Streaming Video Traffic: Characterization and Network Impact Kobus van der Merwe Shubho Sen Chuck Kalmanek
Streaming Media Study: Why ? Lot of streaming on the Internet Quality is getting pretty good Streaming is not well understood User behavior Factors that impact quality Network impact + distribution Reasons: Proprietary protocols WM, Real Very commercial logs files are sensitive and hard to get
On Demand Dates: 12/ /2002 # Requests: 3.5 million # Unique IPs : 0.5 million # Unique ASs : 6600 WM and Real BW:~56 Kbps and ~300 Kbps Live Dates: 02/2002 – 03/2002 # Requests: 1 million # Unique IPs: 0.28 million # Unique ASs: 4000 WM only BW: ~100 Kbps encoded The Data On Demand: prerecorded clips from current affairs & information site Live: commerce oriented continuous live stream Routing data: daily BGP table dumps from Tier-1 ISP Traffic volume : several terabytes
On Demand: Traffic Composition By transport: HTTP : 37% requests, 27% bytes TCP : 29% requests, 45% bytes UDP : 34 % requests, 28% bytes Proprietary Streaming dominates: 63 % requests, 73 % bytes Total TCP dominates: 66 % requests, 72 % bytes - probably because of firewalls By Bandwidth (56 Kbps/300 Kbps) : High BW dominates: 65% requests, 95% bytes Low BW: 35% of sessions account for just 5% of data By protocol (WM/Real): Windows Media dominates: 77% requests, 76% bytes
On Demand: per-AS breakdown by protocol # Requests Traffic volume Most ASs generate more MMS than RealMedia Traffic ASs contributing 80% requests or 80% traffic
On Demand: per-AS breakdown by stream bandwidth # RequestsTraffic volume Most ASs generate more High Bandwidth traffic
Live: Traffic Composition By transport: HTTP : 55% requests, 47% bytes TCP : 17% requests, 38% bytes UDP : 28 % requests, 17% bytes Proprietary Streaming (TCP + UDP) : 45 % requests, 55 % bytes Total TCP dominates: 72 % requests, 85 % bytes - probably because of firewalls Proprietary Streaming, HTTP have similar shares
On Demand: Network Traffic Distribution Significant variability in traffic contributions: 10% ASes account for 82% requests, 85% data # RequestsVolume
Content Distribution Impact Goal: Evaluate different content distribution approaches Centralized + IP peering Centralized + content peering Centralized + replica placement Assume traffic distributed from (originating from) Tier-1 ISP Look at coverage achieved by different approaches Traffic distribution using AS hop-count from Tier-1 ISP as a metric Assumption: for streams originating in Tier-1 ISP small AS-hop count will increase probability of acceptable quality
SetOn-demand % Traffic Vol Live % Traffic Vol <= 1 AS hop <= 2 AS hops AS hops selected ASes AS hops + 15 selected ASes Content Distribution Impact Selected ASes: “consistent contributors” out of 6600 Caveats: Hop count not good metric of anything Limited data set Data set might be self selecting
On Demand :Traffic Time Series Significant variability within/across days Peak = 31 * Mean
On Demand :Rapid Increase in Load Load increases 57 times in 10 minutes !
Live: Traffic Time Series Significant variability within/across days Peak = 9* Mean
Object Popularities Few heavy-hitters account for bulk of traffic Dec 13: top 5 clips account for 85% of traffic Volume 320 clips # Sessions 320 clips
On Demand: Session Characteristics Most sessions download a fraction of the object. A larger proportion of high bw clip is downloaded High Bw mmsLow Bw mms
Summary Windows Media dominates High encoding rates dominate TCP transport dominate Highly skewed request + volume distributions Tier-1 ISPs cover % < 2 AS hops Significant coverage with small # selective arrangements High variability in daily traffic patterns Ramp up in tens of minutes Highly skewed object popularity High bit-rate clips watched longer