1. Lecture 6: Video Streaming 2-1

2. Outline  Network basics:  HTTP protocols  Studies on HTTP performance from different views:  Browser types [NSDI 2014]  Search Provider [SIGCOMM 2013]  Cellular Networks [IMC 2012],[MOBICOM 2014]  Home [IMC 2013]  Network Measurement fundamentals:  Decision Tree (one popular machine learning tool) 2-2

3. Outline  Video Streaming:  Basic Protocols & Tutorials.  Measurement studies:  Impact of video quality on user engagement [SIGCOMM2011]  Structure of Internet Video Quality Problems in the Wild [IMC2013]  Predicative Models  Predictive Model of Quality of Experience [SIGCOMM2013]  Impact of cellular network dynamics on video quality [SIGMETRICS2014] 2-3

4. Outline  Network basics:  HTTP protocols  Studies on HTTP performance from different views:  Browser types [NSDI 2014]  Search Provider [SIGCOMM 2013]  Cellular Networks [IMC 2012],[MOBICOM 2014]  Home [IMC 2013]  Network Measurement fundamentals:  Decision Tree (one popular machine learning tool) 2-4

5. Application Layer 2-5 DNS: domain name system people: many identifiers:  SSN, name, passport # Internet hosts, routers:  IP address (32 bit) - used for addressing datagrams  “name”, e.g., www.yahoo.com - used by humans Q: how to map between IP address and name, and vice versa ? Domain Name System:  distributed database implemented in hierarchy of many name servers  application-layer protocol: hosts, name servers communicate to resolve names (address/name translation)  note: core Internet function, implemented as application- layer protocol  complexity at network’s “edge”

6. Application Layer 2-6 DNS: services, structure why not centralize DNS?  single point of failure  traffic volume  distant centralized database  maintenance DNS services  hostname to IP address translation  host aliasing  canonical, alias names  mail server aliasing  load distribution  replicated Web servers: many IP addresses correspond to one name A: doesn’t scale!

7. Application Layer 2-7 Root DNS Servers com DNS servers org DNS serversedu DNS servers poly.edu DNS servers umass.edu DNS servers yahoo.com DNS servers amazon.com DNS servers pbs.org DNS servers DNS: a distributed, hierarchical database client wants IP for www.amazon.com; 1 st approx:  client queries root server to find com DNS server  client queries.com DNS server to get amazon.com DNS server  client queries amazon.com DNS server to get IP address for www.amazon.com … …

8. Application Layer 2-8 DNS: root name servers  contacted by local name server that can not resolve name  root name server:  contacts authoritative name server if name mapping not known  gets mapping  returns mapping to local name server 13 root name “servers” worldwide a. Verisign, Los Angeles CA (5 other sites) b. USC-ISI Marina del Rey, CA l. ICANN Los Angeles, CA (41 other sites) e. NASA Mt View, CA f. Internet Software C. Palo Alto, CA (and 48 other sites) i. Netnod, Stockholm (37 other sites) k. RIPE London (17 other sites) m. WIDE Tokyo (5 other sites) c. Cogent, Herndon, VA (5 other sites) d. U Maryland College Park, MD h. ARL Aberdeen, MD j. Verisign, Dulles VA (69 other sites ) g. US DoD Columbus, OH (5 other sites)

9. Application Layer 2-9 TLD, authoritative servers top-level domain (TLD) servers:  responsible for com, org, net, edu, aero, jobs, museums, and all top-level country domains, e.g.: uk, fr, ca, jp  Network Solutions maintains servers for.com TLD  Educause for.edu TLD authoritative DNS servers:  organization’s own DNS server(s), providing authoritative hostname to IP mappings for organization’s named hosts  can be maintained by organization or service provider

10. Application Layer 2-10 Local DNS name server  does not strictly belong to hierarchy  each ISP (residential ISP, company, university) has one  also called “default name server”  when host makes DNS query, query is sent to its local DNS server  has local cache of recent name-to-address translation pairs (but may be out of date!)  acts as proxy, forwards query into hierarchy

11. Application Layer 2-11 requesting host cis.poly.edu gaia.cs.umass.edu root DNS server local DNS server dns.poly.edu 1 2 3 4 5 6 authoritative DNS server dns.cs.umass.edu 7 8 TLD DNS server DNS name resolution example  host at cis.poly.edu wants IP address for gaia.cs.umass.edu iterated query:  contacted server replies with name of server to contact  “I don’t know this name, but ask this server”

12. Application Layer 2-12 4 5 6 3 recursive query:  puts burden of name resolution on contacted name server  heavy load at upper levels of hierarchy? requesting host cis.poly.edu gaia.cs.umass.edu root DNS server local DNS server dns.poly.edu 1 2 7 authoritative DNS server dns.cs.umass.edu 8 DNS name resolution example TLD DNS server

13. Application Layer 2-13 DNS: caching, updating records  once (any) name server learns mapping, it caches mapping  cache entries timeout (disappear) after some time (TTL)  TLD servers typically cached in local name servers thus root name servers not often visited  cached entries may be out-of-date (best effort name-to-address translation!)  if name host changes IP address, may not be known Internet-wide until all TTLs expire  update/notify mechanisms proposed IETF standard  RFC 2136

14. Application Layer 2-14 DNS records DNS: distributed db storing resource records (RR) type=NS  name is domain (e.g., foo.com)  value is hostname of authoritative name server for this domain RR format: (name, value, type, ttl) type=A  name is hostname  value is IP address type=CNAME  name is alias name for some “canonical” (the real) name  www.ibm.com is really servereast.backup2.ibm.com  value is canonical name type=MX  value is name of mailserver associated with name

15. Application Layer 2-15 DNS protocol, messages  query and reply messages, both with same message format msg header  identification: 16 bit # for query, reply to query uses same #  flags:  query or reply  recursion desired  recursion available  reply is authoritative identificationflags # questions questions (variable # of questions) # additional RRs # authority RRs # answer RRs answers (variable # of RRs) authority (variable # of RRs) additional info (variable # of RRs) 2 bytes

16. Application Layer 2-16 name, type fields for a query RRs in response to query records for authoritative servers additional “helpful” info that may be used identificationflags # questions questions (variable # of questions) # additional RRs # authority RRs # answer RRs answers (variable # of RRs) authority (variable # of RRs) additional info (variable # of RRs) DNS protocol, messages 2 bytes

17. Application Layer 2-17 Inserting records into DNS  example: new startup “Network Utopia”  register name networkuptopia.com at DNS registrar (e.g., Network Solutions)  provide names, IP addresses of authoritative name server (primary and secondary)  registrar inserts two RRs into.com TLD server: (networkutopia.com, dns1.networkutopia.com, NS) (dns1.networkutopia.com, 212.212.212.1, A)  create authoritative server type A record for www.networkuptopia.com; type MX record for networkutopia.com

18. Objective Score (e.g., Peak Signal to Noise Ratio) Subjective Scores (e.g., Mean Opinion Score) Traditional Video Quality Assessment S.R. Gulliver and G. Ghinea. Defining user perception of distributed multimedia quality. ACM TOMCCAP 2006. W. Wu et al. Quality of experience in distributed interactive multimedia environments: toward a theoretical framework. In ACM Multimedia 2009

19. Objective Scores PSNR Join Time, Avg. bitrate, … Subjective Scores MOS Engagement measures (e.g., Fraction of video viewed) Internet video quality

20. “Binned” rank correlation  Traditional correlation: Pearson  Assumes linear relationship + Gaussian noise  Use rank correlation to avoid this  Kendall (ideal) but expensive  Spearman pretty good in practice  Use binning to avoid impact of “samplers”

21. Information gain background Nice reference: http://www.autonlab.org/tutorials/ Entropy of a random variable: X P(X) A 0.7 B 0.1 C 0.1 D 0.1 X P(X) A 0.15 B 0.25 C 0.25 D 0.25 “high” “low” Conditional Entropy X Y A L B M B N X Y A L A M B N B O “high” “low” Information Gain

22. Why is information gain useful?  Makes no assumption about “nature” of relationship (e.g., monotone, inc/dec)  Just exposes that there is some relation  Commonly used in feature selection  Very useful to uncover hidden relationships between variables!

23. Why naïve regression will not work  Not all relationships are “linear”  E.g., average bitrate vs engagement?  Use only after confirming roughly linear relationship

24. Flash: Enabling Technology For Phase 2 Growth 2005-2011  Client-side: Flash being the de-facto platform  Works in all browsers  Works across OSes  RTMP being the protocol  Maturing CDN technologies to support streaming

25. What is Flash?  Browser plugin developed by Macromedia (acquired by Adobe in 2005)  A programming environment that is independent of browser types and OSes  Rich interactive features and seamless browser integration  desktop application- like features  Flash Player 9.0 + ActionScript 3.0 introduced in 2006 brought a new virtual machine and a just-in- time compiler with 10x performance improvements enabling a wave of more sophisticated and interactive video and web applications  Support for video  Rendering, decoding, introduced H.264 in 2007  Streaming (RTMP),  Content protection (RTMPE, tokenization)  More than 95% penetration on PCs

26. 2010: Apps and Multi-Devices  Rise of apps and the the slow decline of Flash as the universal video platform  iPhone was launched on June 29 th, 2007 without support for Flash  iPad was launched on April 3 rd, 2010 also without support for Flash  Adobe announced removing support for Flash on Android on June 27 th, 2012 (almost exactly 5 years after iPhone launched)

27. 2011 – Present  Four separate screens  PC  Phone  Tablet  TV: connected directly to Internet or via other devices (e.g. Roku, Xbox, PlayStation, AppleTV)  Different user behaviors and video applications for different screens  Fragmented playback software environment  Apps on mobile and tablet devices (iOS, Android)  Apps on game consoles and connected TVs (Xbox, PS3, Roku, AppleTV, Samsung TV, LG TV, etc.)  Both Flash and Silverlight on PCs

28. Internet Video Requirements  Smooth/continuous playback  Elasticity to startup delay: need to think in terms of RTTs  Elasticity to throughput  Multiple encodings: 200Kbps, 1Mbps, 2 Mbps, 6 Mbps, 30Mbps  Multiple classes of applications with different requirements DelayBandwidthExamples 2, N-way conference < 200 ms4 kbps audio only, 200 kbps – 5 Mbps video Skype, Google hangout, Polycom, Cisco Short form VoD< 1-5s300 kbps – 2 Mbps & higherYoutube Long form VoD< 5-30s500 kbps – 6 Mbps & higherNetflix, Hulu, Qiyi, HBOGO Live Broadcast< 5-10s500 kbps – 6 Mbps & higherWatchESPN, MLB Linear Channel< 60s500 kbps – 6 Mbps & higherDirectTV Live

29. Playout Buffer, Delay, Smooth Playback Time Max Buffer Duration = allowable jitter File Position Max Buffer Size Smooth Playback Time Buffer almost empty "Good" Region: smooth playback "Bad": Buffer underflows and playback stops "Bad": Buffer overrflows Buffer Duration Buffer Size

30. 3rd Generation: HTTP Streaming  Key observations: rather than adapt Internet to streaming, adapt media delivery to the Internet  Other terms for similar concepts: Adaptive Streaming, Smooth Streaming, HTTP Chunking  Client-centric architecture with stateful client and stateless server  Standard server: Web servers  Standard Protocol: HTTP  Session state and logic maintained at client  Video is broken into multiple chunks  Chunks begin with keyframe so independent of other chunks  A series of HTTP progressive downloads of chunks  Playing chunks in sequence gives seamless video

31. Adaptive Multi-Bit Rate with HTTP Streaming  Encode video at different levels of quality/bandwidth  Client can adapt to different bit rates within a single session by requesting different sized chunks  Chunks of different bit rates must be synchronized  All encodings have the same chunk boundaries and all chunks start with key frames, so you can make smooth splices to chunks of higher or lower bit rates

32. HTTP Chunking Protocol HTTP Adaptive Player Web browser Web server HTTP TCP … HTTP TCP … A1A1 A1A1 A2A2 B1B1 B2B2 A1A1 B1B1 Cache Client Web server … … A1A1 A2A2 B1B1 B2B2 HTTP GET A 1 Server B2B2 HTTP GET B 2

33. Reasons for Wide Adoption HTTP Adaptive Player Web browser Web server HTTP TCP … HTTP TCP … A1A1 A1A1 A2A2 B1B1 B2B2 B1B1 Cache Client Web server … … A1A1 A2A2 B1B1 B2B2 CDN Infrastructure Reuse the CDN infrastructure Client-driven control enables server/CDN switch Middlebox/firewa ll penetration

34. Example of HTTP Streaming Protocols  Apple HLS: HTTP Live Streaming  Microsoft IIS Smooth Streaming: part of Silverlight  Adobe HDS: HTTP Dynamic Streaming  DASH: Dynamic Adaptive Streaming over HTTP

35. Smooth Streaming IIS Server with Smooth Streaming Extension CharlieBitMe_ 10Mbps.MP4 Mezzanine file 5Mbps.MP 4 Encoders 1Mbp s.MP4 500500 500500.ISMC.ISM Server & Client Manifest files Fetch Client Manifest File HTTP GET http://video.foo.com/CharlieBiteMe.ism/Qu alityLevels(500000)/Fragments(video=0) HTTP GET http://video.foo.com/CharlieBiteMe.ism/Quality Levels(1000000)/Fragments(video=300000)

36. HTTP Live Streaming (HLS) CharlieBitMe_ 10Mbps.MP4 Mezzanine file Encoders 5Mbps.MP 4 1Mbp s.MP4 500500 500500 Stream Segmente rs Per-bitrate.ts media segment files & playlists.m3u8 Master Playlist Web Server Fetch master play list HTTP GET http://media.example.com/segment0.ts Fetch bitrate specific playlist

37. HTTP Dynamic Streaming (HDS) CharlieBitMe_ 10Mbps.MP4 Mezzanine file Encoders 5Mbps.MP 4 1Mbp s.MP4 500500 500500 f4f Packager Per-bitrate.f4f segment &.f4m manifest files.f4m Apache with Adobe HTTP Origin module HTTP GET http://www.example.com/ media/CharlieBitMe.f4m HTTP GET http://www.example.com/media/http_d ynamic_StreamingSeg1-Frag1

38. What We Have Learned So Far? (1)  The detour to Web Internet has a key positive legacy  HTTP chunk will be the new Datagram for Internet video  HTTP chunk switches are the new switches/CDN servers  Many practical problems solved for HTTP after years of evolution  Middle-box support, authentication, firewall penetration, anycast 38 1990 Internet Web Internet Video & Web Internet

39. What Have We Learned So Far (2) ?  No universal QoS support inside the network  CDN a key architectural component to optimize performance  DNS-based names are standard control service access interface  HTTP is standard data plane plane protocol

40. What We Have Learned So Far (3)? 750 Kbps 1.2 Mbps CDNISP GEODevice 750 Kbps 1.2 Mbps 3 Mbps  Adaptive multi-bit-rate video key to address diversity