Page 1

A BENCHMARKING SYSTEM FOR MULTIPATH OVERLAY MULTIMEDIA STREAMING

Svetlana Boudko, Wolfgang Leister

Norwegian Computing Center, Norway

Carsten Griwodz, P˚ al Halvorsen

Simula Research Laboratory, Norway

Department of Informatics, Univ. Oslo, Norway

ABSTRACT

The rapid growth of the Internet multimedia services brings new

challenges to how multimedia streams can be delivered to the users

over bandwidth-constraint networks. Different strategies that exploit

multipath streaming in order to provide better utilization of the Inter-

net resources have been proposed by the research community. How-

ever, there exists no metric that allows us to evaluate how close these

strategies are to the optimal resource utilization. This paper proposes

a static benchmarking system that models the best possible distribu-

tion of streams along multiple paths in an overlay network that is

shared by several senders and receivers. We have tested it with sev-

eral different network topologies, and present the test results in this

paper.

Index Terms— multimedia streaming, resource allocation,

overlay networks, adaptation, multipath routing

1. INTRODUCTION

The rapid growth of the various multimedia streaming services pro-

vided on the Internet in combination with the significant increase

of the Internet population brings new challenges to how multimedia

content can be delivered to the end users.

Although bandwidth in the backbone is generally considered

sufficient today, several challenges for high quality multimedia

streams still persist. Problems that can seriously affect multimedia

streaming can be listed as follows: access networks are still not

provisioned enough; slow convergence of routing protocols after

link or node failures; BGP and OSPF, although designed to support

resource efficient streaming, are still not so used.

Multimedia streaming applications are bandwidth consuming

and delay sensitive. As the routing in the Internet is determined by

policies of the routers and cannot be controlled by the end hosts,

there is no guarantee that the path selected to deliver multimedia

streams is provided with sufficient available bandwidth. At the same

time, an alternative path or several alternative paths can be found

in the network, which utilization gives higher throughput and can

significantly improve the quality of the delivered stream. Construct-

ing several paths for data forwarding can then be implemented in

application level overlay networks.

Recently several different strategies [1,2] have been proposed

that intend to exploit multipath delivery for multimedia streaming by

distributing streams between different paths. However, one cannot

evaluate how good these strategies are in terms of optimal resource

utilization when several multimedia senders and receivers share the

same delivery infrastructure and compete for the same network re-

sources. Therefore, we propose a baseline system that provides us

withthebestpossibledistributionofthestreamsoveravailabledeliv-

ery paths given that the complete knowledge of the network includ-

ing its topology and resource availability is obtainable. The baseline

system can then be used to quantify the difference between the op-

timal solution and solutions provided by algorithms that operate in

dynamically changing networking environments with partial knowl-

edge of the network.

The structure of the paper is organized as follows. Section 2

givesanoverviewofrelatedworkrecentlydonebytheresearchcom-

munity. The model for the baseline system is formulated in Section

3. We present the results of stream distribution among multiple paths

for selected network topologies in Section 4. Our conclusion and

discussion of our future work is given in Section 5.

2. MULTI-PATH STREAMING

Multipath streaming is one of the approaches that have been pro-

posed in several papers [3–5] in order to overcome packet loss and

bandwidth limitations in the Internet during the delivery of multime-

dia streams.

There are mainly two ways of implementing multipath routing.

It can be done either in the network layer exploiting for example

traffic engineering techniques [6] or in the application layer using

overlay networks. When implemented in the network layer sev-

eral routing protocols and algorithms can be used including OSPF

Optimized Multipath protocol [7], Multipath Distance Vector Algo-

rithm [8], Quality of Service (QoS) routing mechanisms [9], etc.

However, these algorithms have serious drawbacks. They require

routers to maintain detailed information about all paths between the

router and every possible destination and therefore scale poorly. Im-

plementing multipathing at the network layer also implies a certain

cooperation between ISPs including that all ISPs choose to use the

same routing protocol. In addition, ISPs are unlikely to allow others

to control the traffic going through their networks.

Implementing multipath streaming in the application layer by

using overlay networks helps to overcome the problems in deploy-

mentrelatedtoISPs. However, itiscertainlylessefficientintermsof

latency. CollectCast [10] is an overlay service that provides forward-

ing of streams from multiple senders to one receiver. Operating at

the application level this service is also able to detect and exploit un-

derlying network topology and network performance characteristics.

In SplitStream [11] multiple trees are built and used for streaming in

order to balance the forwarding load. The content is split into several

stripes, which are then multicast using separate trees.

Using application layer overlays for multipath streaming is also

beneficial since not all data in the multimedia stream is equally im-

portant in terms of user-perceived quality. Therefore, depending on

conditions observed along the overlay links an overlay node can per-

form certain processing of the forwarded streams [10] including ap-

plying error correction techniques, selective drop of packets, or for-

warding more important packets along more reliable paths.

The effectiveness of using overlay networks for multipath

streaming depends on their ability to (1) detect the current net-

Page 2

(a) Total throughput be-

tween one receiver and one

sender.

(b) Shared link.

Fig. 1. Link and path constraints.

work conditions such as latency and available bandwidth and (2)

avoid using points of shared congestion.

sources can be detected using network tomography and inference

techniques [12].

Research in defining correlation techniques for detecting shared

congestion has been done by the authors [13]. However, in spite of

the considerable amount of work that has been done in the area, we

have so far found only examples that demonstrate improved perfor-

mance over other ad-hoc mechanisms or over unicast performance.

We have therefore decided to create the benchmark for optimal per-

formance presented in the next section.

Available network re-

3. BENCHMARK

To build our benchmarking system, we model the overlay network

that includes senders, receivers and overlay nodes as a fully meshed

graphDo= (Vo,Ao), whereVoisthesetofverticesthatrepresents

the nodes of the network, and Aois the set of arcs that represent the

overlay links between the nodes. Sets Vo

Vo

sender nodes, overlay nodes and receiver nodes.

In the overlay graph, we define a set of all possible overlay paths

between the receivers and the senders. Note that these paths should

not contain loops. For each overlay path, we build a corresponding

path in the underlying Internet connecting the vertices of the overlay

path by the shortestpaths in the underlying network. Thus, we define

a set of all possible paths P in the underlay. Then, pk

th path of the subset which contains the paths connecting the sender

i with the receiver j. Ki,j denotes the number of paths from the

sender i with the receiver j.

The intersection of all paths P forms the underlay graph D =

(V,A), where V is the set of vertices representing the nodes of the

underlying network and A is the set of arcs representing the underlay

links. For each path p and each ark a in the underlay graph we define

a function δ as follows:

s ∈ Vo, Vo

o ∈ Voand

r ∈ Voare disjoint subsets of vertices representing respectively

i,jdenotes the k-

δ(a,p) =

(

1,if

0,if

a ∈ p

a / ∈ p

(1)

To model network resources we define two functions on the un-

derlay arcs. One is the bandwidth function that expresses the avail-

able bandwidth of the arc a:

b : A → R+

0

(2)

The second one is the latency function that defines the latency of the

arc a:

l : A → R+

0

(3)

The same functions are defined on the paths. The available band-

width of the path p is defined as the lowest bandwidth among all

arcs that belong to the path p:

b(p) = min{b(a)},a ∈ p

(4)

The latency of the path p is the sum of all latencies of the arcs that

belong to the path p:

l(p) =

X

Streaming requirements of the requested streams are defined by

the matrix R, where the matrix element ri,j represents the required

bitrate at which the stream from the sender vi

receiver vj

the above matrix r, but contains the bitrates for the basic layers. The

basic layer represents the lowest acceptable perceived quality of the

multimedia content.

One more matrix used to specify the receivers’ requirements for

stream delivery is matrix D. Its elements show the acceptable delay

for stream arrival to the receivers. The values for these elements can

differ significantly depending on the nature of the streaming content.

Live streaming is more delay sensitive than on-demand streaming or

video downloading.

The stream distribution problem is specified as the following.

The streams requested by the receivers have to be distributed among

the paths in a way that the overall throughput of the system is max-

imized while the delivery of the basic layer is guaranteed. This

problem can be formulated as a linear programming problem, and

to solve it we apply the Simplex method [14].

We introduce the variable xk

i,jthat denotes a share of the mul-

timedia stream sent from the content provider vi

through the path k. To find the best possible distribution of the send-

ing streams among the available paths, we maximize the following

objective function:

a∈p

l(a)

(5)

sis streamed to the

r. In addition, we define the matrix Rbwhich is similar to

Sto the receiver vj

R

max

X

pk

i,j∈P

xk

i,j· ri,j

(6)

Theobjectivefunctionissubjecttothesetofconstraintsgivenbelow.

As not all data in the multimedia stream is equally important for

the perceived quality, we may drop less important data though the

delivery of the basic layer should be guaranteed for all streams:

∀{i,j} :

X

k=0,Ki,j

xk

i,j· ri,j ≥ rb

i,j

(7)

The next constraint is illustrated in Fig. 1(a). It implies that the

sum of sending rates along all paths from one sender to one receiver

should not exceed the bitrate assigned to this stream.

∀{i,j} :

X

k=0,Ki,j

xk

i,j≤ 1

(8)

As depicted in Fig. 1(b) the links in the underlay can be shared be-

tween several delivery paths. We need to specify that the total send-

ing rate does not exceed the available bandwidth of the shared link.

∀{a} :

X

pk

i,j∈P

xk

i,j· ri,j· δ(a,pk

i,j) ≤ b(a)

(9)

Page 3

Fig. 2. Example for Waxman topology.

Finally, the latency of the paths participating in delivering a partic-

ular stream should not exceed the acceptable delay assigned to this

stream.

l(pk

i,j) ≤ di,j

(10)

The objective function and the set of constraints described in

this section define our benchmarking system. This system provides

a reference for dynamic algorithms since it selects paths in a multi-

source, multi-destination streaming scenario in such a way that the

global resource usage is optimal. Although the system is not prac-

tically applicable because of the algorithm’s execution time and the

amount of information, which is needed to be collected in real time

in order to execute the algorithm, it provides a means of compar-

ing the performance of other algorithms with respect to the optimal

choice of paths. We see it as an alternative to simply comparing with

another ad-hoc strategy.

In the following section, we demonstrate the results of the

benchmark for two example scenarios.

4. USAGE EXAMPLE

Our benchmark computes the optimal bandwidth assigment for a

given static network condition and application situation. It is not a

topic of this paper to compare any existing multipath streaming tech-

nique with this benchmark, but we consider it necessary to demon-

strate that the optimal multipath streaming decision does not have a

trivial solution in the general case. We have done this by applying

the benchmark computation to several topologies and overlay node

placements. Here, we show only two examples.

One of the examples is that of two senders, two receiver and

two overlay nodes that were randomly placed in a Waxman topol-

ogy, a typically considered a fairly realistic generation strategy for

networks. The Waxman topology depicted in Fig. 2 was generated

using the Brite topology generator [15]. In the second example, we

placed the same number of nodes randomly in a chessboard topol-

ogy of edge length 6. This topology, the vertex labels and the link

bandwidth resources are depicted in Fig. 3.

We applied the benchmark to these two situations, where both

senders sent one stream to each receiver. For comparison, we com-

puted also the through achievable in a unicast approach that used a

direct shortest path route between each pair of sender and receiver.

Fig. 3. Example for Chessboard topology.

Shortest

path

from

v1

s

v2

s

sending ratesending rate

distribution for Waxman

to v1

r

42.5

87.0

distribution for Chessboard

to v1

r

23.0

25.0

to v2

r

to v2

r

42.5

39.0

25.0

35.0

Table 1. Total sending rate using shortest path routing.

The throughput achieved in the shortest-path unicast case is then re-

ported in Table 1, the throughput achieved using the benchmark’s

optimization is shown in Table 2. The latter table shows also how

much of the throughput is due to each of the individual paths be-

tween a sender and a receiver. These individual path use option-

ally one or more of the overlay nodes and contribute to the overall

throughput achieved between this pair, while cooperating with the

other sender-receiver pairs in the consumption of the underlay links’

resources to achieve the optimal overall throughput.

We see cleary that the benchmark achieves a much higher per-

pair throughput than the shortest path approach, but this is to be ex-

pected for non-trivial topologies. However more importantly, when

the benchmark assigns the bandwidth shares of the individual over-

lay paths that are available to each sender-receiver pair, it has to do

this in a non-trivial way to achieve the optimum throughput. We

can even see that considerable resources are gained from using non-

trivial paths. In case of the chessboard configuration, we can even

see that a non-trivial path (v2

share to the overall throughput for one pair.

We are therefore confident that meeting the limits of achievable

bandwidth is a challenge for ad-hoc multipath streaming decision

mechanisms. In creating our benchmark, we have gained a means of

evaluating the performance of dynamic mechanisms for their poten-

tial for further improvement.

s→ v2

o→ v1

r) contributes the largest

5. CONCLUSIONS AND FUTURE WORK

We have presented a benchmarking system for multipath media

streaming of scalable media which we have applied to different

Page 4

Overlay

path

from

v1

s

v1

s

v1

s

v1

s

v1

s

sum

v2

s

v2

s

v2

s

v2

s

v2

s

sum

sending ratesending rate

distribution for Waxman

to v1

r

32.0

10.1

9.8

1.0

9.1

distribution for Chessboard

to v1

r

19.0

7.5

2.2

2.0

1.0

throughto v2

r

to v2

r

31.0

10.8

10.1

7.3

5.7

11.0

1.6

10.0

1.8

2.3

v1

v2

v1

v2

o

o

o, v2

o, v1

o

o

62.064.931.726.7

87.0

11.0

0.5

6.0

5.6

39.0

8.0

12.0

3.0

6.0

12.0

6.0

26.0

0.7

1.1

34.0

1.0

4.0

1.1

0.7

v1

v2

v1

v2

o

o

o, v2

o, v1

o

o

110.168.0 45.840.8

Table 2. Distribution of sending rates along overlay paths.

network topologies. For a given topology and set of overlay nodes it

determines, for a given streaming demand, the optimal allocation of

link bandwidth with the goal of allocating the maximum bandwidth

to each of the streams. As bandwidth demands for streams may

vary widely, alternative optimization goals could be achieved by

modifying equation 1, in order to maximize the bandwidth assigned

to a stream relative to its minimal bandwidth.

Sincethecalculationisrathercomputing-intensive, andsincethe

knowledge of the entire system state is necessary for the calculation

this method is not suited for an implementation on a network node

due to scalability reasons. However, our benchmarking system will

be used to compare other algorithms with the optimal case provided

by this work.

Weplantoimplementmultipathstreamingsolutionsfromthelit-

erature and use the results of the benchmark to evaluate how closely

these overlay solutions can track the optimal resource utilization de-

cision in situations where the network conditions are changing dy-

namically.

As a further step, we intend to develop distributed algorithms

that operate under partial knowledge of the network topology and

dynamically changing networking parameters such as available

bandwidth and packet loss. These algorithms are then to be inte-

grated into the overlay node together with other basic streaming

functionalities. These can include stream caching, transcoding of

multimedia content, error correction mechanisms, etc., and will

require considerable extension of the benchmark as well.

Another important issue is the interaction between the overlay

network and wireless access networks especially in connection with

handovers between two access networks [16]. It is important to con-

sider how signaling between the mobile device and the overlay net is

handled. Signaling should be done so that both the terminal mobility

and the session mobility are provided for the users.

6. ACKNOWLEDGMENTS

The work presented in this paper has been conducted as a part of the

Adimus (Adaptive Internet Multimedia Streaming) project funded

by the Nordunet-3 program.

7. REFERENCES

[1] Dan Jurca and Pascal Frossard, “Distributed media rate alloca-

tion in overlay networks,” in 2006 IEEE International Confer-

ence on Multimedia and Expo, July, 2006, vol. 9-12, pp. 1401

– 1404.

[2] Dan Jurca and Pascal Frossard, “Video packet selection and

scheduling for multipath streaming,” IEEE Transactions on

Multimedia, vol. 9, no. 3, pp. 629–641, April 2007.

[3] L.Golubchik, J.Lui, T.Tung, A.Chow, W. Lee, G.Franceschi-

nis, and C. Anglano, “Multi-path continuous media streaming:

What are the benefits,” 2002.

[4] John G. Apostolopoulos and Mitchell D. Trott, “Path diversity

for enhanced media streaming,” Communications Magazine,

IEEE, vol. 42, pp. 80– 87, 2004.

[5] H. Chu and K. Nahrstedt, “Dynamic multi-path communica-

tion for video traffic,” in Proceedings of the Thirtieth Hawaii

International Conference on System Sciences, 1997, vol. 1, pp.

695–704.

[6] AttilaTak´ acs, Andr´ asCs´ asz´ ar, R´ obertSzab´ o, andTam´ asHenk,

“Generic multipath routing concept for dynamic traffic engi-

neering,” in Communications Letters, IEEE, 2006, vol. 10, pp.

126 – 128.

[7] C. Villamizar, “Ospf optimized multipath (ospf-omp),” in Pro-

ceedings of the forty-fourth Internet Engineering Task Force,

1999, pp. 45–54.

[8] P. Narvaez, K.-Y. Siu, and H.-Y Tzeng, “Efficient algorithms

for multi-path link-state routing,” .

[9] R. Guerin, A. Orda, and D. Williams, “Qos routing mecha-

nisms and ospf extensions,” 1996.

[10] Mohamed Hefeeda, Ahsan Habib, Boyan Botev, Dongyan Xu,

and Bharat Bhargava, “Promise: peer-to-peer media streaming

using collectcast,” in MULTIMEDIA ’03: Proceedings of the

eleventh ACM international conference on Multimedia, 2003,

pp. 45–54.

[11] M. Castro, P. Druschel, A. Kermarrec, A. Nandi, A. Rowstron,

and A. Singh, “Splitstream: High-bandwidth multicast in co-

operative environments,” 2003.

[12] Tian Bu, Nick Duffield, Francesco Lo Presti, and Don Towsley,

“Network tomography on general topologies,” in SIGMET-

RICS ’02: Proceedings of the 2002 ACM SIGMETRICS in-

ternational conference on Measurement and modeling of com-

puter systems, 2002, pp. 21–30.

[13] D. Rubenstein, J.F. Kurose, and D.F. Towsley,

ing shared congestion of flows via end-to-end measurement,”

IEEE/ACM Transactions on Networking, vol. 10, no. 3, pp.

381–395, June 2002.

“Detect-

[14] William H. Press, Saul A. Teukolsky, William T. Vetterling,

and Brian P. Flannery, Numerical Recipes 3rd Edition The Art

of Scientific Computing, Cambridge University Press, 2007,

ISBN 13:9780521880688.

[15] A. Medina, A.Lakhina, I. Matta, and J. Byers, “Brite: An

approach to universal topology generation,”

’01: Proceedings of the International Workshop on Model-

ing,Analysis and Simulation of Computer and Telecomunica-

tions Systems, 2001.

in MASCOTS

[16] Tiia Sutinen and Tapio Frantti, “Reference service architec-

tureformultimediaservicesinheterogeneousmulti-accessnet-

works,” in Mobilware’08: Conference on MOBILe Wireless

MiddleWARE, Operating Systems, and Applications, 2008.