can internet video-on-demand be profitable? sigcomm 2007 cheng huang (microsoft research), jin li...
TRANSCRIPT
Can Internet Video-on-Demand Be Profitable?
SIGCOMM 2007
Cheng Huang (Microsoft Research),
Jin Li (Microsoft Research),
Keith W. Ross (Polytechnic University)
Presenter: Junction
Motivation
• VoD such as YouTube, MSN Video, Google Video, Yahoo Video, CNN…
• As the trend of increasing demands on such services and higher-quality videos, it becomes a costly service to provide.
• Using Peer-assisted to replace Server-Client :– By reducing the server’s bandwidth to reduce the cost
that providers pay to ISPs.
Implementation
• Using a nine-month trace from a client-server VoD deployment for MSN Video to gain some observation
• Present a theory for peer-assisted VoD• Simulation• Impact of peer-assisted VoD on the cross-traffic
among ISPs
Characteristics of a Large Scale VoD Service
• Data Collection: – 2006 April to December: MSN Video service– Client-server mode– Covering over 520 million streaming requests for more
than 59,000 videos.• Trace Records
– Client Information Fields (ID, IP address, version…)– Video Content Fields (length, size, bitrate)– Streaming Field (connection, last, interactive…)
Identifying Users and Streaming Sessions
• ID-identified (7%) & hash-identified player• different hashes come from different players
• Streaming session : – A series of streaming requests from the same player to the same
video file. (471/520)
Video Popularity Distribution• The greater the locality of requests to a subset of the
videos, the greater the potential benefit for peer-assisted streaming.
Similar regardless of trafficHigh-degree of locality
Zipf distribution with flat
User Demand and Upload Resources• Estimate the upload bandwidth of a user by download
bandwidths.
Distribution of user download bandwidths Aggregate user demand and upload resources (April 18)
User bandwidth breakdown (KBPS)
Peer-assisted VoD might perform well
User Interactivity• View larger fraction of short videos• A large fraction of the users view videos without
interactivity (> 60%)• It’s important to understand this interactivity while
considering peer-assisted solutions for VoD.
No interactivity does better
95 Percentile Rule
• ISP charges the service provider each month according the service provider’s peak bandwidth usage.
Theory of Peer-Assisted VoD
• Single video & multiple video approach– Single : only redistributes the video currently watching– Multiple : redistribute a video previously viewed
• Three basic operation modes– Surplus mode (S>D)– Balanced mode (S~=D)– Deficit mode (S<D)
Theory of three modes• Video rate : γ bps• M user types• : upload link bandwidth of a type m user• λ : the parameter of Poisson process to describe Users arrival• : the probability that an arrival is a type m user
compound Poisson process
m user types arrive as independent Poisson processes with parameters λ• : The average upload bandwidth of an arriving user • σ : a user’s expected sojourn time in the system
Little’s law
the expected # type users in the system is
in steady state : the average demand is
the average supply is
mw
mp
mp
mmwp
mm p
MD
mmwS
No-Prefetching Policy• Each user downloads content at the playback rate and
doesn’t prefetch content for future needs.– For n = 1, we have s(u1) = r.– For n = 2, we have s(u1, u2) = r + max(0, r-u1).– So on ….
(w1= 768 kbps, w2 = 256 kbps, γ = 512 kbps, σ = 300s)
If (r-u1)<0, still upload r
Bandwidth Allocation Policies for Prefetching
• Surplus upload capacity used to distribute – future content– creating a reservoir of prefetched content – exploited when the system shifts into a deficit state.– Operate better in the balanced mode.
• Water-leveling & greedy
Water-leveling & Greedy
• Ranking by the arrival order• determining required server rate• Allocate and adjust the growth rates
– the growth rate of user k+1 doesn’t exceed user k
@ data demands imposed on the server usually generated by oldest
• Greedy : each user simply dedicates is remaining upload bandwidth to the next user right after itself.
Simulation Result
• Lower bound : a peer can feed content to any peer, not just to the peers that arrived after it.
Real-World Case Study• Three cases:
– All users watch the entire video– With early departures– With both early departures and user interactivity
• Trace Analysis for the Two Most Popular Videos (case 1)
• Typically no server resource are needed
• Valleys• Flash crowd / long-lasting
Impact of Early Departures• Drive the system from the surplus mode, through the
balanced mode, to the deficit mode by scaling the video bitrate.
• Even with early departures, peer-assistance can provide a dramatic improvement in performance.
The Impact of P2P on Internet Server Providers - ISP
• Relationship between ISPs– Transit, sibling, peering– Majority of P2P traffic is crossing entity boundaries