210 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, …iandreop/Andreopoulos_JSAC_Video... ·...

2104 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 24, NO. 11, NOVEMBER 2006

Cross-Layer Optimized Video Streaming OverWireless Multihop Mesh Networks

Yiannis Andreopoulos, Member, IEEE, Nicholas Mastronarde, and Mihaela van der Schaar, Senior Member, IEEE

Abstract—The proliferation of wireless multihop communica-tion infrastructures in office or residential environments dependson their ability to support a variety of emerging applicationsrequiring real-time video transmission between stations locatedacross the network. We propose an integrated cross-layer op-timization algorithm aimed at maximizing the decoded videoquality of delay-constrained streaming in a multihop wirelessmesh network that supports quality-of-service. The key principleof our algorithm lays in the synergistic optimization of differentcontrol parameters at each node of the multihop network, acrossthe protocol layers—application, network, medium access control,and physical layers, as well as end-to-end, across the variousnodes. To drive this optimization, we assume an overlay networkinfrastructure, which is able to convey information on the condi-tions of each link. Various scenarios that perform the integratedoptimization using different levels (“horizons”) of informationabout the network status are examined. The differences betweenseveral optimization scenarios in terms of decoded video qualityand required streaming complexity are quantified. Our resultsdemonstrate the merits and the need for cross-layer optimizationin order to provide an efficient solution for real-time video trans-mission using existing protocols and infrastructures. In addition,they provide important insights for future protocol and system de-sign targeted at enhanced video streaming support across wirelessmesh networks.

Index Terms—Cross-layer strategies, distributed video stream-ing optimization, quality-of-service (QoS), wireless mesh networks.

I. INTRODUCTION

WIRELESS mesh networks are built based on a mixtureof fixed and mobile nodes interconnected via wireless

links to form a multihop ad hoc network. The use of existingprotocols for the interconnection of the various nodes (hops)is typically desired as it reduces deployment costs and also in-creases interoperability [1]. However, due to the network andchannel dynamics, there are significant challenges in the designand joint optimization of application, routing, medium accesscontrol (MAC), and physical (PHY) adaptation strategies for ef-ficient video transmission across such mesh networks.

In this paper, we are addressing some of these challengesby developing an integrated video streaming paradigm enablingcross-layer interaction across the protocol stack and across the

Manuscript received October 1, 2005; revised March 4, 2006 and May 1,2006. This work was supported in part by the National Science Foundation underCareer CCF-0541867 and in part by a grant from Intel IT Research.

Y. Andreopoulos and M. van der Schaar are with the Department of ElectricalEngineering, University of California at Los Angeles (UCLA), Los Angeles, CA90095-1594 USA (e-mail: [email protected]; [email protected]).

N. Mastronarde is with the Department of Electrical and Computer Engi-neering, University of California at Davis (UCDavis), Davis, CA 95616 USA(e-mail: [email protected]).

Digital Object Identifier 10.1109/JSAC.2006.881614

multiple hops. The problem of multihop video streaming has re-cently been studied under a variety of scenarios [2]–[4]. How-ever, the majority of this research does not consider the protec-tion techniques available at the lower layers of the protocol stackand/or optimizes the video transport using purely end-to-endmetrics, thereby excluding a significant amount of improvementthat can occur by cross-layer design [5]–[7]. Consequently, theinherent network dynamics occurring in a multihop wirelessmesh network as well as the interaction among the various layersof the protocol stack are not fully considered in the existingvideo streaming literature. Indeed, recent results concerning thepractical throughput and packet loss analysis of multihop wire-less networks [8], [9] have shown that the incorporation of ap-propriate utility functions that take into account specific param-eters of the protocol layers such as the expected retransmissions,the loss rate, and bandwidth of each link [8], as well as ex-pected transmission time [9] or fairness issues [10], can signif-icantly impact the actual end-to-end network throughput. Mo-tivated by this work, we show that, for delay-constrained videostreaming over multihop wireless mesh networks, including thelower layer network information and adaptation parameters inthe cross-layer design can provide significant improvements inthe decoded video quality.

In this paper, we focus on the problem of real-time trans-mission of an individual video bitstream across a multihop802.11a/e wireless network and investigate: 1) what is thevideo quality improvement that can be obtained if an integratedcross-layer strategy involving the various layers of the protocolstack is performed and 2) what is the performance and com-plexity impact if the optimized streaming solution is performedusing only limited, localized information about the networkstatus, as opposed to global, complete information.

We assume that the mesh network topology is fixed over theduration of the video session and that, prior to the transmission,each application (video flow) reserves a predetermined trans-mission opportunity interval, where contention-free access tothe medium is provided.1 This reservation can be performed fol-lowing the principles of the HCCA2 protocol of IEEE 802.11e[12] and can be determined based on the amount of flowssharing the network. Although the design of such a reservationsystem is an important problem and it affects our results, recentwork showed that scheduling of multiple flows in the contextof a mesh topology can be done such that the average ratefor every flow is satisfied and the interference to neighboring

1Existing IEEE standards [12] already support such quality-of-service (QoS)mechanisms, which, barring interference and environment noise, provide guar-anteed transmission time for each admitted application (video flow).

2HCCA: HCF controlled channel access, where HCF stands for hybrid coor-dinator function [12].

0733-8716/$20.00 © 2006 IEEE

ANDREOPOULOS et al.: CROSS-LAYER OPTIMIZED VIDEO STREAMING OVER WIRELESS MULTIHOP MESH NETWORKS 2105

Fig. 1. A simple topology with three hops.

nodes is minimized [13]. Hence, a similar solution can beapplied for our case and the available nodes and links withinthe entire mesh topology can be preestablished by a centralcoordinator prior to the video streaming session initiation. Thisminimizes the probability of additional delays and link failuresdue to routing reconfigurations during the video streaming andalso decouples the problem of optimized media streaming andoptimized route and link-reservation establishment within thewireless multihop network. Once the available network infra-structure to the video streaming session has been established,we assume that an overlay network topology can convey (infrequent intervals) information about the expected bit-error rate(BER), the queueing delay for each link, as well as the guaran-teed bandwidth under the dynamically changing modulation atthe PHY. Several examples of such application-layer overlaynetworks have been proposed in the literature [19], [20].

Under the above assumptions, this paper makes the followingcontributions. For video packets of each hop in the mesh net-work, we propose an optimization framework that jointly deter-mines per packet: 1) the optimal modulation at the PHY; 2) theoptimal retry limit at the MAC; 3) the optimal path (route) to thereceiver in the remaining part of the mesh network; and 4) theapplication-layer optimized packet scheduling, given a predeter-mined topology and time reservation per link using the conceptsof IEEE 802.11e HCCA.

This paper is organized as follows. Section II defines the sce-narios examined in this work and provides the necessary defi-nitions and formulations for the expected bandwidth, transmis-sion error rate as, well as the expected delay for streaming undervarious network paths. Section III presents the cross-layer op-timization problem. The proposed solutions are presented inSection IV. Section V analyzes the complexity and feedbackrequirements of the proposed approaches. Section VI presentsindicative results, including comparisons with other well-knownapproaches from the literature. Our conclusions are presented inSection VII.

II. PROPOSED INTEGRATED CROSS-LAYER VIDEO STREAMING

Consider that nodes (hops) of a wireless multihop meshnetwork decide to participate in a video streaming session. Ex-ample topologies with and are shown in Figs. 1and 2. Node represents the original video source, while node

is the destination node (video client). Each link is associ-ated with the corresponding allocated bandwidth for the videotraffic , the error rate observed on the link , as wellas the corresponding delay due to the video queue .Within the reserved time for the video traffic, each link exhibits acertain throughput given the chosen modulation strategy. Video

packets are lost due to the experienced BER. This error is due tonoise and interference in the wireless medium stemming frombackground noise, node mobility, or simultaneous link transmis-sions. In addition to this error, under delay-constrained videostreaming, packets are discarded due to delays incurred in thetransmission, e.g., the queueing delay of each link. Notice thatFig. 2 displays different connectivity structures for the networktopology, as specified by the indicated links. Obviously, thetightly connected multihop mesh topologies T1 and T2 of Fig. 2offer more alternative paths for the video traffic that topologyT3; however, the overall reserved time across the various nodesof the network is also increased. In general, the decision on theconnectivity as well as the number of nodes participating in thevideo streaming session depends largely on a number of system-related factors that transcend the video streaming problem (e.g.,node cooperation strategy/incentives and network coordinationand routing policies imposed by the utilized protocols). Hence,in this paper, we investigate cross-layer optimization for videoover multihop wireless mesh networks given the network spec-ification (participating nodes and connectivity), as well as theavailable reservation time on each link for the video traffic.

Under the existence of feedback from an overlay network in-frastructure, the BER and queueing delay per link can be dis-seminated to the remaining network hops at frequent intervals(via a hop-to-hop feedback mechanism3), or when the incurredchange in network parameters is larger than a preset threshold.Thus, they can be considered to be known (Fig. 1). However,in certain cases, feedback from remote hops may arrive with anintolerable delay, or, alternatively, it can be deemed unreliabledue to the rapidly changing network conditions. As a result, acertain “horizon” of information retrieval can be envisaged foreach hop (Fig. 2), where network information within the horizonis deemed reliable and can be received in a timely manner, whileinformation beyond the horizon can only be theoretically esti-mated based on average or previous measurements.

A. Wireless Multihop Mesh Topology Specification

For a generic multihop wireless mesh network, we considerthe connectivity structure

(1)

where each element is the connectivity vector(end-to-end network path) given by

(2)

where each component indicates a partic-ular wireless link (the th link of path ), and is thetotal number of links participating in the network path . Forexample, for the topology of Fig. 1 with , we have

(3)

3For example, in order to utilize the medium more efficiently, it is possible topiggyback feedback about the link status information onto the acknowledgmentpackets.


Fig. 2. More complex topologies with seven hops. In these cases, the link information up to hop h is directly conveyed by the overlay network infrastructure,while other link information is inferred based on theoretical estimates using average or past information (where m indicates the estimated or average value forthe metric m ;m = fg; e; dg).

with and

(4)

Notice that (1) and (2) apply both for the end-to-end topology ofinterest but also for the topology between any intermediate nodeand the terminal (client) node in the mesh network utilized forvideo transmission. For example, if we consider the subnetworkof topology T2 of Fig. 2 consisting of nodes , and ,there are two paths from to , and the equivalent definitionsapply locally. Hence, the subsequent problem specification andanalysis is inherently scalable and can be applied in a similar

fashion to either the entire end-to-end topology or only part ofthe topology (subnetwork). Finally, it is important to mentionthat all the proposed algorithms in this paper assume the nonex-istence of routing loops, i.e., the mesh network between the cur-rent hop and the destination hop can be represented by a treegraph.

B. Link and Path Parameter Specification

For each link , given a certain modulation at thephysical layer, we denote the expected BER as . Noticethat this error is usually estimated based on channel modeling,as well as experimental studies in the network which analyzethe effects of interference [11]. As a result, the higher layersof the protocol stack can assume to be independent and


randomly distributed [21]. Under a predetermined negotiationof traffic specification parameters for each link in the meshnetwork (e.g., following the HCCA protocol [12]), each linkcan provide a guaranteed bandwidth at the applicationlayer. Following the HCCA specification [12], this bandwidthis linked with the traffic specification parameters by [14]:

(5)

where is the transmission opportunity duration pro-vided by the HCCA admission control for the video flow trafficof link is the nominal MAC service data unit (MSDU)size,4 is the specified duration of the service interval[12] for the video flow traffic at link is the phys-ical-layer rate, and represents the duration of the re-quired overheads corresponding to polling and acknowledgmentpolicies. As demonstrated by (5), even though the negotiatedtransmission opportunity duration is constant per link, the guar-anteed bandwidth depends on the provided physical-layer rate

, which in turn makes it dependent on the chosenmodulation5 . Finally, depending on the chosen mod-ulation, may change for each MSDU. Hence, theguaranteed bandwidth of (5) can be determined for eachMSDU.

Under the aforementioned assumptions for the error model ofeach link, the probability of error for the transmission of MSDU

of size bits is

(6)

Consequently, the probability of error for the packet transmis-sion in path is

(7)

Under a single (successful) MSDU transmission via each link, the transmission delay for path can be calculated as

(8)

where depends on the transmitting-link queuelength and will be discussed in the next subsection.

Considering an end-to-end scenario, if we denote the retrans-mission limit for each MSDU (transmitted via path ) as

, the average number of transmissions over path until

4In this paper, we assume that one video packet is encapsulated in one MSDUand the two terms are used interchangeably.

5For notational simplicity, we do not particularly indicate the dependence ofe(l ) and g(l ) on the modulation m(l ).

the packet is successfully transmitted, or the retransmissionlimit is reached, can be calculated as

(9)

Consequently, the (end-to-end) expected delay for the transmis-sion of an MSDU of size through undertransmissions can be approximated by

(10)

The last equation derives the end-to-end delay estimate byjoining all links of path via the summation terms, therebyforming a “virtual” link from the sender to the receiver nodein the multihop network. We follow this approach since themaximum number of retransmissions required on eachpath can only be defined end-to-end, based on the maximumpermissible delay from the sender node to the receiver. Weremark that in our experiments, the retransmission limit forany part of a path or even for one link is set equal to ,since, in principle, all possible retransmissions (until the MSDUexpires due to delay violation) could occur at an individuallink. Following the analysis of (7)–(10), it is straightforward todefine the average MSDU transmissions and the expected delayfor subpaths that include only a subset of links, or even for anindividual link. This will be proven to be very useful for someof the derivations of this paper.

C. Application and Network-Layer Parameter Specification

Since we are considering real-time video streaming throughthe multihop wireless network, each MSDU is associated witha corresponding delay limit , before which the videodata encapsulated in the MSDU should arrive at the destinationnode . In addition, decoding the video data at the videoreceiver incurs a reduction in the perceived distortion, which isrepresented by . Several models exist for the definition of

(e.g., see [15]). Recent results [16], [17] demonstrated thatacknowledgment-based transmission of scalable video undera strict distortion-reduction prioritization of the video packetsleads to an additive distortion-reduction model at the receiverside under packet losses in the multihop network. This additivemodel is codec-specific and typically expresses the expectedmean square error (MSE) reduction at the video decoder insteadof the visual distortion reduction, since the latter is harder toquantify. See [15] for an example and [18] for further detailson linking the distortion-reduction estimates with thepacketization process at the application layer. In this paper,we assume that an optimized packet scheduling is performedat the application layer, where all packets with the same


delay deadline are ordered at the encoder (sender)side according to their expected distortion reduction [15],[3]. The delay deadline will also be considered as aparameter in the proposed optimization strategy and it will bedefined based on the application requirements.

At each link in the mesh wireless network, each video flowis subjected to a queueing delay,6 which depends on: 1) theMSDUs from a particular flow (user) that are scheduled fortransmission via the link of interest at the moment when MSDU

arrives and 2) the queue output rate. The queue output rate de-pends on the quality of the link [error probability given in (6)]and the average number of retransmissions for each MSDU inthe particular link [given by (9) with the replacement of by

]. If we assume that link is shared among multiple paths,then at the arrival of MSDU at the queue of link , another

MSDUs (where, typically, ) will be in the samequeue. For each , by indicating the group of MSDUs byvector , the queueing delay can be estimated as

(11)

For the optimization of the routing strategy of each MSDU(presented in the next section), the determination of (11) canbe performed dynamically during the path estimation, under theknowledge of the previous decisions for the MSDUs that weretransmitted by the current node. Alternatively, each node canindependently calculate (11) based on the queue contents of theparticular link and disseminate the result at frequent intervals inthe mesh network via the overlay infrastructure.

III. PROBLEM FORMULATION

Assume a set of wireless hops (nodes), with being thevideo encoder (server) and the video decoder (client), and aconnectivity structure with paths, where each path

consists of hops. In addition, assume a predefinedHCCA transmission opportunity duration for eachlink , with , and a link adaptation mechanismat the physical layer that can operate at an MSDU granularity.The end-to-end cross-layer optimization which determines thechosen path (routing), the maximum MAC retry limit, and thechosen modulation (at the PHY layer) for the transmission ofeach MSDU is

(12)

where

(13)

6In this paper, we assume that the MSDUs of each flow are accommodatedwith an independent queue at each link.

with given by (7), the maximum number of re-transmissions for MSDU if scheduled via path , andcorresponding to the remaining time interval for which linkcan support the video-flow traffic under HCCA. For the trans-mission opportunity intervals belonging to the current serviceinterval, can be calculated as

(14)

Under the constraints set by the video codec and the mesh wire-less network infrastructure, the optimization of (12) attemptsto find the cross-layer parameters that maximize a capacity-distortion utility function. This function is formulated as theproduct of the minimum path capacity (expressed by the re-maining reserved time within the current service interval at themost congested link) and the expected source distortion-reduc-tion of (13). In this way, we minimize congestion across the var-ious links [since the path whose worst link is having the highestcapacity is selected under given by (13)], and con-currently maximize the expected distortion reduction (under thecurrent path’s minimum link capacity ).The granularity of this optimization is one MSDU. However,coarser granularities could also be considered, in order to re-duce complexity. The problem constraints can be expressed foreach MSDU as

(15)

i.e., the maximum transmission delay through each possiblepath must be below or equal to the MSDU deadlinein order for the video data to be useful to the decoder. Moreover,the timing constraint set from HCCA scheduling is

(16)

The two constraints of (15) and (16) can provide two boundsfor the maximum number of retries for each MSDU for eachlink . Since both the MAC-layer scheduling and the appli-cation-layer deadline constraints are concurrently imposed, if

, we set the tightestbound for the maximum retry limit

(17)

Obviously, if there is no path for which

, then the current MSDU may be dropped.


Fig. 3. Exhaustive algorithm for the determination of the cross-layer optimized mesh-network path selection, MAC retry limit, and physical-layer modulation.The algorithm is applied for each MSDU existing in the queue of each node in the multihop wireless network.

IV. VIDEO STREAMING OPTIMIZATION IN THE

MULTIHOP MESH NETWORK

In this section, we derive an algorithm that determines theoptimal parameters for (12) under a predetermined deadline foreach MSDU (given by ) and a predetermined trans-mission opportunity duration per link, given by ,which is set by the HCCA admission control once the video flowis scheduled for transmission. Moreover, although the condi-tions of the various links vary over time, we assume the networktopology to be fixed for the duration of the video transmission.

A. End-to-End Optimization

The optimization of (12) can be performed for each node ofthe mesh wireless network under the assumption that, for everylink , the parameters are determined based onthe chosen modulation , and the experienced signal-to-interference-plus-noise ratio (SINR) [21]. In addition, we as-sume that is communicated to the sender node viafrequent feedback using an overlay network infrastructure [19],[20] that uses real-time protocols for conveying informationfrom different layers.

The proposed optimization algorithm is given in Fig. 3. No-tice that, although an entire path is selected at the sender node,the algorithm is executed for each node in the network inde-pendently by assuming each node is the sender and consid-ering only the network (and MSDU) subset corresponding tothe node of interest. This ensures that the algorithm can scalewell under a variety of topologies. In addition, in this way, po-tential network variations that invalidate the error, bandwidth, orqueueing-delay assumptions used when scheduling at the sendernode, can be incorporated/corrected during the scheduling ofa subsequent node. Finally, the independent algorithm execu-tion at each node ensures that expired MSDUs will not propa-gate through the entire network unnecessarily. This facilitatesthe conservation of network resources in the mesh topology andreduces link congestion.

The algorithm of Fig. 3 searches through all the possiblerouting configurations (line 4) that emerge under varying modu-lation strategies (line 6) and determines the retransmission limitfor each case (line 8). The utility function is evaluated (line 9)

and the overall maximum is retained. Although this is a greedyapproach, it is guaranteed to obtain the maximum under dy-namic feedback from the overlay network (parameters calcu-lated in line 7).

B. Optimization Under a Certain Horizon of NetworkInformation

In this case, we are only considering the part of the mesh net-work topology that immediately connects to the node of interest.This may be advantageous in comparison to the previous case,since a limited set of network parameters needs to be commu-nicated to the sender node.

For analytical purposes, this can be considered as the previouscase with , where is the total lengthof the path that was used in the end-to-end optimization of theprevious section. In this case, every path originating from thecurrent node consists of one or more links, but we do not con-sider the entire path to the destination. The advantage offeredby this scenario is that the required information for the MSDUscheduling is localized (limited).

For each path , we assume that the information for the opti-mization process is known only for links . Forthe remaining links of each path , we as-sume that the allocated transmission opportunity duration avail-able for the MSDUs of each link is known, as well as the limitsfor the SINR experienced by each link. Our goal is to establish

for the video transmission up to links , i.e., theknown network “horizon,” in order to perform the optimizationof Fig. 3 locally. With respect to IEEE 802.11a networks, it canbe shown [21], [22] that the physical-layer throughput of eachlink can be approximated by

(18)

where is the maximum achievable data rate for eachmodulation is the observed SINR, and areconstants whose values for each modulation can be ex-tracted based on the observation for and predetermined exper-imental points [21]. Assuming that, for every link , the SINR


is a random variable following a certain probability distribu-tion , we have

(19)

where are the bounds of the observable values forthe SINR for each link . In addition, under a given or esti-mated probability distribution , (5) can be used in orderto derive the expectation for the guaranteed bandwidth of eachlink, which, after a few straightforward manipulations is shownin (20) at the bottom of the page, since the remaining parame-ters of (5) are constants (in our analysis, we consider a nominalMSDU size ). In a similar manner, the expected error of thesubpath within each path is

(21)

The last equation was derived based on (7) under the assumptionthat the bit-error probability can be approximated by [21], [22]

(22)

where are derived experimentally depending on the ob-servation for and the chosen modulation [21],with .

Having the expected values for the full path’s guaranteedbandwidth, the maximum transmission delay for an MSDUtransmitted through links can be derivedbased on (10) as

(23)

Notice that (23) involves also the knowledge of the queueingdelay of each link , i.e.,for the subsequent links after the “horizon.” For each link ,

the expected can be updated within intervals of(that correspond to

the expected time required for MSDU to reach link , afterit passes the link which is at the “horizon”) as

(24)

where operator indicates the update of quantityby and

(25)

(26)

The derivation of (24) is performed as follows. The queueingdelay of the previous iteration is incremented by the productof the factor which indicates the expected delay due to retriesfor the new MSDU in link with the probability that theMSDU will reach link successfully (after maximallyretries are performed at links ), which is givenby , defined in (25). At the same time, the queueing delayis decremented by the product of the factor indicating the timeduration for the possible successful MSDU transmissions withthe probability of a successful MSDU transmission, which isgiven by , defined in (26).

Assuming that the value for is pro-vided based on (24), (23) can be used in the constraint of (15)by updating the delay deadline

(27)

and the optimization process follows, as explained in the begin-ning of this section.

The analytical formulation of this section is also useful indefining low-complexity scheduling algorithms at each nodewithout the need for real-time network feedback. For example,

(20)


Fig. 4. Algorithm for cross-layer optimization under an estimation-based framework.

if we assume that, due to the random interference caused by thesimultaneous operation of the wireless nodes in the mesh net-work, the probability distribution for each link is uni-form within an interval of , we have

(28)

With the explicit expression of from (28), we can derivethe expected values for the physical-layer rate, the guaranteedbandwidth and the path error rate by solving (19) and (20),and approximating the solution of (21) with numerical methods.Based on these values, we derive a less complex solution for thescheduling of each group of MSDUs corresponding to a videoGOP. The algorithm is given in Fig. 4. Based on this algorithm,for every new MSDU, all the cross-layer parameters are estab-lished analytically for each path (lines 1–3 of Fig. 4) and only thesearch through all the possible paths (i.e., line 4 of Fig. 4) is re-quired in order to derive the optimal solution. Consequently, thisoptimization has minimal complexity. In the following section,we formulate the complexity requirements of the three differentoptimization solutions, while Section VI presents comparativeexperimental results.

V. COMPLEXITY AND INFORMATION REQUIREMENTS OF THE

DIFFERENT ALTERNATIVES

Each proposed cross-layer optimization approach explores adifferent search space in order to determine the optimal parame-ters and also requires a varying amount of feedback on the con-ditions of the various links in the multihop mesh network. Thisresults in varying computational and communication require-ments for the presented algorithms.

Consider the case of a mesh network consisting of nodes.Each node , is the origin of paths. Eachpath stemming from node consists of nodes, with

. For each link of these paths, withand , there are possible mod-

ulations at the physical layer, which result in a different errorrate and different guaranteed bandwidth at the MAC layer. Forthe end-to-end cross-layer optimization with network feedbackfrom each node (Section IV-A), the overall complexity for thescheduling of an MSDU at node is

(29)

where represents the complexity for the dissem-ination of the necessary network information from node , aswell as the execution of the algorithm of Fig. 3.

Similarly, considering a scenario with partial network infor-mation, i.e., when the overlay network provides feedback onlyuntil node (with ), we have

(30)

where represents the complexity for the es-timation of the various parameters based on the analysis ofthe previous section and is the number of different paths(within the partial network topology under examination) orig-inating from node , with . Finally, for theoptimization of Fig. 4 where the best modulation strategy is apriori determined

(31)

where is the number of links that are directly connectedto node . As an indication of the different complexity re-quirements, as well as the different information requirements ofeach case, Table I presents numerical results for the three meshnetwork topologies of Fig. 2 based on (29)–(31), and we set

[21]. The normalized information requirements areexpressed in terms of the number of links in all possible paths(whose error, guaranteed bandwidth, and queueing delay is con-veyed by the overlay network) multiplied by the total numberof times this information is updated by the overlay network perMSDU ( with ). First, we considered thecase of the first node [ in (29)–(31)] since this includesall the possible paths and all the links in the mesh topology (topof the table). Hence, the results of the top part of Table I showthe worst case complexity and information requirements fromthe viewpoint of an individual node.

Notice that the information cost depends on the frequency ofupdates received by the overlay network per MSDU, denotedby . Given and the required bytes for conveyingthe status of each link via hop-to-hop feedback, it is straight-forward to convert the provided information cost for each caseinto actual bandwidth overhead for the overlay infrastructure inthe multihop wireless mesh network. Since for the


TABLE ICOMPLEXITY COMPARISON AND THE ASSOCIATED INFORMATION REQUIREMENTS OF DIFFERENT ALTERNATIVES FOR CROSS-LAYER

OPTIMIZATION FOR: (TOP) THE FIRST NODE OF EACH OF THE THREE TOPOLOGIES OF FIG. 2; (BOTTOM) ALL NODES

IN THE TOPOLOGY. THE BASIC COMPLEXITY UNIT FOR (29)–(31) IS SET TO C (1) = 1 AND WE

ADDITIONALLY SET C (1) = 0:2 � C (1) BASED ON EXPERIMENTAL OBSERVATIONS

estimation-based case, the information cost of this case is prac-tically negligible.

As a second step, we considered the cumulative complexityand information requirements for all the nodes in the multihopmesh network in order to estimate the streaming complexity andinformation overhead at the system level; the results are pre-sented in the bottom of Table I. We remark that, depending onthe topology specification (i.e., average node connectivity) andthe chosen method, the estimated complexity scales up to threeorders of magnitude. Similarly, there is a large gap between thelowest and highest information requirements for the various ap-proaches among the different topologies. As expected, the morecomplex the mesh topology, the higher the rate of increase ofcomplexity and information requirements.

VI. EXPERIMENTAL RESULTS

Although we experimented with a variety of video content,we present results with one typical video sequence (“Foreman,”300 frames, CIF format, with 30 Hz replay rate) since this ex-periment captures the average behavior of our system very well.We used a fully scalable codec [23] and the produced bitstreamwas extracted at an average bitrate of 2 Mb/s and packetizedinto MSDUs of data payload not larger than 1000 bytes. Theend-to-end delay for the MSDUs of each GOP was set to0.54 s, which corresponds to the replay duration of one GOP.We remark that although the utilized video coder is not amember of the MPEG family of coders, the assumptions madein Section II-C for the distortion-reduction estimation andthe application-layer packet scheduling are also valid for thescalable coder currently standardized by the JSVM group ofMPEG/VCEG [24] since it is based on open-loop motion-com-pensated prediction and update steps followed by embeddedquantization and context-based entropy coding. Hence, ourmethods and experiments are relevant to future systems thatwill utilize such scalable video coding technology in the contextof mesh networks.

We simulated the cases of the multihop mesh network topolo-gies of Fig. 2, labeled T1–T3, under predetermined transmis-sion intervals for each link. Our simulation took into account

the different parameters for the various layers, such as varyingSINR, transmission overheads at the MAC layer due to MSDUacknowledgments, and polling overheads, as well as queueingand propagation delays in the various links of the mesh network.In order to incorporate the effect of noise and interference, weperformed a number of simulations using random values for theSINR of each link, chosen between 15 and 25 dB. Network feed-back via the overlay network was conveyed to each hop when-ever a significant change in the experienced channel conditionoccurred. For the end-to-end optimization with network feed-back (termed “end-to-end” in our results) this includes the in-formation conveyed from all hops. However, we also consid-ered a localized case where the information horizon was set tothe direct neighborhood of each hop (termed “localized” in ourresults—this information horizon is shown pictorially in Fig. 2)and the remaining network parameters were estimated, as ex-plained in Section IV-B. In addition, a purely estimation-basedcase was also considered with no “horizon,” where the onlyavailable information is the channel SINR range [ of(28)] for each link, communicated by the overlay network infra-structure whenever the channel variation exceeded 2 dB (termed“estimation based”) from the estimated value given by (28). Thisensured that the information cost for the dissemination of thenetwork information is minimal compared with the other alter-natives, as indicated in Table I. Notice that, both for the “local-ized” case, as well as for the “estimation-based” case, the theo-retical framework of (18)–(28) was used.

Apart from the various alternatives of the proposed optimiza-tion, we also derive results with streaming under two other opti-mization algorithms. The first case is optimization based on theexpected transmission count (ETX) [8], where the utility func-tion is chosen such that the retransmission limit of each MSDUis set based on the effective network bandwidth and the expectederror rate. This case considers the MSDU delay deadline from apurely network-centric approach [8], i.e., it does not use the con-straints set in (15) and (16), but rather restricts the MSDU delaydeadline based on link loss ratios and the available throughput[8]. It was termed “ETX optimized” in our results. Second, thecase of selecting the link with the highest effective bandwidth


TABLE IIAVERAGE PSNR RESULTS (Y-CHANNEL�25 RUNS WITH 300 VIDEO FRAMES PER RUN) FOR VIDEO STREAMING IN THE MULTIHOP NETWORKS OF FIG. 2

TABLE IIIAVERAGE LINK UTILIZATION FACTOR AND PLR FOR EACH LINK OF THE THREE TOPOLOGIES OF FIG. 2 FOR THE CASE OF

MEDIUM- AND LOW-BANDWIDTH SIMULATION. THE NOMINAL MSDU SIZE �L = 1000 BYTES WAS ASSUMED

was realized for the routing of each MSDU, since it correspondsto the popular solution for optimized routing [25] (termed as the“highest bandwidth” solution). Notice that, in both cases, thebest modulation was established as in the “end-to-end” case, andeach link’s status information was also used for these cases, asconveyed by the overlay network infrastructure. As a result, thedifferences in performance stem purely from the different per-formance utilities that were chosen during the MSDU routingand path selection. Effectively, this separates the fully network-aware methods (proposed “end-to-end,” “highest bandwidth”[25], and “ETX optimized” [8]) from the partial network-awareapproaches (proposed “localized” and “estimation based”). Inaddition, within the fully network-aware methods, the differ-ence in the performance utilities means that only the “end-to-end” approach fully utilizes application-layer, MAC, and PHYparameters via the optimization framework of (12)–(16).

Indicative results for the obtained average peak-signal-to-noise ratio (PSNR) of each method are given in Table II (25 runsper method/case). Two representative cases of medium and lowaverage transmission bandwidth were chosen. For each case,the average percentage of bandwidth utilization and the packetloss rate (PLR) per link is shown in Table III. These results were

generated with the “end-to-end” case but similar results wereobtained for the remaining cases as well. The results of Table IIIdemonstrate that for each case (medium or low bandwidth), thepredetermined time reservation for the video flow packets perlink leads to similar average link utilization in all three topolo-gies, i.e., within a 4%–8% margin. Nevertheless, the obtainedaverage PLRs differ for each topology, as well as the obtainedaverage PSNR for each method, as indicated in Table II. In gen-eral, since the average link utilization is similar within each ofthe medium- and low-bandwidth utilization cases, the topologywith the most active links for the video traffic is expected toprovide the highest video quality. This is indeed true for thecorresponding results of each method as seen from the rows ofTable II. Nevertheless, the relationship between PSNR and PLRfor each topology (as reported in Table III) is not obvious.

In order to understand better the relationship between the ob-tained PLR for each case and the derived PSNR, the percentageof losses for the video packets when clustered into eight distinctdistortion categories is presented in Fig. 5. The second topologyof Fig. 2 was used for these results; similar results have beenobtained for the remaining topologies. Notice that our choice ofeight distinct categories is only performed for illustration pur-


Fig. 5. Percentage of losses for each packet distortion-reduction class (Cat:1 = least signi�cant packets; Cat:8 = most signi�cant). (a) Medium-band-width case. (b) Low-bandwidth case. The results correspond to topology T2 of Fig. 2.

poses, since each packet is associated with its own distortion-reduction. In our simulations, the packet losses were mainly dueto deadline violation, since each hop drops the packets whichhave already expired. The results of Fig. 5 indicate, for all thescenarios under consideration, that scheduling at the applicationlayer by expected distortion-reduction leads to reduced lossesfor the most significant classes of packets. This justifies our useof a scalable video coder that permits such a scheduling. How-ever, each method achieves different PSNR performance andPLRs depending on its chosen utility and the presence of net-work feedback.

As shown in the results of Table II, the “end-to-end” case out-performs all other methods by a significant margin. The “ETXoptimization” appears to perform relatively well, even though itis outperformed by approximately 1.3 dB by the “end-to-end”case. The “localized” case appears to outperform the popular“highest bandwidth” case in the vast majority of cases, eventhough the latter uses full feedback for the status of all the links ineach multihop topology. Although the “estimation-based” caseperforms worst, this case requires almost no network feedbackand, as shown in Table I, has the lowest complexity. Moreover, inthe case of low average bandwidth, this case is comparable to the“highest bandwidth.” This is expected since the “highest band-width” approach provides less intelligent decisions when most ofthe links have low effective throughput. Finally, a comparison ofthe results for the different topologies reveals that, as expected,the higher the connectivity, the better the average performanceof all methods. Nevertheless, this comes at a higher allocationof resources in the multihop mesh topology, and it additionallyhas higher complexity and requires more feedback for thecondition of all the links, as demonstrated in Table I.

Our results highlight several important issues in network de-sign and infrastructure. First, it was shown that having frequentfeedback via an overlay network about the link conditions andperforming end-to-end optimization with the appropriate utilityfunction offers significant improvements in the achievablevideo quality. Indeed, the “end-to-end” and “ETX optimized”cases outperform the remaining algorithms by 3–5 dB, inall cases (Table II). Second, the importance of choosing across-layer distortion-capacity utility function is highlighted bythe fact that both methods outperform the conventional “highestbandwidth” scenario. Moreover, the proposed utility of (12) andthe derivation of the MSDU retransmission limit based on thedelay limit for the video transmission [(15)–(17)] appear to be

the best choice for video streaming applications. Third, higherconnectivity in the multihop mesh topology leads to bettervideo streaming performance, at the expense of complexity andnetwork feedback requirements.

Fourth, the study of the PLRs reported in Fig. 5 in conjunc-tion with the PSNR results of Table II reveals that prioritizationof video packets with respect to distortion-reduction incurredin the decoded video is extremely important. For example, eventhough the “estimation-based” case has lower average PLR fromthe “highest bandwidth” case, it performs worse in terms ofPSNR since the latter achieves lower PLR for the most signifi-cant classes of packets. This result emphasizes the fact that, inthe case of analysis of multimedia transmission over wireless,average PLRs that do not consider the significance of the var-ious packets for the application are not always relevant metricsfor the system performance.

Finally, it appears that even a limited horizon of informationin the network infrastructure can be extremely beneficial. We be-lieve that the determination of an appropriate “horizon” of infor-mation that provides the optimal tradeoff between the overheadat the overlay network versus the improvement offered by uti-lizing dynamic network feedback is an interesting research di-rection. Moreover, the dynamic adaptation of such a “horizon”in function of the network variations or the mesh topology spec-ification (i.e., simple versus complex mesh networks and staticversus dynamic scenarios) could be examined.

VII. CONCLUSION

Delay-constrained video streaming over multihop wirelessmesh networks is an application that deserves considerable at-tention due to the research challenges imposed by such a service,as well as due to the important role that robust and efficient mul-timedia services have when it comes to commercial deploymentof such networks in office and residential areas. We investigateda framework where QoS guarantees are provided for video trans-mission over a variety of links in a multihop network using IEEE802.11a/e. The integrated cross-layer solution that maximizesthe product of the expected video quality with the link utilizationappears to provide significant improvement over other optimizedsolutions. Moreover, the utilization of network information (forthe dynamically changing conditions of the various hops) gath-ered via overlay-network feedback, appears to be of paramountimportance for the overall video quality at the receiver hop.


Although the proposed algorithm operates per video packetand can potentially incur significant complexity and communi-cation overhead for the overlay network infrastructure, there is asignificant potential for improved video streaming performance.This motivates us to investigate the problem further and attemptto explore the best granularity for the optimization, as well as thenetwork feedback that provides optimal quality/complexity/ro-bustness in a distributed video streaming scenario over the hopsof the mesh network. Finally, under the proposed paradigm, theissues of collaborative streaming of multiple flows and fairnessdeserve significant attention in future research.

REFERENCES

[1] K. Holt, “Wireless LAN: Past, present, and future,” Proc. IEEE Design,Autom., Test in Europe, vol. 3, pp. 92–93, 2005.

[2] W. Wei and A. Zakhor, “Multipath unicast and multicast video commu-nication over wireless ad hoc networks,” in Proc. Int. Conf. BroadbandNetw., 2002, pp. 496–505.

[3] P. A. Chou and Z. Miao, “Rate-distortion optimized streaming of pack-etized media,” IEEE Trans. Multimedia, vol. 8, no. 2, pp. 390–404, Apr.2006.

[4] E. Setton, T. Yoo, X. Zhu, A. Goldsmith, and B. Girod, “Cross-layerdesign of ad hoc networks for real-time video streaming,” IEEE Wire-less Commun. Mag., vol. 12, no. 4, pp. 59–65, Aug. 2005.

[5] M. van der Schaar, S. Krishnamachari, S. Choi, and X. Xu, “Adaptivecross-layer protection strategies for robust scalable video transmissionover 802.11 WLANs,” IEEE J. Sel. Areas Commun., vol. 21, no. 10,pp. 1752–1763, Dec. 2003.

[6] A. Butala and L. Tong, “Cross-layer design for medium access controlin CDMA ad-hoc networks,” EURASIP J. Appl. Signal Process., to bepublished.

[7] Y. Wu, P. A. Chou, Q. Zhang, K. Jain, W. Zhu, and S.-Y. Kung, “Net-work planning in wireless ad hoc networks: A cross-layer approach,”IEEE J. Sel. Areas Commun., vol. 23, no. 1, pp. 136–150, Jan. 2005.

[8] D. S. J. De Couto, D. Aguayo, J. Bicket, and R. Morris, “A highthroughput path metric for multi-hop wireless routing,” in Proc. ACMInt. Conf. Mobile Comput. Netw., 2003, pp. 134–146.

[9] R. Draves, J. Padhye, and B. Zill, “Routing in multi-radio, multi-hopwireless mesh networks,” in Proc. ACM Int. Conf. Mobile Comput.Netw., 2004, pp. 114–128.

[10] V. Gambiroza, B. Sadeghi, and E. W. Knightly, “End-to-end perfor-mance and fairness in multihop wireless backhaul networks,” in Proc.ACM Int. Conf. Mobile Comput. Netw., 2004, pp. 287–301.

[11] R. L. Cruz and A. V. Santhanam, “Optimal routing, link schedulingand power control in multihop wireless networks,” in Proc. Joint Conf.IEEE Comput. Commun. Society, 2003, vol. 1, pp. 702–711.

[12] Draft Supplement to Part 11: Wireless Medium Access Control(MAC) and physical layer (PHY) specifications: Medium AccessControl (MAC) Enhancements for Quality of Service (QoS), IEEE802.11e/D5.0, Jun. 2003.

[13] M. Kodialam and T. Nandagopal, “Characterizing achievable ratesin multi-hop wireless networks: The joint routing and schedulingproblem,” in Proc. ACM Int. Conf. Mobile Comput. Netw., 2003, pp.42–54.

[14] P. Ansel, Q. Ni, and T. Turletti, “An efficient scheduling scheme forIEEE 802.11e,” in Proc. IEEE Workshop on Model. and Opt. in Mobile,Ad-Hoc and Wireless Netw., Cambridge, U.K., Mar. 2004.

[15] M. Wang and M. van der Schaar, “Operational rate-distortion mod-eling for wavelet video coders,” IEEE Trans. Signal Process., to bepublished.

[16] D. Taubman and J. Thie, “Optimal erasure protection for scalably com-pressed video streams with limited retransmission,” IEEE Trans. ImageProcess., vol. 14, no. 8, pp. 1006–1019, Aug. 2005.

[17] M. van der Schaar and D. Turaga, “Cross-layer packetization and re-transmission strategies for delay-sensitive wireless multimedia trans-mission,” IEEE Trans. Multimedia, to be published.

[18] Y. Andreopoulos, R. Keralapura, M. van der Schaar, and C.-N. Chuah,“Failure-aware, open-loop, adaptive video streaming with packet-leveloptimized redundancy,” IEEE Trans. Multimedia, to be published.

[19] D. Krishnaswamy and J. Vicente, “Scalable adaptive wireless networksfor multimedia in the proactive enterprise,” Intel Technol. J., vol. 8, no.4, Nov. 2004. [Online]. Available: www: http://developer.intel.com/technology/itj/2004/volume08issue04/art04_scalingwireless/p01_ab-stract.htm

[20] S. Roy, M. Covell, J. Ankcorn, S. Wee, and T. Yoshimura, “A systemarchitecture for managing mobile streaming media services,” in Proc.IEEE Distrib. Comput. Syst. Workshop, 2003, pp. 408–413.

[21] D. Krishnaswamy, “Network-assisted link adaptation with power con-trol and channel reassignment in wireless networks,” in Proc. 3G Wire-less Conf., 2002, pp. 165–170.

[22] K.-B. Song and S. A. Mujtaba, “On the code-diversity performance ofbit-interleaved coded OFDM in frequency-selective fading channels,”in Proc. IEEE Veh. Technol. Conf., 2003, vol. 1, pp. 572–576.

[23] Y. Andreopoulos, A. Munteanu, J. Barbarien, M. van der Schaar, J.Cornelis, and P. Schelkens, “In-band motion compensated temporalfiltering,” Signal Process.: Image Commun., Special Issue on “Sub-band/Wavelet Interframe Video Coding”, vol. 19, no. 7, pp. 653–673,Aug. 2004, .

[24] J. R. Ohm, “Advances in scalable video coding,” in Proc. IEEE, Jan.2005, vol. 93, no. 1, pp. 42–56.

[25] D. B. Johnson and D. A. Maltz, “Dynamic source routing in ad hocwireless networks,” in Chapter in Mobile Computing, T. Imielinski andH. Korth, Eds. Norwell, MA: Kluwer, 1996.

Yiannis Andreopoulos (M’00) was born inAeghion, Greece, in 1977. He received the ElectricalEngineering Diploma and the M.Sc. degree in signaland image processing systems from the Universityof Patras, Patras, Greece, and the Ph.D. degree fromthe Vrije Universiteit Brussel, Brussels, Belgium.

Currently, he is working as a Postdoctoral Re-searcher at the University of California at LosAngeles (UCLA). During 2002–2003, he con-tributed to the ISO/IEC JTC1/SC29/WG11 (MPEG)committee (Scalable Video Coding Group). His

research interests are in the fields of transforms, complexity modeling formedia systems, video coding, and video transmission through unreliable media,e.g., wireless networks and the Internet.

Nicholas Mastronarde received the B.Sc. and M.Sc.degrees in electrical engineering from the Universityof California at Davis (UCD), in 2005 and 2006, re-spectively. He is currently working towards the Ph.D.degree at the University of California at Los Angeles(UCLA).

During his graduate study at UCD, he conductedresearch in video streaming over multihop wirelessmesh networks and published several internationalconference papers and a journal paper. His researchinterests include wireless multimedia transmission,

video coding, signal processing, and information theory.

Mihaela van der Schaar (SM’04) received thePh.D. degree from the Eindhoven University ofTechnology, Eindhoven, The Netherlands, in 2001.

She isnow anAssistant Professor in the Departmentof Electrical Engineering, University of California atLos Angeles (UCLA). Prior to this, she was a SeniorResearcher at Philips Research, The Netherlands,and the U.S., where she led a team of researchersworking on multimedia compression, networking,communications, and architectures. In 2003, she wasalso an Adjunct Assistant Professor at Columbia

University. From 2003 to 2005, she was an Assistant Professor in the Departmentof Electrical and Computer Engineering, University of California at Davis. Since1999, she was an active participant to the ISO Motion Picture Expert Group(MPEG) standard to which she made more than 50 contributions and for whichshe received three ISO recognition awards. She was also chairing for three yearsthe ad hoc group on MPEG-21 scalable video coding, and also co-chairing theMPEG Ad Hoc Group on multimedia testbed. She holds 22 granted U.S. patentsand several more pending. She has published extensively on multimediacompression, processing, communications, networking, and architectures.

Prof. van der Schaar received the National Science Foundation (NSF) CareerAward (2004) and the IBM Faculty Award (2005). She was an Associate Ed-itor of the IEEE TRANSACTIONS ON MULTIMEDIA and SPIE Electronic ImagingJournal from 2002 to 2005. Currently, she is an Associate Editor of the IEEETRANSACTIONS ON CIRCUITS AND SYSTEM FOR VIDEO TECHNOLOGY and anAssociate Editor of the IEEE SIGNAL PROCESSING LETTERS.

Date post:	14-Aug-2020
Category:	Documents
Upload:	others
View:	0 times
Download:	0 times

210 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, …iandreop/Andreopoulos_JSAC_Video... ·...

Documents