Scalable Group Communications and Systematic Group Modeling

Jun-Hong Cui University of Connecticut

jcui@cse.uconn.edu http://www.cse.uconn.edu/~jcui

Cool Application 1 : Teleconferencing

Cool Application 2 : Telemedicine

Cool Application 3 : Net Gaming

Jun-Hong Cui (c) UCONN 2004

Multicast: What and How? Multicast:

One to many or many to many communications (group communications)

To achieve multicast: Multiple unicast (one to one) Network multicast---IP multicast Overlay multicast (using proxies) Application layer multicast (end host)

Jun-Hong Cui (c) UCONN 2004

Outline of this talkScalable Group Communications--- Aggregated Multicast

Systematic Group Modeling--- GEM Model

Research Directions

Jun-Hong Cui (c) UCONN 2004

IP Multicast Group: IP D class address

Use Tree delivery structure

Routers: keep forwarding entries per-group/source (multicast state)

IP multicast Resource efficient Scalable to group size

Customer Networks, Domain D

D1Domain X

Domain A

Domain C

Domain Y

X1

Ab

AaA1

A2Domain B

B1

A3

C1Y1

group NHopg1 Ab, A3

Jun-Hong Cui (c) UCONN 2004

The Problem: Not Scalable to the Number of Groups

More groups more trees

More forwarding entries More tree maintenance

overhead IP multicast NOT scalable

to the number of groups State Scalability problem Serious in transit domains

Our solution Aggregated multicast to

improve state scalabilityCustomer Networks, Domain D

D1Domain X

Domain A

Domain C

Domain Y

X1

Ab

AaA1

A2Domain B

B1

A3

C1Y1

group NHopg1 Ab, A3g2 Ab, A3

Jun-Hong Cui (c) UCONN 2004

Key Insight

There are many overlaps among multicast trees in transit domains

Customer Networks, Domain D

D1Domain X

Domain A

Domain C

Domain Y

X1

Ab

AaA1

A2Domain B

B1

A3

C1Y1

group NHopg1 Ab, A3g2 Ab, A3

Jun-Hong Cui (c) UCONN 2004

Aggregated Multicast Key idea:

Force multiple groups share a single delivery tree (aggregated tree)

Benefits: Reduce state at core

routers Reduce tree

maintenance overhead Push complexity to

edge Target:

Multicast provisioning in transit domainsCustomer Networks, Domain D

D1Domain X

Domain A

Domain C

Domain Y

X1

Ab

AaA1

A2Domain B

B1

A3

C1Y1

Tree NHopT1 Ab, A3

Jun-Hong Cui (c) UCONN 2004

Aggregated Multicast (cont.) Core routers:

Keep state per-tree Edge routers:

Keep group state Groups:

Aggregate at incoming edge router

De-aggregate at outgoing edge routers

Customer Networks, Domain D

D1Domain X

Domain A

Domain C

Domain Y

X1

Ab

AaA1

A2Domain B

B1

A3

C1Y1

Tree NHopT1 Ab, A3

Aggregation

De-aggregation

De-aggregation

Jun-Hong Cui (c) UCONN 2004

Perfect Match vs. Leaky Match

Group-Tree match Perfect match Leaky match

Bandwidth waste in leaky match Data delivery to non-member nodes

Customer Networks, Domain D

D1Domain X

Domain A

Domain C

Domain Y

X1

Ab

AaA1

A2Domain B

B1

A3

C1Y1

Tree NHopT1 Ab, A3

Discard Packets

Jun-Hong Cui (c) UCONN 2004

Aggregation Control Leaky match

Good for tree aggregation But waste bandwidth

There is a trade-off Static group-tree matching: NP hard A dynamic group-tree matching algorithm to control the trade-off Under a given bandwidth waste threshold

Jun-Hong Cui (c) UCONN 2004

Group-Tree Matching

Domain X

Domain A

Customer Networks, Domain D

Domain C

Domain Y

X1

AbAa

D1

A1

A4

Domain BB1

A3

C1Y1

Domain E E1 A2

Jun-Hong Cui (c) UCONN 2004

Group-Tree Matching

Domain X

Domain A

Customer Networks, Domain D

Domain C

Domain Y

X1

AbAa

D1

A1

A4

Domain BB1

A3

C1Y1

Domain E E1 A2

Jun-Hong Cui (c) UCONN 2004

Implementation Issues Multiplex multiple groups over a shared tree

IP encapsulation MPLS (Multi-Protocol Label Switching)

Tree management and group-tree matching Tree Manager (need to know group membership) Distributed or centralized solutions

Have designed and implemented protocols: ASSM for source specific multicast (SSM) BEAM for shared tree multicast (ASM) AQoSM for QoS multicast provisioning

Jun-Hong Cui (c) UCONN 2004

Extend to Overlay and Adhoc Net

Overlay multicast Implement multicast in overlay net

A collection of proxies (or gateways) Processing power, memory & bandwidth more critical Aggregated multicast reduces management overhead

Wireless multicast Implement multicast in wireless adhoc net

No infrastructure, self-organized Energy, memory, bandwidth, resilience very critical Aggregated trees help to improve performance

Jun-Hong Cui (c) UCONN 2004

Overlay Network

Jun-Hong Cui (c) UCONN 2004

Adhoc Network

Jun-Hong Cui (c) UCONN 2004

Outline of this talkScalable Group Communications--- Aggregated Multicast

Systematic Group Modeling--- GEM Model

Research Directions

Jun-Hong Cui (c) UCONN 2004

The Problem: Group Modeling

The locations of the group membersGiven a graph, where should we place them?

Current assumptions: uniform random model (unproven)All members uniformly distributed Not realistic for many applications

Jun-Hong Cui (c) UCONN 2004

Group Modeling is Critical Some studies have shown the locations of members have significant effects on Scaling properties of multicast trees Aggregatability of multicast state Performance of state reduction schemes

Realistic group models Improve effectiveness of simulation Guide the design of protocols

Jun-Hong Cui (c) UCONN 2004

Our Contributions

Measure real group membership properties MBONE (IETF/NASA) and Netgames (Quake)

Design a model to generate realistic membership GEneralized Membership Model (GEM) Use Maximum Enthropy: a statistical method

Jun-Hong Cui (c) UCONN 2004

Roadmap

Membership Characteristics Measurement and Analysis Results

Model Design and Validation

Jun-Hong Cui (c) UCONN 2004

Beyond Uniform Random Model

How close are the members in a group?

Are all the members same in group participation?

What are the correlations between members in group participation?

Jun-Hong Cui (c) UCONN 2004

An Illustration (Teleconference)

Internet

Edge Router Member Router

Seattle

Boston

AtlantaLos Angeles

0.5

0.5

0.7

0.4

1.0

Jun-Hong Cui (c) UCONN 2004

Membership Characteristics Member clustering

Capture proximity of group members Use network-aware clustering method

Group participation probability Show difference among members/clusters

Pairwise correlation in group participation Capture joint probability of two members/clusters Use correlation coefficient (normalized covariance)

Jun-Hong Cui (c) UCONN 2004

Measure Membership Properties MBONE applications (from UCSB)

IETF-43 (Audio and Video, Dec. 1998) NASA Shuttle Launch (Feb. 1999) Cumulative data sets on MBONE (1997-1999)

Net Games (using QStat) Quake I (query master server) Choose 10 most popular servers (May. 2002)

Examine three membership properties

Member Clustering

CDF of cluster size for MBONE and net games

MBONE cumulative data sets

MBONE real data sets

Net game data sets

(3, 0.64)

Group Participation Probability

CDF of participation probability for Net Game data sets

Group Participation Probability

CDF of participation probability for MBONE applications

Pairwise Correlation in Group Participation

CDF of correlation coefficient for Net Game data sets

Pairwise Correlation in Group Participation

CDF of correlation coefficient for MBONE applications

Jun-Hong Cui (c) UCONN 2004

Generalized Membership Model--- GEM (An Overview)

Network topologyCluster methodGroup behavior

Distr. of participation prob. Distr. of pairwise correlation Distr. of member cluster size

1. Create clusters in given topology2. Select clusters as member clusters According to input distributions3. Choose nodes for each member

clusters

Desired number of multicast groupsthat follow the given distributions

Inputs

GEM

Outputs

Jun-Hong Cui (c) UCONN 2004

Member Distribution Generation

Definition: K clusters: C1 , C2 , … , Ci , … , CK

Prob. pi for any i in [1, K] Joint prob. pi,j for any i, j in [1, K] X=(X1 ,X2 , … , Xi , … , Xk): Xi is a binary indicatorXi = 1 if Ci is in the group Xi = 0 if Ci is not in the group

Objective:Generate vectors x=(x1 , x2 , … , xk) satisfying P(Xi = 1) = pi and P(Xi = 1 , Xj = 1) = pi,j

Jun-Hong Cui (c) UCONN 2004

Maximum Entropy Method To calculate the distribution of (X1,X2, …, Xk) requires O(2K) constraints

But we only know O(K+K2) constraints We use Maximum Entropy Method

Entropy is a measure of randomness We construct a maximum entropy distr. p*(x)

Satisfy constraints in specified dimensions Keep as random as possible in unconstrained

dimensions i.e. maximize entropy while match given constraints

Jun-Hong Cui (c) UCONN 2004

Three CasesConsidering P(Xi=1)= pi and P(Xi=1, Xj=1)=pi,j

1. Uniform distr. without correlation (easy) pi,j = pi * pj , and pi = pj

2. Non-uniform distr. without correlation (easy) pi,j = pi * pj , but pi = pj not necessary

3. Non-uniform distr. with pairwise correlation Neither pi,j = pi * pj nor pi = pj necessary Need to calculate the maximum entropy distr. p*(x)

Entropy decreases from case 1 to case 3

Jun-Hong Cui (c) UCONN 2004

Experimental ValidationConsider all membership properties

Consider three cases Figures omitted …Our experiments show

GEM can regenerate groups satisfying given distributions (from real measurement)

Jun-Hong Cui (c) UCONN 2004

Summary Uniform random model

Can capture net games approximately But not realistic for MBONE applications

GEM: a generalized membership model Three cases (case 1: uniform random model) Realistic membership can be regenerated

Beyond multicast Peer-to-peer network modeling

Beyond wired networkWireless adhoc networks, sensor networks …

Jun-Hong Cui (c) UCONN 2004

Outline of this talkScalable Group Communications--- Aggregated Multicast

Systematic Group Modeling--- GEM Model

Research Directions

Jun-Hong Cui (c) UCONN 2004

Networking: Expanding Visions

(from Jim Kurose)

Peer-to-Peer Networking

Peer-to-Peer NetworkingFocus at the application level

Jun-Hong Cui (c) UCONN 2004

Applications & Challenges Applications

P2P file sharing (Napster, Gnutella, Freenet, etc.) Application-layer multicast

Characteristics each node potentially same responsibility, functionality

logical connectivity rather than physical connectivity Why P2P?

High resource utilization (bandwidth, memory, CPU) Challenges

Self-organized and large scale (routing) Reliability and security

Jun-Hong Cui (c) UCONN 2004

Research Directions Overlay multicast

Scalability, QoS, security, pricing, … Multicast modeling

Systematic multicast evaluation Peer-to-peer networks

measurement & modeling, complex queries Wireless adhoc networks

Mobility modeling, scalable multicast Sensor networks

Sensor deployment and security Very large scale sensor network design

Jun-Hong Cui (c) UCONN 2004

Questions?jcui@cse.uconn.eduhttp://www.cse.uconn.edu/~jcui

Jun-Hong Cui (c) UCONN 2004

THANKS!!!

Jun-Hong Cui (c) UCONN 2004

Network Characteristics No fixed infrastructure, instantly deployable Node portability, mobility Error-prone channel Limited resources

bandwidth, energy supply, memory and CPU. Heterogeneous nodes

big/small; fast/slow etc Heterogeneous traffic

voice, image, video, data Wireless multihop connection

to save power, overcome obstacles, enhance spatial spectrum reuse, etc

Jun-Hong Cui (c) UCONN 2004

Calculate the Maximum Entropy Distribution

dxxpxpxp logmax arg*

jiwhenpdxxpxx jiji ,,

ii pdxxpx

The maximum entropy distr. p*(x) is the solution for:

1 dxxp

Subject to

and and

Use lagrange multipliers and numerical method to construct p*(x), Then Gibbs Sampler to sample it

Group Participation Probability

Participation probability distribution for IETF43-Video

Pairwise Correlation in Group Participation

Joint probability distribution for IETF43-Video