1. Peer-assisted VoD for set-top box based IP network
Vaishnav Janardhan & Henning Schulzrinne
Dept. of Computer Science
Columbia University
New York, NY
August 31, 2007
P2P-TV SIGCOMM

2. Costs in providing video content DVRs Architecture Challenges
Overview
Costs in providing video content
DVRs
Architecture
local DHTs and pre-fetching
Challenges
August 31, 2007
P2P-TV SIGCOMM

3. Economics of VoD Transit bandwidth $40/Mb/s/month ~ $0.125/GB
Transit bandwidth $40/Mb/s/month ~ $0.125/GB
US colocation providers charge $0.30/GB to $1.75/GB
Netflix postage cost: $0.70 round-trip
Typical PPV charges: $4/movie (7 GB)
August 31, 2007
P2P-TV SIGCOMM

4. Cost for providing content
across provider boundaries
possibly another step when crossing oceans
within campus/AS
(multiple L2s)
same L2 switch
(non-blocking)
distance
within home
August 31, 2007
P2P-TV SIGCOMM

5. Example: FiOS TV architecture
Fiber
Serving
Office
Hub
Super
Headend
Broadcast Video
Voice, Data, IP TV
Splitter
Joseph Savage, Nov. 29, 2006
2 national super headends
9 video hub offices
292 video serving offices
J. Savage (Telecom ThinkTank), Nov. 2006
August 31, 2007
P2P-TV SIGCOMM

6. Verizon’s FTTP Architecture
CUSTOMER PREMISE
Voice & Data Downstream 1490 nm
Voice, Data & Video nm, 1310 nm, 1550 nm
OLT
Optical Line Terminal
ONT
Optical Network Terminal
Optical Couplers (WDM)
Optical Splitter
Upstream 1310 nm
Video 1550 nm
1x32
CENTRAL OFFICE
EDFA
Erbium Doped Fiber Amplifier
Bandwidth & Services
by Brian Whitton, Verizon Networks
Upstream
Downstream
1310 nm
1490 nm
1550 nm
Voice & Data at 155 to 622 Mbps
Voice, Data & VOD at 622 Mbps
Broadcast Video
54 MHz
864 MHz
Analog TV
Digital TV and HDTV
Brian Whitton, Verizon
August 31, 2007
P2P-TV SIGCOMM

7. Properties of DVRs Storage of 80-250 GB (Tivo 3)
Probably on-line 24/7 already
Often, directly connected to network (“home gateway”)
May be owned by cable or DSL company
August 31, 2007
P2P-TV SIGCOMM

8. (P2P) video variants
Lots of variants - with very different requirements
Mode
start-up time
VCR controls
content
near VoD
minutes - hours (~ BlockBuster)
full, including skip ahead
movies, UGC
VoD
seconds
full
live streaming
none
TiVo-like (pause, rewind)
news, sports
August 31, 2007
P2P-TV SIGCOMM

9. VoD approaches Internet video ISP-provided VoD
network
WAN
LAN
content provider
Internet video
(YouTube, Netflix, ...)
ISP-provided VoD
end users
Classical P2P (BitTorrent, ...)
this approach
servers
August 31, 2007
P2P-TV SIGCOMM

10. VoD requirements short clips < 10’ (long tail) feature-length
Example: Superbad grossed $33M during August 17 weekend (in US)
= roughly 3M viewers
= roughly 1% of US population
 if VoD, each neighborhood has likely one copy
2 problems:
get initial copy to neighborhood
multicast, OTA
distribute in neighborhood
only viable for top 1000 content
feature-length
avoid Netflix queue
avoid stocking 20,000 DVDs
BlockBuster has 5,192 stores in the US
NetFlix has 85,000 movie titles on DVD
August 31, 2007
P2P-TV SIGCOMM

11. Assumptions Every P2P scheme needs to address those DRM is orthogonal
i.e., access to bits  access to content
may not work if DRM assumes individualized content
keying or fingerprinting
Upstream bandwidth is sufficient to deliver >= 1 stream
true for modern FTTH and FTTC networks
if not, P2P systems only work if ∑ upstream > ∑ consumption
if near-VoD, averaging interval may be whole day, rather than peak viewing period
but still need time to buffer content  delay and no feedback on FF
DVRs have spare capacity
likely true for PCs
may be optimistic for DVRs using LRU-style storage management
may be able to leverage content having been viewed by user
if owned by ISP, cheating problems disappears (no need for tit-for-tat)
DVRs can’t store all content
85,000 DVDs  595 TB
August 31, 2007
P2P-TV SIGCOMM

12. Notes on cost shifting Servers vs. bandwidth
Fixed vs. incremental costs
for VoD providers, each (peak) stream incurs additional cost
for end systems, generally $0
Bandwidth
providers - ~ peak usage
ISP - want to avoid paid (= non-local) traffic
users - may not care, but may be rate-limited or violate contract
no cost impact as long as downstream >> upstream bandwidth
e.g., Columbia severely limits student bandwidth
“Quotas are 350M/hr download and 180M/hr upload” (= 400 kb/s)
not much extra upstream bandwidth left
August 31, 2007
P2P-TV SIGCOMM

13. Example: Columbia University
Our commodity routers are nyser111-gw-1 which connects to Broadwing(now Level(3) at 500 Mbps), nevis-gw-1 which connects to Qwest at 45Mbps, Vortex which connects to RCN (private peering) at 100 Mbps,nyser32-gw-1 which connects to NYSERNet at 200 Mbps and Ultralight(CERN) at 1 Gbps.
ratio not much upstream capacity left
August 31, 2007
P2P-TV SIGCOMM

14. Network architecture National Backbone Los Angeles Chicago New York
Dallas
Regional Data Center
Server services:
DNS
DHCP
August 31, 2007
P2P-TV SIGCOMM

15. Architecture Try to find content locally (AS)
using a local (provider-internal) DHT by identifier
identify peer with available capacity
cf. Aggarwal (CCR 7/07) to identify candidate nodes
If local, stream from peer
assume single server upstream bandwidth is sufficient
otherwise, piece together multiple servers
could use standard RTSP VCR controls
Use extra upstream capacity for pre-fetching content
first, retrieve key frames and anchor points for fast-forward
MPEG: 1/15th of frames
then, rest of video
handles bandwidth variability & releases server earlier for other uses
If not local, contact ISP (caching) video server
e.g., RTSP redirect
August 31, 2007
P2P-TV SIGCOMM

16. Pre-fetching t (sec) 60 seconds 60 seconds 60 seconds
Adjust to anchor point
Adjust to anchor point
t (sec)
5 sec
5 sec
5 sec
Anchor
point
Seek point
Anchor
point
Seek point
Anchor
point
60 seconds
60 seconds
60 seconds
August 31, 2007
P2P-TV SIGCOMM

17. Pre-fetching Peer 1 [seed] Peer 2 [leech] Peer 4 Peer 3 Peer 5 Tracker
Sliding Window module
Pre-fetching module
August 31, 2007
P2P-TV SIGCOMM

18. Conclusion Need careful analysis of cost trade-off
P2P may only be optimal if you ignore network costs
compare to classical proxy architectures
clearly identify assumptions -- more than one “P2P video”
Presented combination of different approaches
Locally popular content remains local
Mid-list content at end users
“Long tail” content at ISP
Back list at content provider
What is the minimal set of tools and building blocks?
August 31, 2007
P2P-TV SIGCOMM

19. Admission control DVR has small upload capacity
during busy time, may have > 50% DVR utilization
Content replication converges to popularity
But also hosts rare content only available once in network
Allow client displacement
new client indicates rare content (“last resort”)
DVR tries to find alternative source for existing user
and serves new client
August 31, 2007
P2P-TV SIGCOMM