Infocom 2006, April 23-29

1D. Pompili, L. Lopez, C. Scoglio, Multicast Algorithms in Service Overlay Networks, Infocom 2006

Dario Pompili*, Luca Lopez^, Caterina [email protected], ^[email protected], [email protected]

*Broadband and Wireless Networking LaboratoryGeorgia Institute of Technology, Atlanta, GA 30332, USA

^Dipartimento di Informatica e SistemisticaUniversity “La Sapienza,” Rome 00184, Italy

+Department of Electrical and Computer EngineeringKansas State University, Manhattan, KS 66506, USA

Multicast Algorithms in Service Overlay Networks

Infocom 2006, April 23-29


Outline

Introduction to Overlay Networks Native Network vs. Virtual Overlay Network Unicast vs. Multicast Transmissions Multicast Applications Proposed Overlay Multicast Algorithms: DIMRO

and DIMRO-GS Performance Evaluation Conclusions and Future Work


Introduction to Overlay Networks

Overlay routing enhances IP network reliability and performance by forwarding traffic through intermediate overlay nodes [1-3]. This way:– It can bypass congestion– It can overcome transient outages

Past research [4,5] focused on techniques for:– Building overlay networks– Evaluating their performance

[1] H. Zhang, J. Kurose, and D. Towsley, “Can an overlay compensate for a careless underlay?” in Proceedings of IEEE INFOCOM 2006, Barcelona, Spain, Apr. 2006

[2] Y. Zhu, C. Dovrolis, and M. Ammar, “Dynamic overlay routing based on available bandwidth estimation: A simulation study,” To appear in the Computer Networks Journal, 2006

[3] S. Y. Shi and J. S. Turner, “Routing in overlay multicast networks,” in Proceedings of IEEE INFOCOM 2002, New York, NY, USA, June 2002

[4] D. Andersen, H. Balakrishnan, M. Kaashoek, and R. Morris, “Resilient overlay networks,” in Proceedings of ACM Symposium on Operating Systems Principles, Banff, Canada, Oct. 2001

[5] A. Nakao, L. Peterson, and A. Bavier, “A routing underlay for overlay networks,” in Proceedings of ACM SIGCOMM, Karlsruhe, Germany, Aug. 2003


Native Network vs. Virtual Overlay Network

We consider two layers of network infrastructure: – The native network, includes end-systems, routers, links, and

the associated routing functionality, and provides best-effort datagram delivery between its nodes

– The virtual overlay network, formed by a subset of the native layer nodes interconnected through overlay links to provide enhanced services

Overlay links are virtual in the sense that they are IP tunnels over the native network


Unicast vs. Multicast Transmissions Unicast transmissions:

– The sender transmits data to a single receiver and, if multiple receivers want the same data content, the sender has to transmit multiple copies of data

Multicast transmission:– The sender transmits only one copy of data that is delivered to multiple

receivers

A challenging objective in multicasting is to minimize the amount of network resources to compute multicast trees

Unicast Multicast

3 copies of the same message

Only 1 copy of the

message

S

R2

R3 R3

R2

S


Multicast Applications

We present two algorithms to build virtual multicast trees on an overlay network for as many applications as:– Live video, software and file distribution, replicated database, web site

replication, videoconference, distributed games, file sharing, periodic delivery

Source Specific Group Shared

Real-time Live video distributionVideoconference, distributed games

Non Real-time

Software and file distribution, replicated database, server and web site replication,

periodic data delivery (sport scores, magazines, newspapers)

File sharing, collaborative

groupware, replicated database


Proposed Overlay Multicast Algorithms

DIMRO: DIstributed Multicast algorithm for Internet Resource Optimization – It builds virtual source rooted multicast trees for source specific

applications

– It takes the virtual link available bandwidth into account to avoid traffic congestion and fluctuation, which cause low network performance

DIMRO-GS: DIstributed Multicast algorithm for Internet Resources Optimization in Group Shared applications– It constructs a virtual shared tree for group shared applications

– It connects each member node to all the other member nodes with a source rooted tree computed using DIMRO

Both DIMRO and DIMRO-GS algorithms offer service differentiation, i.e., they provide QoS at the application layer without IP-layer support


DIMRO: Objectives

Most of the algorithms in the literature focus on computing multicast trees for real-time applications (delay sensitive)– Bandwidth is often taken into account only as a constraint

– Optimization involves only one tree, and not all trees for different multicast groups

DIMRO optimizes the overlay network performance when multiple multicast comunications should take place

The objective is to exploit the network resource in an efficient way

The available bandwidth on each virtual link plays a key role in the multicast tree research


DIMRO: Motivations

If the overlay available bandwidth is not esplicitely taken into account the risk is to:– Overload some link

– Leave some other link unexploited

Load balancing is NOT achieved if cost does not explicitely take available bandwidth on links into account

R1A

SA

R2AR1B

R2B

SB

R3B

Overloaded Link

Unexploited Link


DIMRO: Steps of the Algorithm1. Receivers are ordered according to decreasing

requested bit rates 2. The optimum path between the sender and the

receiver with the highest requested rate is computed

3. Then, paths connecting other receivers are computed, according to the decreasing order

4. The kth receiver rk is connected to the sender s by using that path p(s,rk) that minimizes the following objective function:

)},(),{( 1),(

krspvu uv

uvk

arspf

r1

1024 Kbps

r2

512 Kbps

r3

256 Kbps

s 0 00

0

0

0

Step 1Step 1 Step 1

Step 2Step 3

Step 3

uv

uvuv

uv

kuvuvuv B

B

B

FbB

)(

• Buv: total bandwidth of virtual link (u,v)• buv: available bandwidth of virtual link (u,v) • Fk: cumulative rate by the kth receiver rk

• auv: binary variable that equals 0 if link (u, v) already belongs to the tree, 1 otherwise

• V and E: number of vertexes and edges in the overlay network

Tvu

Tvuauv ),(,1

),(,0

),|,||,(| BFEV


DIMRO-GS: the Algorithm

DIMRO-GS (Group Shared DIMRO) is the extension of DIMRO to the multicast shared tree case

For M group members, the virtual shared tree is set up by building M virtual source rooted trees, each one having a different group member as root and all the other members belonging to the tree

Each source rooted tree is built using DIMRO If each group member has the same bandwidth requirement, the

first step of DIMRO is skipped When M source rooted trees are computed, the virtual shared tree is

completed

DIMRO computational complexity: O(M · |V| · |E|) – It builds the virtual multicast tree by computing for as many as M times

the spanning tree using the Bellman-Ford algorithm, whose complexity is O(|V| · |E|)

DIMRO-GS computational complexity: O(M2 · |V| · |E|) – It runs DIMRO M times


Simulation Scenario

A random overlay network has been generated following the Waxman’s model [6], with parameters (α,β)

The bandwidth capacity Buv of each virtual link (u, v) is randomly generated using a uniform distribution with mean B = 100Mbps

Two different one hundred node random overlay networks are generated

Network 2 has a higher number of links than Network 1

Two metrics are used to compare the competing algorithms, the Rejection Rate and the Network Load

α β B nodes

Network 1

0.2 0.4 100Mbps 100

Network 2

0.3 0.6 100Mbps 100

),(

Re#

Re#Re

vuuv

Treesquested

TreesjetedRatejection

[6] B. M. Waxman, “Routing of multipoint connections,” IEEE Journal on Selected Areas in Communications, vol. 6, no. 9, pp. 1617–1622, Dec. 1988


DIMRO: Performance Evaluation in Network 1

In Network 1, the DIMRO Rejection Rate is lower than the Rejection Rate of the Optimal solution of the Steiner Tree Problem (OSTP) with cuv= 1The DIMRO Network Load is the same as the OSTP Network Load until the Rejection Rate is the same. Then, since DIMRO rejects less trees, its Network Load becomes higher than the OSTP one


DIMRO: Performance Evaluation in Network 2

In Network 2, the DIMRO Rejection Rate is significantly lower than the OSTP Rejection Rate (less unavoidable bottlenecks exist). Since Network 2 has a higher number of links, the number of possible paths between two nodes increases. Thus, it is easier for DIMRO to avoid bottlenecksA lower Network Load with a higher Rejection Rate proves that OSTP does not efficiently use the available network resources


DIMRO-GS: Performance Evaluation in Network 1

When the Number of Requested Trees increases, the FTM Rejection Rate becomes significantly higher than the one of DIMRO-GS, since FTM uses a higher amount of network resourcesIn fact, when the Number of Requested Trees increases, the FTM Network Load becomes significantly higher than the DIMRO-GS Network Load


DIMRO-GS: Performance Evaluation in Network 2

Network resources saturate later when DIMRO-GS is used, which causes a lower Rejection RateIn Network 2, DIMRO-GS achieves a lower Rejection Rate and Network LoadNote that in Network 1 bottlenecks do not allow efficiently exploiting all available network resources


Conclusions and Future Work

Two algorithms for multicast applications in service overlay networks were presented

The first builds virtual source rooted multicast trees for source specific applications

The second constructs a virtual shared tree for group shared applications

Their objective is to achieve traffic balancing on the overlay network to avoid traffic congestion and fluctuation, which cause low network performance

The algorithms actively probe the underlay network and compute virtual multicast trees by dynamically selecting the least loaded available paths on the overlay network

Future research will focus on dynamic multicast groups, and on the interactions between the overlay and the underlay network

Infocom 2006, April 23-29

Documents

Transcript of Infocom 2006, April 23-29