Delving into Internet Streaming Media Delivery A quality

Delving into Internet Streaming Media Delivery: A quality and Resource Utilization Perspective Zhen Xiao Joint work with Lei Guo, Enhua Tan, Songqing Chen, Oliver Spatchcheck, and Xiaodong Zhang ACM SIGCOMM Internet Measurement Conference (IMC'06), October 2006 1

Multimedia on the Internet • Education and research • News media • Entertainment and gaming • Advertisement 2

Streaming Media Streaming server CDN/MDN 3

Pseudo Streaming Web server http: //www. You. Tube. com/ http: //video. google. com/ HTTP meta file 4

Streaming Media • Thousands of concurrent streams • Flexible response to network congestion • Efficient bandwidth utilization • High quality to end users • Challenges and techniques 5

Existing Measurements • Access pattern and user behaviors – A bunch of measurement studies – Server clusters, media proxies • Streaming mechanism and delivery quality – Few studies • Traffic volume … – Downloading > pseudo streaming > streaming (WWW’ 05, cookie talk 2005) – P 2 P >> all other media delivery systems 6

Our Measurement • Investigate modern streaming services – The delivery quality and resource utilization • Collect a large streaming media workload – From thousands of home users and business users – Hosted by a large ISP – Packet level instead of server logs • Analyze commonly used streaming techniques – Automatic protocol switch – Fast Streaming – MBR encoding and rate adaptation 7

Outline • Traffic overview • Protocol rollover • Fast Streaming • Rate adaptation • Conclusion 8

Traffic Overview • User communities – Home user – Business user • Media hosting services – Self-hosting – Third-party hosting 9

Number of requests Business users access more audio than home users 10

On-demand media: File length Audio music previews Video pop songs Business users tend to access longer audio/video files 11

On-demand media: Playback duration Audio Video pop songs music previews Business users tend to play audio/video longer 12

Live media: Playback duration Audio Video Business users tend to access live audio/video longer 13

Traffic Overview • User communities – Home user News and entertainment sites – Business user – Working environment affects access pattern • Media hosting services – Self-hosting – Third-party hosting 14

Traffic Overview • User communities – Business users tend to access streaming media longer than home users – Working environment affects access pattern • Media hosting services – Self-hosting – Third-party hosting 15

Media hosting services 16

Outline • Traffic overview • Protocol rollover • Fast Streaming • Rate adaptation • Conclusion 17

Protocol Rollover Streaming server Media player RTSP/UDP RTSP/TCP HTTP/TCP Embed RTSP commands in HTTP packets Traffic volume: UDP: 23% TCP: 77% HTTP: rare 18

Protocol rollover time Startup latency = protocol rollover time + transport setup time + startup buffering time Windows media service Real. Networks media service Protocol rollover increases user startup time significantly 19

Protocol selection and rollover avoidance • Most streaming traffic are TCP-based – The usage of NAT? – MMS clients report private IP address in clear text • Home user: 98. 3% report 192. 168. *. * • Business user: 89. 5% report 192. 168. *. * • Protocol rollover sessions are minor – Home user: 7. 37% – Business user: 7. 95% • Most streaming sessions use TCP directly – Why? 20

Protocol selection and rollover avoidance • Windows media service – Specify the protocol in the media meta file Use URL modifiers to avoid protocol rollover Ex: rtspt: //xxx. com: /xxx. wmv – More than 70% • Real. Networks media service – NAT transversal techniques 21

Outline • Traffic overview • Protocol rollover • Fast Streaming • Rate adaptation • Conclusion 22

Fast Streaming • Fast Streaming: deliver media data “faster” than its encoding rate – – Fast start Fast cache Fast recovery Fast reconnect • Always TCP-based 23

Media objects delivered with Fast Cache (Vo. D home user workload) File length Encoding rate Fast Cache is more widely used for media files with longer length and higher encoding rate. 24

Bandwidth Utilization PLAY RTSP/1. 0 Bandwidth: 1. 12 Mbps Speed: 20. 5 RTSP /1. 0 200 OK Speed: 5 Fast Cache Normal TCP 25

Fast Cache smooth bandwidth fluctuation Rebuffer ratio = rebuffer time / play time Fast Cache Normal TCP 26

Fast Cache produces extra traffic Most streaming sessions only request the initial part of a media object Over supplied data Fast Cache: 55% Normal TCP: 5% 27

Server response time DESCRIBE foo. wmv RTSP/1. 0 SRT RTSP /1. 0 200 OK SDP Third party media service sniffer Self-hosting media service 28

Server Load Windows Server 2003 Win XP Windows media load simulator Ethernet … Server log 1 X 4 X Some CDNs/MDNs do not support Fast Cache at all 29

Outline • Traffic overview • Protocol rollover • Fast Streaming • Rate adaptation • Conclusion 30

Rate Adaptation Multiple-bit-rate encoding 96 Kbps 128 Kbps 320 Kbps … 1. 128 Mbps Stream switch WM: Intelligent streaming RM: Sure. Stream thinning: deliver key frame only Video cancellation 31

MBR encoding on-demand audio live audio stream in video objects video stream in video objects 42% on-demand video are MBR encoded 32

Stream switch Play-out buffer Streaming switch latency 40% Low quality duration 30 sec 60% 3 sec Stream switch is often not smooth 33

Stream thinning 30 sec 70% 34

Fast Cache and stream switch Do not work with each other: fewer stream switches than MBR encoded objects When network congestion occurs … fill play-out buffer playing buffering 5 sec playing buffering time Like pseudo streaming When rebuffer occurs 35

Streaming quality and playback duration Home user business user >100 sec 88% Longer duration sessions have higher prob. of quality degradation Business user workload has more quality degradation 36

Streaming quality summary The quality of media streaming on the Internet leaves much to be improved 37

Coordinating caching and rate adaptation • Fast Cache: aggressively buffer data in advance – Over-utilize CPU and bandwidth resources – Neither performance effective nor cost-efficient • Rate adaptation: conservatively switch to lower bit rate stream – Switch handoff latency • Coordinated Streaming Lower bound Prevent switch latency Upper bound Prevent aggressive buffering 38

Coordinated Streaming Rebuffering ratio Over-supplied data Switch latency 39

Conclusion • Quality of Internet streaming – Often unsatisfactory – Need to improve • Modern streaming media services – Over-utilize CPU and bandwidth resources – Not a desirable way to improve quality • Coordinated Streaming – Combine merits of both caching and rate adaptation – Simple but effective 40