peer-to-peer television for the ip multimedia subsystem

Author

Advisor

Universidad Carlos III de Madrid

Peer-to-Peer Televisionfor the

IP Multimedia SubsystemAlex Bikfalvi

Jaime García-Reinoso

Peer-to-Peer Television for the IP Multimedia Subsystem 2

OutlineI

II

III

IV

BackgroundMotivation • Contributions • Peer-to-peer streaming • IP Multimedia Subsystem

Peer-to-Peer Television for the IMSService architecture • Signaling protocol • Support for mobility

The User Activity in IPTVData and modeling • Synthesis

Enhancements at the Application ServerSignaling delay • Multiple TV channels

V Performance EvaluationBasic evaluation • Hybrid streaming

July 18, 2012

VI ConclusionsSummary • Publications • Future enhancements


BackgroundPart I

July 18, 2012


MotivationInternet Protocol Television• Increasing interest in the recent years• Deployments and triple-play packages• Competition with Internet Services

Next Generation Network• Flexible platform for any type of service• Decouples service provisioning from the

network• Adequate level of quality of service (QoS)

IP Multimedia Subsystem• Framework for IP multimedia services• Session control using the Session

Initiation Protocol (SIP)

TISPAN• Extended IMS NGN to multiple access

technologies• Services standardization, including IPTV

1

2

3

4

July 18, 2012


Motivation

July 18, 2012

• High performance• Availability with existing protocols and

equipments• Commercial IPTV deployments: walled

gardens

In TISPAN, broadcast television uses IP multicast

• Large number of TV channels• Static multicast: inefficient for many TV

channels• Dynamic multicast: delay and scalability

issues• Administrative and economic reasons• Support for multiple transport protocols


Going Peer-to-Peer…

July 18, 2012

Peer-to-peer television (P2PTV) service for the IP Multimedia Subsystem

Viewer

TV set

Broadcast server

Transport network (telco)

Set-top box

IPTV

Internet

Unused

• Exploit the unused download and upload capacity


Color Key Broadcast server Viewer

TV set

Broadcast server


Peer-to-peer

Set-top box

Going Peer-to-Peer…

July 18, 2012


• Exploit the unused download and upload capacity

• Dedicated user equipments (set-top boxes)

• Streaming transparent to the user


P2PTV-AS

IP transport network

IP Multimedia Subsystem

TV set

Viewer

User equipment (peer)

Application Server• Control service access• Manage peer participation• Implement enhancements

User Equipment (peer)• Download current

channel• Download other streams• Upload streams to other

peers

… with the IMS

July 18, 2012



Challenges

July 18, 2012

We don’t break new ground in peer-to-peer streaming

However…• Integration with a commercial-grade service• Multiple TV channels increase peer churn (over

60% of channel changes in 10 seconds[1,2])• IMS signaling requirements increase the setup

delay

Two enhancements…Fast signalingInactive uploading

sessions with committed QoS resources

Low churnPeer participation on

multiple TV channels[3]

1 2

[1] Cha et al., Watching Television Over an IP Network, 2008[2] Qiu et al., Modeling user activities in a large IPTV system, 2009[3] Wu et al,. View-upload decoupling: A redesign of multichannel P2P video systems, 2009


BackgroundPart I

July 18, 2012

Peer-to-Peer Streaming

11Peer-to-Peer Television for the IP Multimedia Subsystem

Peer-to-Peer Streaming• Emerged as a response to IP multicast

issues

July 18, 2012

P2P

CDN

IP multicast

Support Scalability

Cost Resources

Complex

Low

Med

High

Med

High

High

Low

High

Low

Low

Med

High

Low

Med

High

• Initially P2P emulated IP multicast• Application-level multicast• Data forwarded along a tree overlay between

hosts



July 18, 2012

Broadcast server

Interior peer

Leaf peer

Overlay tree between end-hosts or peers

• Data flow may be push or pull

• Accommodates one stream or channel


Broadcast server

Departing peer

Affected peers


July 18, 2012

Overlay tree between end-hosts or peers

• Data flow may be push or pull

• Accommodates one stream or channel

• Churn: interruptions due to the departing peers


Multiple TV Channels

July 18, 2012

Uploading peers

Color Key Blue channel Orange channel

Broadcast server

1

2 3

4 5

A

B

A B C D E

C

D E

1 2 3 4 5 Viewing peers

Increased peer churn due to channel changes

• View-upload decoupling[1]

[1] Wu et al., View-upload decoupling: A redesign of multichannel P2P video systems, 2008


BackgroundPart I

July 18, 2012

The IP Multimedia Subsystem



July 18, 2012

Next Generation NetworksIntegrated broadband IP networks for multimedia

servicesThird Generation

Partnership Project (3GPP)

• Quality of service• Service

implementation• Seamless mobility• Authentication, policy

and charging


Decouples service from the transport

network

Functional entities and standardized

interfaces



July 18, 2012

Mi

PCRF

P-CSCF S-CSCF

SLF

HSS

Gx

Rx

Mw Mw

Cx

Cx Dx

Dx

ISC Sh

UE

Sh

ISC

IP network

Mr

IM-SSF

Sh

SIP-AS

OSA-SCS

ISC MRFC

MRFP

Mp

Mj

BGCF

MGCF

I-CSCF

Gm

Mn

Legacy terminal

3GPP terminal

IMS terminal

IMS gateway

MGW

SGW

PSTN

Transport plane

Call-session control

Control plane

Application plane

Subscriber database

Policy and charging

Gateways

Applications servers

Proxy Serving


Peer-to-Peer Television for the IP Multimedia Subsystem

Part II

July 18, 2012


Service Architecture

July 18, 2012

Common platform for IPTV streaming using peer-to-peer technology: P2PTV

Signaling connection

Color Key Media connection

Set-top box (STB)

Viewer

TV set IPTV provider(s)


Call Session Control Functions

Application Servers


Broadcast servers

User Equipment

P2PTV Application

Server

Broadcast Servers


Application Server

July 18, 2012

SIP SignalingManage the IMS

multimedia sessions with the UE

INVITE sip:[email protected]: sip:[email protected]: sip:[email protected]: 1000

UE

sip:[email protected]

P2PTV-AS

sip:p2ptv-as.example.net

1

S-CSCF

UAS

2

User agent server• Sessions terminating

at the P2PTV-AS• Streaming by the

broadcast server


From: sip:[email protected]: sip:[email protected]: 4000


Application Server

July 18, 2012



UE


P2PTV-AS


Back-to-back user agent• Peer-to-peer

streaming

S-CSCF

B2BUA

UE



Channel Streaming

July 18, 2012

The P2P push-pull streaming generates an overlay

Physical layer

UE UE

UE

BS

Overlay layer

UE

UE

UE

UE UE

UE

A TV channel is divided in multiple (e.g. 3)

streams

Ideally, a UE peer downloads all streams

when tuning to the channel

A UE peer may upload one or more streams to

overlay neighbors

This strategy works well with multiple description codecs such as H.264/SVC


Streaming Enhancements

July 18, 2012

We face the following challenges…• The IMS session signaling is an expensive

operation• Streaming multiple channels with a classic

experienceWe propose two solutions…Fast signaling

Inactive uploading sessions with committed

QoS resources

Low churnPeer participation on multiple TV channels

1 2

• The first targets the control plane• The second targets the media plane


Fast Signaling

July 18, 2012

The P2P streaming requires two multimedia sessions

• Downloading side, low dynamics, reusable during channel changes

• Uploading side, high dynamics, non-reusable during channel changes

Inactive session

Inactive session

Upload Download

Foster peer

Stream 1

Stream 2

Stream 3

Uploading neighbors Downloading neighbors The performance bottleneck is on the

upload side

Introduce foster peers with inactive upload

sessionsThese accommodate new requests without

establishing new sessions


Stream 100 Stream 200

P1

P2

P3 P

4

P5

P6

BS BS

Fast Signaling

July 18, 2012

1

2

Situations that benefit from foster peers

Stream 100 Stream 200

P1

P2

P6

P3 P

5

P4

BS BS

1

• Fast stream change when the user changes the current channel

• Fast recovery to accommodate peer churn when occurs

Changing stream

Fast stream change

Fast recovery

Number of inactive sessions


Fast Signaling

July 18, 2012

Does not require a new session on the upload side

• Initiated by the P2PTV-AS as a response to user demand

S-CSCF P-CSCF PCRFP2PTV-

ASUE

INVITE1

INVITE sip:[email protected]: sip:[email protected]=stream 100a=inactive

INVITEINVITE

183 Session Progress

183 Session Progress183 Session Progress

sip:[email protected] sip:[email protected]

AA-Request

AA-Answer 2


Fast Signaling

July 18, 2012

Changing the TV stream using a foster peer

P-CSCF S-CSCFUEdownP2PTV-

ASUEup


UPDATE1

UPDATE sip:[email protected]: sip:[email protected]: sip:[email protected]=stream 100c=IN IP4 10.0.0.1a=curr:qos local recva=curr:qos remote send

UPDATE UPDATE UPDATE

UPDATE sip:[email protected]: sip:[email protected]: sip:[email protected]=stream 100c=IN IP4 10.0.0.1a=curr:qos local senda=curr:qos remote recv

200 OK200 OK200 OK200 OK


Session Coordination Algorithm

User activity Time sample k

Channel Sessions

1

i ( , )w k i

( ,1)w k

Blocking ratio Utilization

Fast Signaling

July 18, 2012

It sounds simple, but…

How many inactive sessions accommodate the TV channel demand?

• Too few, no fast signaling and high channel change delay

• Too many, waste network resources with reserved bandwidth

On a given TV channel

P2PTV-AS

w2 w3

w6 w5 w4

w7 w8 w9

w1

Subscribers

Peer inactive sessions

wi

1 User activity

2Inactive sessions


Peer Churn

July 18, 2012

Peers download streams from multiple TV channels

• Primary streams correspond to the current TV channel

• Secondary streams from other TV channelsOnly the primary streams are affected by

the channel change churn

Upload Download

Primary streams

Secondary streams

Active

Active

Inactive

Inactive

10

20

30

40

10

10

20

30

30

30

40

Current channel

1

2

3

UE peer

TV set

Primary streams are used for viewing

Secondary streams are used for uploading

The cost is the increased bandwidth

usage


Peer Churn

July 18, 2012

This is view-upload decoupling: can we do better?

• Complete decoupling wastes bandwidth• Upload primary streams for peers with free

bandwidth

Upload Download

UE peer

10

20

2

2

10

10

10

20

20

1

2

3

TV set

Primary streams

Secondary streams

Free bandwidth

Active

Active

Inactive

Inactive

Peers may need a new download session at

the next channel change

Primary streams become secondary

Self-organizing, no centralized assignment of secondary streams

Higher delay: use inactive download

sessions


Peer Coordination

July 18, 2012

Peer Sessions

Download Upload

Primary

Secondary

Active

Inactive

PCA: peer selection

PCA: session coordination

User activity

Self-organizing/PCA

• The session coordination computes the number of inactive sessions for a channel wi

The peer coordination assigns peer resources…

• … to accommodate the demand

We discuss both algorithms in further detail in part IV


Peer-to-Peer Television for the IP Multimedia Subsystem

Part II

July 18, 2012

Support for Mobility


Support for Mobility

July 18, 2012

We examine the performance in roaming situations

Minimize the loss of streaming data…• Buffering mechanism compensating for

connectivity loss• Reducing the handover delay

Existing solutions…SIP

Establish a new session after

roaming to the new network

Optimized SIPTransfer the

session context between the old and new P-CSCF

to meet the session

preconditions

Mobile IPTunnel the video

data from the home to the

visited network

1 2 3


Proactive Context Transfer

July 18, 2012

The UE must reestablish the session in the new network

• We exploit the handover delay when the UE is disconnected

• The network takes an active participation in the handover

• Use the IEEE 802.21 (MIH) standard

Unfortunately…

Proactive Context Transfer Service Application Server

• One in every network• Part of the MIH point-of-

service• Notified by the UE before the

handover• Installs the session context

in the network at the P-CSCF• Applies to SIP or MIP mobility


Performance Evaluation

July 18, 2012

Comparing the handover delay with previous scenarios

ho sip mip attach pdpp pdpsT T T TT T • Delay components

SIP and MIP handover delay Total handover delay for UMTS

SIP

and

MIP

dela

yT

sip

+T

mip

[s]

0

0.5

1

1.5

2

2.5

3Tsip

Tmip

SIPopt

MIPhomeMIPvisited

PCTSSIPPCTSM

Han

dove

rde

lay

Tho

[s]

SIPopt

MIPhomeMIPvisited

PCTSSIPPCTSMIP

0

2

4

6

8

10TattachTact,PDPpTact,PDPsTsipTm

Home

Visited

SIP

MIP


The User ActivityPart III

July 18, 2012


Objectives

July 18, 2012

Essential for system design and performance evaluation

Measurement studies…• Internet-based services like PPLive, PPStream,

SopCast[1,2,3,4,5]

• Telco IPTV such as Telefonica Imagenio, AT&T U-Verse[6,7,8]

[1] Ali et al., Measurement of commercial peer-to-peer live video streaming, 2006[2] Hei et al., A measurement study of a large-scale P2P IPTV system, 2007[3] Silverston et al., Measuring P2P IPTV systems, 2007[4] Xie et al., A measurement of a large-scale peer-to-peer live video streaming system, 2007[5] Vu et al., Measurement of a large-scale overlay for multimedia streaming, 2007[6] Cha et al., Watching television over an IP network, 2008[7] Qiu et al., Modeling user activities in a large IPTV system, 2009[8] Qiu et al., Modeling channel popularity dynamics in a large IPTV system, 2009

SimulwatchSynthetic workload generator by Qiu

et al.

• Limited number of properties• Low accuracy for some metrics• Some flaws


User Activity

July 18, 2012

The state of the user equipment and current channel[1]

Offline session

Online session

Channel session

[1] Qiu et al., Modeling user activities in a large IPTV system, 2009

Peer-to-Peer Television for the IP Multimedia Subsystem 39onx offx chx

User Activity

July 18, 2012

The state of the user equipment and current channel[1]

ond offd

Online events

Offline events

1onx 1

offx

• Session length• Session rate

ond offd chd


Session Length• A hyper-exponential distribution

July 18, 2012

c1

bc1

c2bc2

c3 bc3

c4 bc4

don [s]CC

DF

1−

F(d

on)

100 1050

0.2

0.4

0.6

0.8

1TraceModelPoints at ciPoints at bci

0 0.5 10

0.5

1

Q-Q p

1

( ) 1 e i

n

id

i

F d p

i

ip

n

• Fitting algorithm of Feldmann et al.[1]

• Fit each exponential on exponentially spaced intervals

,i ic bc [1] Feldmann et al., Fitting mixtures of exponentials to long-tail distributions to analyze network performance models, 1998

Online session interval length


Session Rate• Has a complex daily and weekly pattern

July 18, 2012

t [h]

x on

[min

−1]

0 3 6 9 12 15 18 21 240

0.02

0.04

0.06

0.08

0.

Online session rate (normalized)• Qiu et al. model the spectrum with a continuous

distribution• Limited accuracy: does not include phase information

• Propose dominant frequency components based on power

f [min−1]

Xx(f

)

0 0.05 0.1 0.15

×10−3

0

2

4

f [min−1]φ

x(f

)

0 0.05 0.1 0.15−π

−π2

0

Frequency spectrum

60 min

30 min

20 min

15min

10min

Difficult to model

6.66min


t [h]

x on

(t)

[min

−1]

0 3 6 9 12 15 18 21 240

0.05

0.1Trace

t [h]

x on

(t)

[min

−1]

0 3 6 9 12 15 18 21 240

0.05

0.1Model 6.66 minutesModel 30 minutes

Session Rate• Has a complex daily and weekly pattern

July 18, 2012

Online session rate (normalized)• Qiu et al. model the spectrum with a continuous

distribution• Limited accuracy: does not include phase information

• Propose dominant frequency components based on power

f [min−1]

Xx(f

)

0 0.05 0.1 0.15

×10−3

0

2

4

f [min−1]φ

x(f

)

0 0.05 0.1 0.15−π

−π2

0

Frequency spectrum

60 min

30 min

20 min

15min

10min

Difficult to model

6.66min

59+57 parameters

9+7 parameters


• Include weekly pattern using a modulating function

• Stochastic properties• Difference between trace and model: normal

distribution

• Based on the number of online viewers

Session Rate

July 18, 2012

t [day]

r on

0 1 2 3 4 5 6 70.5

1

1.5Fraction of online viewersContinuous and fundamental

onr

0 1 ,1cos 2 w YY Y f t

* *,on onX x N * *

,off offX x N ,ch chX x N


Workload Synthesis• Generate workload based on analytical

model• Incomplete measurement data on IPTV user

activity[1,2,3]

• Rescale workload dimensions like users or channels

• Conclusions• Better approximation of the activity data• Exclude some details like user preference

July 18, 2012

[1] Cha et al., Watching television over an IP network, 2008[2] Qiu et al., Modeling user activities in a large IPTV system, 2009[3] Qiu et al., Modeling channel popularity dynamics in a large IPTV system, 2009

Online intervalOnline eventOffline event

Offline intervalChannel event Channel interval

Timeline


The Application ServerPart IV

July 18, 2012


Application Server Functions

July 18, 2012



1

Session CoordinationCompute the number of inactive upload sessions

2

Peer CoordinationAssignment of peer

bandwidth across TV streams

3


Session Coordination• The system is like a queue for every TV

channel

July 18, 2012

Service dch

Active sessions wa

Inactive sessions wi

Arrival zch

Blocking

On a given TV channel

From the perspective of

the fast signaling

InputBlocking ratio

β and utilization ρ

Disturbance

User activity u(t), zch(t)

OutputNumber of sessions

w(t)=wa(t)+wi(t)


Simulation of the user arrival

Session Coordination• Finding a relationship between input and

output• No simple distribution for the user activity

dynamics• The temporal dimension is an important element

July 18, 2012

If the arrival would be a Poisson process and the

service rate have an exponential distribution

The number of upload session computed using the Erlang-B

equation

λ/µ

wop

t

0 20 40 60 80 1000

20

40

60

80

100

120

140DataErlang-B: rb =10 −2

Erlang-B: rb =10 −3

Erlang-B: rb =10 −4

M/M/w/w

0 20 400

20

Arrivals served

Arrivals blocked

No strong correlation between the channel user activity and the optimal number of upload

sessions


Session Coordination

July 18, 2012

Use an adaptive algorithm with a feedback loop

SCAInput: r Output: w(k)

Controller System

Delay

Reference: r

Input: w(k)

User activity: u(k)

Output: y(k)Error: e(k)

Design tasks…• Selection of the control signals: input, output and

reference• Determine the controller transfer function:

Time-discrete

system: t k

( ) ( )w k f e k


Session Coordination• Performance evaluation• A large P2PTV deployment with 100 000

subscribers• Synthetic workload spanning 1 week

July 18, 2012

Block

ing

ratio

b

Channel 1 Channel 10 Channel 100Channel 500

×10−3

0

0.5

1

1.5

2β =10 −3

β =10 −4

β =10 −5

β =0

Util

izat

ion

ρ

Channel 1 Channel 10 Channel 100Channel 5000.8

0.9

1

β =10 −3

β =10 −4

β =10 −5

β =0

Unu

sed

sess

ions


50

100β =10 −3

β =10 −4

β =10 −5

β =0

Blocking ratio b(k) Session utilization ρ(k)

Blocking ratio around the desired reference

value

The session utilization is between 90% and 100%


Based on the view-upload decoupling idea…[1]

• Assign peers to secondary and inactive streams• Subject to peer bandwidth constraints

Peer Coordination

July 18, 2012

We leverage the session coordination to estimate the channel demand and allocate

peer resources[1] Wu et al., View-upload decoupling: A redesign of multichannel P2P video systems, 2008

Primary streams: np

Secondary streams: ns

Free download bandwidth

Active streams: na

Inactive streams: ni

Free upload bandwidth

Download Upload


Peer Coordination• Peers self-organize on secondary streams

based on• Requests for active and inactive sessions• Available bandwidth

July 18, 2012

1

2

3

4

4

5

5

4

10

11

12

13

20

30

40

On t

he d

ow

nlo

ad s

ide

Primary Secondary Freepn

10

11

12

13

20

30

40

10

11

12

13

20

30

40

50

51

52

53

13

20

30

40

Initial peer state

Receive request to upload stream 13

Reserve bandwidth for one additional primary

streamAt the next channel change the primary

stream becomes secondary


• Key differences to view-upload decoupling

• Peers report their uploading capabilities• The algorithm uses a resource level metric r to

select peers

Peer Coordination

July 18, 2012

Channel changes Channel changes

View-upload decoupling Peer coordination algorithm

Primary

Secondary Uploading primary

,u f

u

br

b ir w

When looking up peers with free

bandwidth

When looking up peers with

inactive sessions


Peer Coordination• Performance evaluation• Three bandwidth scenarios based on DSL access

July 18, 2012

FractionPoor Middle-class Rich

Download Upload Downloa

d Upload Download Upload

15 % 15 000 544 23 576 3 328 20 000 10 000

20 % 13 000 800 21 576 2 944 20 000 10 000

50 % 10 000 2 584 18 576 2 548 20 000 10 000

15 % 12 000 2 584 14 576 2 548 20 000 10 000ADSL2 and ADSL 2+

ADSL 2+Fiber or other

broadband


Peer Coordination• Impact of the peer bandwidth

• Uplink constrained for poor and middle-class• Increased download for poor peers• Middle-class close to necessary bandwidth

July 18, 2012

t [day]

Dow

nloa

d[G

bps]

1 2 3 4 5 6100

200

300

400Poor Middle Rich

t [day]

Upl

oad

[Gbp

s]

1 2 3 4 5 6100

200

300

400Poor Middle Rich

t [day]

Dow

nloa

d[G

bps]

1 2 3 4 5 60

1000

2000Poor Middle Rich

t [day]U

ploa

d[G

bps]

1 2 3 4 5 60

500

1000Poor Middle Rich

Committed peer bandwidth Free peer bandwidth

Always free bandwidth

Free bandwidth

only for rich

Poor peers may not support many

neighbors

Middle-class close to rich


Peer Coordination• Streaming overhead

• Overhead diminishes for rich peers• Server contribution significant only for resource

limited• Inactive sessions use a small fractionJuly 18, 2012

Band

wid

th(n

orm

aliz

edto

dow

nloa

dpr

imar

y)

Download Download secondary Upload server0

0.5

1

1.5

2

2.5PoorMiddleRich

Band

wid

th(n

orm

aliz

edto

dow

nloa

dpr

imar

y)

Upload Upload active Upload inactive0

0.5

1

1.5

2PoorMiddleRich

Peer download and server Peer upload

Secondary streams

Inactive sessions

1 2Overhead of the

secondary streams

Server uploadInactive sessions


Performance EvaluationPart V

July 18, 2012


Experimental Setting

July 18, 2012

Performance evaluation using computer simulations

Simulation

Emulation

Planet Lab

Scalability Complexity Accuracy Reproducibility

Medium

Low

High

High

Medium

Low

High

Medium

Low

Low

High

High

• Large-scale measurements with up 100 000 viewers

• We implemented two distinct delay models[1,2,3]

• Deterministic and queuing [1] Pesch et al., Performance evaluation of SIP-based multimedia services in UMTS, 2005[2] Ulvan et al., Analysis of Session Establishment Signaling Delay in IP Multimedia Subsystem, 2009[3] Munir, Analysis of SIP-based IMS session establishment signaling for WiMax-3G networks, 2008


Delay and Churn• System performance from the end-user

perspective

• Depends on model of signaling processing at IMS functions

• Below the threshold of acceptable viewer experience[1]

• Churn eliminated for over 90% of channel sessions

July 18, 2012

Min

imum

dela

y[s

]

Deterministic model Queueing model0

0.5

1PoorMiddle-classRich

Max

imum

dela

y[s

]

Deterministic model Queueing model0

0.5


Churn events per session

Frac

tion

ofse

ssio

ns

0 1 2 3 4 5 6 7 8 90

0.5


Churn rate (normalized) [s−1]

CDF

10−4 10−2 1000

0.5


10−3 10−2 10−10.6

0

Channel connection delay Churn performance

[1] Kooij et al., Perceived quality of channel zapping, 2005

Less than 700 ms

The remaining due to users turning off their

UE


Hybrid Streaming

July 18, 2012

Bandwidth poor scenarios are difficult for P2PTV

• Peer uploading bandwidth is a limiting factor• Broadcast server participation remains significant

Use P2P in conjunction with IP

multicast?

• Scalability with the number of TV channels

• IP multicast for popular TV channels

• P2P for unpopular TV channels


Hybrid Streaming• Comparison of bandwidth usage

• Use IP multicast for most popular channels• Use peer-to-peer for the majority of unpopular

channels• Always have some overhead due to secondary

streamsJuly 18, 2012

TV channels using IP multicast

BC,m

[Gbp

s]

0 70 140 210 280 350 420 490 560 630 7000

20

40

ModelSimulation


BC,u

[Gbp

s]

0 70 140 210 280 350 420 490 560 630 7000

50

100

150ModelSimulation


BA

,u[G

bps]

0 70 140 210 280 350 420 490 560 630 7000

20

40

60ModelSimulation


BA

,d[G

bps]

0 70 140 210 280 350 420 490 560 630 70010

20

30

40ModelSimulation

Core network bandwidth Access network bandwidth

Benefit for popular

channels

Use P2P for unpopular channels

Always overhead in the access


Hybrid Streaming• Comparison of scalability issues

• Number of routing entries approximated as

July 18, 2012


ne,

IGM

P

0 70 140 210 280 350 420 490 560 630 7000

4

8

ModelSimulation


ne,

PIM

0 70 140 210 280 350 420 490 560 630 700

×104

0

5

10

15

ModelSimulation

×1


BS

[Gbp

s]

0 70 140 210 280 350 420 490 560 630 7001.38

1.39

1.4

1.41

ModelSimulation

TV channels using IP multicastB

[Gbp

s]

0 70 140 210 280 350 420 490 560 630 7000

100

200ModelSimulation

Multicast entries Server and total bandwidth

Multicast routing entries

Server usage in rich

scenarios

Gain limited for unpopular channels

,IGMPe ii

n pu

M }

,PIM-SMe m ii

n l pu

M }


Conclusions

July 18, 2012


Service architecture

Summary of Contributions

July 18, 2012

Multi-layer streaming

1

2

Mobility support

Workload generator

User activity model

Synthesis algorithms

Enhancements

Fast signalingBandwidth assignmentPerformance

evaluationBasic concept

Extended scenarios

3

4

5

6

7

8

9

10

11

12

Bikfalvi et al., Nozzilla: A Peer-to-Peer IPTV Distribution Service for an IMS-based NGN, ICNS, 2009

Bikfalvi et al., A Peer-to-Peer IPTV Service Architecture for the IP Multimedia Subsystem, IJCS, 2009

Vidal et al., Enabling Layered Video Coding for IMS-based IPTV Home Services, IEEE Network, 2009

Vidal et al., Supporting Mobility in an IMS-based P2P IPTV Service, Computer Communications, 2010

Bikfalvi et al., P2P vs. IP Multicast: Comparing Approaches to IPTV Streaming Based on TV Channel Popularity, Computer Networks, 2011

Submissions in progress


Future Enhancements

July 18, 2012

Extending the Experimental Evaluation• Assessment of playback-quality and viewer

experience• Consider packet-based flows• Include codec characteristics

3

Enhancing the Peer Coordination• Peer churn when viewers turn off their

equipment• Estimate the reliability of user connection• Use this information during the peer selection

2

Enhancing the Session Coordination• Peer churn not included in evaluation of inactive

sessions• Use a feedback loop over the peer coordination• Superior blocking ratio performance

1

Thank You


Backup Slides

July 18, 2012


Enhancement Algorithms• Fast signaling• Reduced peer churn• Support for mobility

Peer-to-Peer• Uploading by user

equipments• Inherent scalability• Complexity and churn

IP Multimedia Subsystem• Convergence of services• Guaranteed quality of

service• Session signaling using SIP

Internet Protocol Television• Audio/video streaming• Commercial grade service• Open platform vs. walled

gardens

Summary• The main topics covered by this dissertation

July 18, 2012


Multiple Trees• Addresses the issues of single trees[1]

• Improves peer participation• Increases robustness to peer churn

July 18, 2012

Interior peers for blue stream

Broadcast server

Interior peers for orange stream

Color Key Blue stream Orange stream

Download Upload

2 streams 2 streams

[1] Castro et al., SplitStream: high-bandwidth multicast in cooperative environments, 2003


Data-Driven Streaming• Focus on video streaming

• Missing video pieces (segments, chunks)

• Heterogeneous bandwidth and delay

• Mesh overlay• Unstructured

protocol

July 18, 2012

Streaming buffer

Playback point

Color Key Missing video segment Available video segment

Broadcast server

Peer

Segment

Segment buffer bitmap

Scheduled segment requests

1

2

Uploading peer Downloading peer


Walled Gardens

July 18, 2012

Color Key All TV channels 1 or 2 TV channels Physical link

DSLAM

Viewer

Set-top box

IP router with multicast enabled

Telco core network

Broadcast server

TV set

Residential gateway

PC

Phone

Customer premise

Commercial IPTV offering a classic viewing experience

• Static IP multicast for all TV channels, unsustainable for future growth[1]

[1] Cha et al., On next-generation telco-managed P2P TV architectures, 2008


Detailed Architecture

July 18, 2012

PCRF

P-CSCF I-CSCF

S-CSCF

P2PTV-AS

SLF

HSS

SGSN

IP CN

Gn

Gi

Gi

Gx

Rx

Mw Mw

Cx

Cx Dx

Dx

ISC Sh

UE (STB)

Gm

GGSN Mobile UE

Viewer

TV set DSLAM

GW

BS

Xd

DSL access network

UMTS access networkUE – IMS

interface (Gm)

P2PTV-AS – BS interface


Application Server

July 18, 2012



INVITE sip:[email protected]: sip:[email protected]: sip:[email protected]: 1000

User Service ProfileInitial Filter CriteriaFilter CriterionFilter Criterion

Filter Criterion

Trigger Point

Application Server

Request URI

=sip:[email protected]

sip:p2ptv-as.example.netUE


P2PTV-AS


1

2

S-CSCF

UAS

3

User agent server• Sessions terminating

at the P2PTV-AS• Streaming by the

broadcast server


Application Server

July 18, 2012



1


P2PTV-AS


User agent client• Sessions originating

at the P2PTV-AS• Inactive uploading

sessions

S-CSCF

UAC

UE



Session Establishment

July 18, 2012

Multimedia sessions with the Session Initiation Protocol

P-CSCF S-CSCF

INVITE INVITE

INVITE sip:[email protected]: sip:[email protected]

Evaluation of the Initial Filter

Criteria

1

2


UE

sip:as.example.net

AS

INVITE

Service control3



Authorize resources

4


Session Establishment

July 18, 2012

Multimedia sessions with the Session Initiation Protocol

P-CSCF S-CSCF


UE

sip:as.example.net

AS

Resource reservation

5200 OK200 OK200 OK

PRACK PRACK PRACK

UPDATE UPDATE UPDATE200 OK200 OK200 OK200 OK200 OK200 OK

ACK ACK ACK

6

Content

Server

Multimedia content


Session Signaling

July 18, 2012


AS

• The UE peers maintain a multimedia session for each downloaded or uploaded stream

UEup


INVITE INVITE1

INVITE sip:[email protected]: sip:[email protected]=stream 100

Evaluation of the Initial Filter

Criteria

2

INVITE

Selection of the uploading peer

3

INVITEINVITE sip:[email protected]: sip:[email protected]


Session Signaling

July 18, 2012


ASUEup




Authorize resources

5



6

PRACK PRACK PRACKPRACK


7

200 OK200 OK200 OK200 OK

Stream 100


Business Model• The P2PTV relies on unused network

capacity• Bandwidth available in the access network• But not contracted by the customer

July 18, 2012

Computer

TV set and STB

IMS phone

ADSL modem/RGW

DSLAM

IP network

Subscriber household An ADSL example

Requires an agreement between the P2PTV and the transport provider

(telco)Multiple IPTV content providers may use the

P2P streaming infrastructure

Service Level Agreement

Downlink Uplink

Voice Internet P2PTV


Business Model

July 18, 2012

IPTV content providers

P2PTV provider

Transport provider (telco)

Service package

Subscriber

The NGN business model facilitates the convergence of multiple services

Customers…• Have a contract with a service packager (telco)• Service providers establish relationships with the

transport provider


Streaming Architecture

July 18, 2012

We propose a push-pull mechanism

• A trade-off between data-driven and session requirements

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

99

98

114 115

116

109 110

111

Playback buffer

Playback point

Peer (UE)

1Pull: request streaming from segment 109

2 Push: send segments in order

• Prevents playback gaps and interruptions• Helps synchronizing multiple streams of the same

TV channel• Streaming boost for buffer under runs


P2PTV for Mobile Devices

July 18, 2012

PCRF

P-CSCF I-CSCF

S-CSCF

AS

SLF

HSS GGSN SGSN

IP network

UTRAN

Gn

Gi

Gi

Gx

Rx

Mw Mw

Cx

Cx Dx

Dx

ISC Sh

Control plane

Application plane

Transport plane

UE

Gm

• A mobile user equipment in an UMTS access network

• Mobile UEs do not upload content due to limited bandwidth


Proactive Context Transfer

July 18, 2012

MIIS

PoS PoS

PoS

PoS

Mobile UE

Current network

Surrounding networks

1Contact MIIS to obtain the surrounding networks

2Contact PoS to obtain the

information on the surrounding networks

3Upon the interaction with the PoS from the surrounding networks

Obtaining the destination configuration


• Include weekly pattern using a modulating function

• Stochastic properties• Difference between trace and model: normal

distribution

• Based on the number of online viewers

• The new session rate is

• Where we have

* ( ) ( ) ( )on onx t M t x t

* ( ) ( ) ( ( ))1 onoff offx t M t x t x t

Session Rate

July 18, 2012

t [day]

r on

0 1 2 3 4 5 6 70.5

1

1.5Fraction of online viewersContinuous and fundamental

onr

0 1 ,1cos 2 w YY Y f t

0 1( ) cos 2 w MM t M M f t

* *,on onX x N * *

,off offX x N ,ch chX x N


Channel Popularity• Selection of the TV channel at a channel

change[1]

• Popularity model for target switching• Mean: zipf/exponential distribution• Instantaneous: mean reversion model

July 18, 2012

Channel selectio

n

Target : 44 %tqSequential : 56 %sqTurn-on : 4 %s oqqBrowsing : 52 %s cqq

Browsing forward : 37 %s c fqqqBrowsing backward : 15 %s c bqqq

[1] Qiu et al., Modeling channel popularity dynamics in a large IPTV system, 2009

( ) ( ) ( )tP i P i P i


Workload Evaluation• Session rate :

• Keeps the dominant low and high frequency components

• Adds the weekly modulation and stochastic properties

July 18, 2012

t [h]

x on

[s−

1]

0 4 8 12 16 20 24

×10−3

0

0.5

1

1.5Trace

t [h]

x on

[s−

1]

0 4 8 12 16 20 24

×10−3

0

0.5

1

1.5Synthetic (Thursday)Synthetic (Sunday)

t [h]

x off

[s−

1]

0 4 8 12 16 20 24

×10−4

0

2

4Trace

t [h]

x off

[s−

1]

0 4 8 12 16 20 24

×10−4

0

2

4Synthetic (Wednesday)Synthetic (Sunday)

onx offxDeviation due to the weekly modulating

function

Online session rate Offline session rate


Workload Evaluation• Session rate :

• Keeps the dominant low and high frequency components

• Adds the weekly modulation and stochastic properties

July 18, 2012

chx

t [h]

x ch

[s−

1]

0 4 8 12 16 20 24

×10−3

0

2

4

6Trace

t [h]

x ch

[s−

1]

0 4 8 12 16 20 24

×10−3

0

2

4

6Synthetic (Wednesday)Synthetic (Sunday)

Very high frequency components are lost

Channel session rate


• Session length :

• Follows closely the model probability distribution• Errors due to limited data and algorithm

approximations

doffCC

DF

1−

F(d

off)

100 1050

0.2

0.4

0.6

0.8

1TraceSynthetic

don

CCD

F1

−F

(don

)

100 1050

0.2

0.4

0.6

0.8

1TraceSynthetic

Workload Evaluation

July 18, 2012

ond offdError due to the trace mismatch with the

session rate

Error due to the workload generation algorithm

Online session length Offline session length


• Session length :

• Follows closely the model probability distribution• Errors due to limited data and algorithm

approximations

Workload Evaluation

July 18, 2012

chd

dch

CCD

F1

−F

(dch

)

100 1050

0.1

0.2

0.3

0.4

0.5

0.6

0.7TraceSynthetic

Channel session length


• Number of online viewers :

• Number of online viewers determined from the session rate

• Pattern similar to the trace data

DayThu Fri Sat Sun Mon Tue

u on

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5TraceSynthetic

Workload Evaluation

July 18, 2012

onu

Some days approximated poorly

Peaks lost due to missing of high frequency rates

Some days approximated well

Number of online viewers (normalized)


• Session event error :

• Approximations introduced by the synthesis algorithm

• Very small, do not substantially affect the workload

on

CDF

10−6 10−4 10−2 100 1020

0.5

1

off

CDF

10−6 10−4 10−2 100 1020

0.5

Workload Evaluation

July 18, 2012

onò offòChannel on eventOnline event

onòOffline eventChannel off event

offò

Less than 1.5 % of errors are greater than

1 second

Session event error


Simulwatch• Session length• Use the Simulwatch model parameters by Qiu et

al.[1]

July 18, 2012

[1] Qiu et al., Modeling user activities in a large IPTV system, 2009

Parameter p1 p2 p3 λ1 λ2 λ3

don 0.3 0.66 0.041.3 10-2

3.3 10-3

2.3 10-4

doff 0.19 0.75 0.063.2 10-2

2.5 10-3

2.4 10-4

dch 0.23 0.64 0.13 2.12.6 10-2

3.2 10-3


Simulwatch• Session length• Use the Simulwatch model parameters by Qiu et

al.[1]

July 18, 2012

[1] Qiu et al., Modeling user activities in a large IPTV system, 2009Online session length Offline session length

Parameter p1 p2 p3 λ1 λ2 λ3

don 0.3 0.66 0.041.3 10-2

3.3 10-3

2.3 10-4

doff 0.19 0.75 0.063.2 10-2

2.5 10-3

2.4 10-4

dch 0.23 0.64 0.13 2.12.6 10-2

3.2 10-3

t [s]

CCD

F1

−F

(don

)

100 10510−15

10−10

10−5

100

TraceOur modelSimulwatc

t [s]

CCD

F1

−F

(doff

)

100 10510−15

10−10

10−5

100

TraceOur modelSimulwatc


Simulwatch• Session rate• Power spectrum is a Weibull distribution• Individual spikes at frequencies of 1 hour, 30

min, 15 min

July 18, 2012

1( / )2( e)

kk

fx

fkX f

Paramete

r k μ 1 hour 30 min

xon0.003

6 278 1.76 1.41


Simulwatch• Session rate• Power spectrum is a Weibull distribution• Individual spikes at frequencies of 1 hour, 30

min, 15 min

July 18, 2012

1( / )2( e)

kk

fx

fkX f

Paramete

r k μ 1 hour 30 min

xon0.003

6 278 1.76 1.41

f [min−1]

Xx

60min 30min

0 0.02 0.04 0.06 0.08 0.110−2

100

102

Trace: Xx = | F(x)|Simulwatch

f [min−1]

φx

0 0.02 0.04 0.06 0.08 0.1−π

0

πTrace: φx =arg F(x)Simulwatch

Session rate spectrum

No phase

t[h]

x’ on

0 3 6 9 12 15 18 21 240

0.02

0.04

0.06

0.08

0.1

0.12Trace: x’on (t) = F −1(F (xon (t)))Simulwatch: x’on (t) = F −1(Xx)

x60min

x30min

x60min+x3

Session rate


Session Coordination

July 18, 2012

Keep the channel blocking ratio at a desired reference

Without going into much details…

However…• The blocking ratio is a slow metric• Cannot respond effectively to spikes in the

channel activity

• We assume the systems are linear and time invariant

• Proportional-integral controllers

Fast control loop1

Slow control loop2

• Output:• Reference:

( ) ( ) ( )ffy k w k u k

( ) 0fr k

• Output:• Reference:

( ) ( )sy k b k

( )sr k Desired blocking

ratio

0

( ) ( ) ( )P I

k

i

w K Kk e k e i


Session Coordination• Priority versus non-priority mode• The slow control loop drives the blocking ratio

b(k) toward the reference β• Non-priority: disable the control loop when

July 18, 2012

Block

ing

ratio

b


×10−3

0

0.5

1

1.5

2β =10 −3

β =10 −4

β =10 −5

β =0

Block

ing

ratio

b

Channel 1 Channel 10 Channel 100 Channel 500

×10−4

0

2

4

6

8

10β =10 −3

β =10 −4

β =10 −5

β =0

Priority mode Non-priority mode

( )b k

Allow blocking ratio smaller than

the reference


Session Coordination• Priority versus non-priority mode• It has a limited impact on session utilization• Non-priority mode preferred for better

performance

July 18, 2012

Util

izat

ion

ρ


0.9

1

β =10 −3

β =10 −4

β =10 −5

β =0

Unu

sed

sess

ions


50

100β =10 −3

β =10 −4

β =10 −5

β =0

Util

izat

ion

ρ


0.9

1

β =10 −3

β =10 −4

β =10 −5

β =0

Unu

sed

sess

ions


50

100β =10 −3

β =10 −4

β =10 −5

β =0

Priority mode Non-priority mode


Peer Coordination• Peers are organized in peer pools

• Double mapping between peer and its resource level

July 18, 2012

Global pool

Stream pool

Inactive pool

All peers with free bandwidth

1

2

3

Peers with free bandwidth downloading the corresponding stream

Peers with inactive sessions for the corresponding stream

0.05

0.11

0.19

0.22

0.28

0.30

0.35

0.35

0.25

0.36

0.38

0.41

001

055

098

124

354

364

457

612

791

830

874

965

Resource level

Peer identifier


Peer Coordination• Peer selection based on pool membership

July 18, 2012

Channel connect

Inactive pool

Stream pool

Broadcast server

Channel recovery

Stream pool

Broadcast server

Inactive session

Stream pool

Global pool


• Create a model of the bandwidth usage• Function of the number of multicast channels set

M and unicast channels set U

Hybrid Streaming

July 18, 2012

Bandwidth

Server S si

B b

i M UCore network

Core network multicast

Stream bandwidth

,C m m i si

B l pu b

M i

Multicast tree size

Channel popularity

Core network unicast , 1C u u i si

B l pu b

Ui

Unicast path length

Peer-to-peer

overhead

Access network

Access network upload , 1A u i si

B pu b

UiAccess network download


• Create a model of the bandwidth usage• Function of the number of multicast channels set

M and unicast channels set U

Hybrid Streaming

July 18, 2012

Bandwidth

Access network

Access network upload , 1A u i si

B pu b

UiAccess network download

Primary streams , ,A d p i si

B pub

Ui

Secondary streams , ,A d s i si

B pu b

Ui

Multicast streams , ,A d m i si

B pub

M i

peer-to-peer television for the ip multimedia subsystem

Documents