIDENTIAL #05 MADRID JAZZ TELECOM RESIDENTIAL #06 ALLEUR EDPNET RESIDENTIAL #07 BREMEN DEUTSCHE TELEKOM RESIDENTIAL #08 SHIZUOKA BIGLOBE NEC RESIDENTIAL #09 CERN CERN RESEARCH #10 BREMEN DFN NREN #11 TIMISOARA ROEDUNET NREN #12 LOUVAIN BELNET NREN #13 BREMEN DFN NREN #14 HELSINKI FUNET NREN #15 LONDON BSKYB-BROADBAND LAB #16 TORINO TELECOM ITALIA LAB #17 MADRID BT ESPANA LAB #18 IPSWICH BT UK LAB #19 NIIGATA NDAC IXP #20 BRAUNSCHWEIG GAERTNER DATENSYSTEME BUSINESS #21 OLTEN INIT SEVEN BUSINESS

Measuring YouTube Data Analysis 4 Tie results are derived from measurements conducted for 20 days in September 2014. Data Analysis 4 Appendix Conclusion Google Global Caches Tiroughput and Stall Events Happy Eyeballs TCP Connect Times Speed Tests Measurement Trials from Dual-Stacked Measurement Setup Measurement Setup Metrics Methodology Research Contributions Research Question Motivation Introduction Hosts 12 / 25