Informationally Decentralized Video Streaming Over Multihop

IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 6, OCTOBER 2007 1299

Informationally Decentralized Video StreamingOver Multihop Wireless Networks

Hsien-Po Shiang and Mihaela van der Schaar, Senior Member, IEEE

Abstract—Various packet scheduling, dynamic routing,error-protection, and channel adaptation strategies have been pro-posed at different layers of the protocol stack to address multi-uservideo streaming over multihop wireless networks. However, thesecross-layer transmission strategies can be efficiently optimizedonly if they use accurate information about the network conditionsand hence, are able to timely adapt to network changes. Due tothe informationally decentralized nature of the multihop wirelessnetwork, performing centralized optimization for delay-sensitivevideo streaming application based on global information about thenetwork status is not practical. Distributed solutions that adaptthe transmission strategies based on timely information feedbackneed to be considered. To acquire this information feedback forcross-layer adaptation, we deploy an overlay infrastructure, whichis able to relay the necessary information about the network statusand incurred delays across different network “horizons” (i.e.,across a different number of hops in a predetermined period oftime). In this paper, we propose a distributed streaming approachthat is optimized based on the local information feedback acquiredfrom the various network horizons. We investigate the distributedcross-layer adaptation at each wireless node by considering theadvantages resulting from an accurate and frequent networkinformation feedback from larger horizons as well as the draw-backs resulting from an increased transmission overhead. Basedon the information feedback, we can estimate the risk that packetsfrom different priority classes will not arrive at their destinationbefore their decoding deadline expires. Subsequently, the varioustransmission strategies such as packet scheduling, retransmissionlimit and dynamic routing policies are adapted to jointly considerthe estimated risk as well as the impact in terms of distortion ofthe different priority classes. Our results show that the proposeddynamic routing policy based on timely information feedbackoutperforms existing state-of-the-art on-demand routing solutionsby more than 2 dB in terms of the received video quality.

Index Terms—Decentralized information feedback, dynamicrouting, multihop wireless networks, video streaming.

I. INTRODUCTION

EMERGING multihop wireless LAN (WLAN) networksprovide a low-cost and flexible infrastructure that can be

simultaneously utilized by multiple users for a variety of ap-plications, including delay-sensitive multimedia transmission.However, these wireless networks provide only limited Quality

Manuscript received December 9, 2006; revised February 11, 2007. Thiswork was supported under NSF Career Grant CCF-0541867 and by UCMicro and HP Research. The associate editor coordinating the review of thismanuscript and approving it for publication was Dr. Anna Hac.

The authors are with the Department of Electrical Engineering, Universityof California, Los Angeles, CA 90095 USA (e-mail: [email protected];[email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TMM.2007.902845

of Service (QoS) support for real-time multimedia applications.Hence, efficient solutions for multimedia streaming must ac-commodate time-varying bandwidths and probabilities of errorintroduced by the shared nature of the wireless medium andquality of the physical connections. In the studied distributedtransmission scenario, multimedia users proactively collabo-rate in sharing the available wireless resources to maximizetheir video quality. To enable optimal usage of the multihopinfrastructure, the various network entities (source nodes, relaynodes, etc.) can timely and accurately exchange informationabout channel statistics, expected delays, or even packet lossprobabilities (due to the deadline expiration of video packets)incurred by previously transmitted multimedia packets fromdifferent users and distortion classes across the network. How-ever, this network information feedback usually shares thesame resources allocated for the payload (e.g., multimedia)transmission and thus, the resulting overheads need to beexplicitly considered for optimized transmission.

Prior research on multi-user multimedia transmission overwireless networks has focused on centralized, flow-basedresource allocation strategies based on pre-determined rate re-quirements and usually neglects the overheads associated withthe network information gathering and dissemination [1], [2],[6]. These solutions are not very adaptive to the network size orthe number of users and attempt to solve the global end-to-endrouting and path selection problem as a MulticommodityFlow [3] problem in a nonscalable fashion. This flow-basedoptimization does not take into account the fact that videoapplications are loss tolerant and hence, they can gracefullyadjust their quality to accommodate a larger number of usersas channel conditions are changing. Importantly, they do notguarantee that the packet delay constraints are met for videoapplications, since they do not timely adapt the transmissionstrategies for the various packets based on their delay deadlinesand the available network information, which captures thechanging network conditions such as congestion/interference.

Alternatively, the majority of the multimedia-centric researchoptimizes the video streaming using purely end-to-end metricsand does not consider the protection techniques available atthe lower layers of the protocol stack. Hence, they do nottake advantage of the significant gains provided by cross-layerdesign [2], [5], [6]. In [7], an integrated cross-layer optimiza-tion framework was proposed that considers the video qualityimpact based on different information horizons. However, theproposed solution in [7] considers only the single user case,where a set of paths and transmission opportunities are staticallypre-allocated for each video application. This leads to a subop-timal, nonscalable solution for the multi-user case. Importantly,

1520-9210/$25.00 © 2007 IEEE

1300 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 9, NO. 6, OCTOBER 2007

the overhead induced by the various information horizons arenot investigated in [7], which have essential impact for thedelay-sensitive multimedia applications. To enable efficientdistributed multi-user video streaming over a wireless multihopinfrastructure, nodes need to timely collect and disseminatenetwork information based on which, the various nodes cancollaboratively adapt their cross-layer transmission strategies.For instance, based on the available information feedback, anetwork node can timely choose an alternate (less congested)route for streaming the packets that have a higher contributionto the overall distortion or a more imminent deadline.

Although the information feedback is essential to thecross-layer optimization, the cost of collecting the informationis seldom discussed in the literature. Due to the informationallydecentralized nature of the multihop wireless network, it isimpractical to assume that the global network information andthe time-varying application requirements can be relayed tothe central (overlay) network manager in a timely manner.Distributed suboptimal solutions that adapt the transmissionstrategies based on well-designed localized information feed-back should be adopted for the delay-sensitive applications.

In summary, no integrated framework has been developedthat explicitly considers the impact of accurate and frequent net-work information feedback from various horizons, when opti-mizing the resource allocation and the cross-layer transmissionstrategies for multiple collaborating users streaming real-timemultimedia over a wireless multihop network. In this paper, webuild on the previous work [24] and investigate the impact ofthis information feedback on the distributed cross-layer trans-mission strategies deployed by the multiple video users. We as-sume a directed acyclic overlay network [8] that can be super-imposed over any wireless multihop network to convey the in-formation feedback. Our solution relies on the users’ agreementto collaborate by dynamically adapting the quality of their mul-timedia applications to accommodate the flows/packets of otherusers with a higher quality impact and/or higher probability tomiss their decoding deadlines. Unlike commercial multi-usersystems, where the incentive to collaborate is minimal, we in-vestigate the proposed approach in an enterprise network settingwhere source and relay nodes exchange accurate and trustableinformation about their applications and network statistics.

To increase the number of users that can simultaneously sharethe same wireless multihop infrastructure as well as to improvetheir performance given time-varying network conditions, wedeploy scalable video coding schemes [21] that enable a fine-granular adaptation to changing network conditions and a highergranularity in assigning the packet priorities. We assume eachreceiving node performs polling-based contention-free mediaaccess control (MAC) [16] that dynamically reserves a trans-mission opportunity interval in a service interval. The networktopology and the corresponding channel condition of each linkare assumed to remain unchanged within the service interval.

In this paper, we discuss the required information/parameterexchange among network nodes/layers for implementing a dis-tributed solution for selecting the following cross-layer trans-mission strategies at each intermediate node – the packet sched-uling, and the next-hop relay (node) selection based on routingpolicies similar to the Bellman-Ford routing algorithm [10], and

the retransmission limit at the MAC layer. In performing thecross-layer adaptation, we explicitly consider the packet dead-lines and the relative priorities (based on the quality impact ofthe packets) encapsulated in the packet headers. Each interme-diate node maintains a queue of video packets from varioususers and determines the cross-layer transmission strategies in adistributed fashion through the information feedback from otherintermediate nodes within a certain network horizon and witha certain frequency. While a larger horizon/frequency can pro-vide more accurate network information, this also results in anincreased transmission overhead that can have a negative im-pact on the video performance. Hence, we aim at quantifyingthe video quality benefit derived by the various users for dif-ferent network conditions and video application characteristicsbased on various information feedbacks.

Our paper makes the following contributions.

A. Decentralized Information Feedback Driven Cross-LayerAdaptation

In this paper, we show how the various cross-layer strategiescan be adapted based on the information feedback. The solu-tions of centralized flow-based optimizations [1]–[3] have sev-eral limitations. First, the video bitstreams are changing overtime in terms of required rates, priorities and delays. Hence, itis difficult to timely allocate the necessary bandwidths acrossthe wireless network infrastructure to match these time-varyingapplication requirements. Second, the delay constraints of thevarious packets are not explicitly considered in centralized solu-tions, as this information cannot be relayed to a central resourcemanager in a timely manner. Third, the complexity of the cen-tralized approach grows exponentially with the size of the net-work and number of video flows. Finally, the channel character-istics of the entire network (the capacity region of the network)need to be known for this centralized, oracle-based optimiza-tion. This is not practical as channel conditions are time-varying,and having accurate information about the status of all the net-work links is not realistic.

Alternatively, we focus on a fully distributed packet-based so-lution, where timely information feedback can efficiently drivethe cross-layer adaptation for each individual multimedia streamas well as the multi-user collaborations in sharing the wirelessinfrastructure. To cope with the delay sensitivity of the videotraffic, we explicitly consider the delay deadlines of the variouspackets (packets are dropped whenever their deadlines expire)and estimate the remaining transmission time based on the avail-able information feedback. This approach is better suited for theinformationally decentralized nature of the investigated multi-user video transmission problem over multihop infrastructures.

B. Impact of Various Information Horizons/Frequencies

We define the mechanism of information feedback conveyedthrough a multihop overlay infrastructure and investigate the im-pact of different information horizons/frequencies on the videoquality derived by the various multimedia users. We discussthe tradeoff between the increased transmission overhead andthe benefit of larger information horizons, which result in im-proved predictions of network conditions. More information al-lows nodes in the network to better estimate the time for each

SHIANG AND VAN DER SCHAAR: INFORMATIONALLY DECENTRALIZED VIDEO STREAMING 1301

packet to reach its destination and hence, the chance of missingits deadline.

C. Information Feedback Driven Packet Scheduling andRetransmission Strategies

We introduce the concept of risk estimation based on theavailable information feedback that determines the probabilitythat a packet will miss its delay deadline. Based on the estimatedrisk and the quality impact of the video packet, we proposednovel information feedback driven scheduling and retransmis-sion strategies for each node in the network.

D. Dynamic Routing Policies versus On-Demand SourceRouting

We present a novel dynamic routing policy that adapts basedon the information feedback to maximize the overall quality ofdelay-sensitive multimedia. Besides exploring different infor-mation horizons under various network conditions to enablea dynamic routing adaptation, we also compare our solutionagainst on-demand source routing that does the informationfeedback prior to the video transmission (e.g., the route dis-covery probing in DSR [25], and AODV [26]). Hence, unlikeour dynamic routing solution, where the information feedbackis performed during the entire video transmission, in theserouting schemes there is no transmission overhead associatedto the real-time information feedback.

The paper is organized as follows. Section II defines thevideo and network specification for multi-user video trans-mission over multihop wireless networks and provides across-layer distributed optimization scheme based on the infor-mation feedback. In Section III, we discuss the impact of theinformation feedback with different information horizons andpresent an integrated cross-layer adaptation algorithm for thereal-time multi-user streaming problem. Section IV introducesa novel information feedback driven scheduling algorithm thattakes advantage of the larger information horizons. Section Vintroduces our information feedback driven retransmissionlimit calculation, and Section VI presents our dynamic routingalgorithms based on the information feedback. In Section VII,we discuss the overheads of the information feedback of var-ious parameters. Simulation results are given in Section VIII.Section IX concludes the paper.

II. MULTI-USER VIDEO TRANSMISSION – PROBLEM

FORMULATION AND SYSTEM DESCRIPTION

A. Video Classes

We assume that video users with distinct source and des-tination nodes are sharing the same multihop wireless infra-structure. In [21], it has been shown that partitioning a scalablevideo flow (stream) into several prioritized classes can improvethe number of simultaneously admitted stations in a congested802.11a/e WLAN infrastructure, as well as the overall receivedquality. Hence, we separate each scalable encoded video streaminto a certain number of classes (quality layers). Similarly, inthis paper, we adopt an embedded 3-D wavelet codec [23] andconstruct video classes by truncating the embedded bitstream.We assume that the packets within each video class have the

same delay deadline, similar to [11], [21]. The number of classesfor a video sequence equals , and the total number of classesacross all users in the network equals . Eachvideo class is characterized by the following.

• , the expected quality impact of receiving the packets inthe class . We prioritize the video classes based on thisparameter. In the subsequent part of the paper, we label the

classes (across all users) in descending order of theirpriorities, i.e., .

• , the average packet lengths of the class . The ex-pected quality improvement for receiving a video packetin the class is defined as (see e.g., [11] for moredetails).

• , the number of packets in the class in one GOPduration of the corresponding video sequence.

• , the probabilities of successfully receiving thepackets in the class at the destination. Thus, the ex-pected number of the successfully received packets of theclass is .

• , the delay deadlines of the packets in the class .Due to the hierarchical temporal structure deployed in3-D wavelet video coders (see [11] and [21]), for a videosequence , the lower priority packets also have a lessstringent delay requirement. This is the reason why weprioritize the video bitstream in terms of the quality im-pact. However, if the used video coder did not exhibitthis property, we need to deploy alternative prioritizationtechniques that jointly consider the qualityimpact and delay constraints (see more sophisticatedmethods in e.g., [27]).

At the client side, the expected quality improvement for videoin one GOP can be expressed as

(1)

Here, we assume that the client implements a simple errorconcealment scheme, where the lower priority packets are dis-carded whenever the higher priority packets are lost [11]. This isbecause the quality improvement (gain) obtained from decodingthe lower priority packets is very limited (in such embeddedscalable video coders) whenever the higher priority packets arenot received. For example, drift errors can be observed whendecoding the lower priority packets without the higher prioritypackets [21]. Hence, we can write

if andotherwise,

(2)

where we use the notation in [27] – to indicate that theclass depends on . Specifically, if and are classesof the same video stream, means due to thedescending priority ( ). represents the end-to-endpacket loss probability for the packets of class . repre-sents the experienced end-to-end delay for the packets of class

. is an indicator function. Note that the end-to-end proba-bility depends on the network resource, competing users’


Fig. 1. Directed acyclic multihop overlay network for an exemplary wirelessinfrastructure. (a) Actual network topology that has two source-destinationpairs, five relay nodes. (b) Overlay network topology that has two source-des-tination pairs, six relay nodes (with one virtual node in the 1-hop intermediatenodes).

priorities as well as the deployed cross-layer transmission strate-gies. In addition, at the intermediate node , we assume that thevideo packets are scheduled in a specific order according tothe prioritization associated with the video content characteris-tics.

B. Overlay Network Specification

We assume that the source-destination pairs are connectedby a directed acyclic multihop overlay network, which is super-imposed upon the physical wireless network. This overlay net-work consists of hops with intermediate nodes at the thhop ( ). The number of source and destinationnodes are the same, i.e., , and each node will betagged with a distinct number ( ). Fig. 1 illus-trates the overlay network. We define as the fractionof packets of class at node to select the node asits relay at the th hop. We refer to this term as the relayselecting parameter. Note that multiple paths could be selectedfor a class, i.e., . Whenever an interme-diate node is not reachable for class at node then

. Since the total number of intermediate nodesin the th hop is , we have .

Importantly, note that the deployed structure is very generaland any multihop network that can be modeled as a directedacyclic graph can be modified to fit into this overlay structureby simply adding virtual nodes (virtual hops for different users)[9]. We introduce virtual nodes with zero service time for users

that have a smaller number of hops, and fix the path for par-ticular classes to pass through the virtual node (by enforcing

). Fig. 1 gives an example of a 3-hop overlay net-work with two users ( , , , ,

). Methods to construct such overlay structures givena specific multihop network and a set of transmitting-receivingpairs can be found in [19], [20]. Through the multistage overlayinfrastructure, the information feedback is performed from theintermediate nodes to all the connected nodes ( )in the previous hop.

To describe the channel conditions of each transmission linkfrom node to , i.e., , ), we assume as in [18]that each wireless link is a memoryless packet erasure channel.

represents the packet error rate over the link ,) for the class and represent the effective

transmission rate (goodput). They can be approximated usingthe sigmoid function [18]

(3)

(4)

where SINR is the measured Signal-to-Interference-Noise-Ratio(SINR), and and are empirical constants corresponding tothe modulation and coding schemes for a given packet length

of class . represents the maximum trans-mission rate supported by the optimal modulation and codingscheme. To cope with the packet error, at the MAC layer, weassume the network deploys a protocol similar to that of IEEE802.11a/e, which enables packet-based retransmission. Let

represent the maximum number of retransmissionsfor the packets of class over the link ( , ).

C. Centralized Cross-Layer Optimization for Multi-UserWireless Video Transmission

We define as the cross-layer transmission strategyvector for packets at the node consisting of the packet sched-uling policy , the relay selecting parametersfor routing, the MAC retransmission limit per link,i.e., . And

represents the set includingall the feasible cross-layer transmission strategy vector, where

is the set of all feasible packet scheduling strategies,is the set of all possible selections of relays, and

is an integer set from 0 to the maximum retransmission limitsupported by the MAC protocol. Then, assuming the global in-formation is available, the investigated multi-user wire-less video transmission problem can be formulated as a central-ized delay-driven cross-layer optimization

(5)


where , and represents the setof nodes at which the transmission strategies decisions can bemade for the video packets. is the numberof the nodes in . Since the successfully received packets ofeach class must have their end-to-end delay smaller thantheir corresponding delay deadline , the constraint of the opti-mization is Due to the priorityqueuing and the error concealment scheme in (2), the optimalsolution of (5) serves the more important packets instead oftransmitting as many packets as possible. Although the central-ized optimization provides optimal solution for the multi-uservideo streaming problem, it suffers from the unrealistic assump-tion of collecting timely global information across the multihopnetwork for the delay-sensitive applications. Due to the informa-tionally decentralized nature of the multihop wireless networks,the centralized solution is not practical for the multi-user videostreaming problem. For instance, the optimal solution dependson the delay incurred by the various packets across the hops,which cannot be timely relayed to a central controller. More-over, the complexity of the centralized optimization grows ex-ponentially with number of classes and nodes in the network.Hence, the optimization might require a large amount of time toprocess and the collected information might no longer be accu-rate by the time transmission decisions need to be made.

D. Proposed Distributed Cross-Layer Adaptation Based onInformation Feedback

Instead of gathering the global information , we pro-pose a distributed suboptimal solution that collects the local in-formation feedback at the node to maximize the ex-pected quality of the various users sharing the same multihopwireless infrastructure

(6)

where represents the number of packets of classpresent in the queue at the node .

In this paper, we define with the following informationfeedback parameters.

• SINR, the SINR to calculate the channel conditions overeach link of the overlay network [see (3) and (4)].

• , the packet loss probability of the class throughthe intermediate node . The parameter illustrates thebottleneck identification for various video classes. This in-formation can be used by the application layer to decidehow many quality layers are transmitted or to adapt its en-coding parameters (in the case of real-time encoding) toimprove its video quality performance given the currentnumber of users, priorities of the competing streams andnetwork conditions, but also, importantly, to alleviate thenetwork congestion.

• , the expected delay from the intermediatenode to the destination node of the class to conveythe congestion information of the network, which is essen-tial for the delay-sensitive applications.

Fig. 2. Illustrative example of an application layer overlay network with infor-

mation horizon h = 2.

Let us consider the simple example in Fig. 2 that illustrateshow information feedback is deployed. The term informationhorizon will be defined in Section III. In this example, node n1 isan intermediate node that needs to relay multiple video classesfrom various users. In order for the relay n1 to determine theoptimized cross-layer transmission strategies, at least 1-hop in-formation feedback is required. The network status informationcan be disseminated at frequent intervals over the overlay infra-structure, and it is considered to be known at the decision relayn1. However, in certain cases, feedback information from somehops (beyond the information horizon) may arrive with an intol-erable delay, and may be unreliable due to the rapidly-changingnetwork conditions.

In this paper, we make the following assumptions forperforming the information feedback and the delay esti-mation . First, we assume a polling-basedcontention-free media access (which is similar to the deployedIEEE 802.11e [16] and 802.11s [17] standards) that dynam-ically reserves transmission opportunities within a serviceinterval [16], and the network status (such as the topology,the transmission rate and the packet error rate

for each link) remains unchanged in . Second,because of the retransmission in the MAC layer protection,the effective packet transmission time can be formulated as ageometric distribution [30] with , , andpacket length (as discussed in Section III-B). Third, forsimplification, the arrival of the packets at each intermediatenode is regarded as a Poisson arrival process, which is reason-able if the number of intermediate nodes is large enough andthe selection of paths is relatively balanced. Fourth, we assumethat the queue waiting time dominates the overall delay. Underthese assumptions, we can estimate the risk that packets fromdifferent priority classes will not arrive at their destinationbefore their decoding deadline expires (see Section IV for moredetail). The adaptation of , , and the dynamicrouting policies for can be deployed in a distributedmanner based on the information feedback. Next, we discussthe mechanism of performing the information feedback throughthe directed acyclic overlay network.


III. IMPACT OF ACCURATE NETWORK STATUS

Since the network conditions can rapidly vary in multihopnetwork infrastructures, the performance of any video streamingsolution will significantly depend on the availability of accuratenetwork information. Three key aspects for multi-user videostreaming are influenced by the availability, accuracy, and time-liness of this information feedback.

• Decentralized decision making – network nodes can be im-plemented to improve their adopted cross-layer strategiesbased on information feedback about the channel condi-tions and regional network congestion to avoid unneces-sary queuing delay and hence, packet drops.

• Timely adaptation – information feedback enables timelyadaptation to network changes (e.g., nodes leaving orsources of interference appearing or disappearing), whichis essential for delay-sensitive multimedia transmission.

• Inter-user collaboration – based on information feedback,network resources can be effectively managed and usersare able to effectively collaborate to achieve the desiredglobal optimal utility. For instance, in the absence of suchinformation, an intermediate node may waste precious re-sources by allocating time to packets from classes thatwill miss their deadlines, thereby preventing other classeswhich can meet their delay constraint from being trans-mitted.

A. Information Feedback Frequencies and InformationHorizon

The information feedback should be performed in a dis-tributed (per hop) fashion that explicitly considers the dis-semination delay. We assume that the information feedbackis periodically transmitted to the previous hop every 1

seconds during each ( ). We defineas the frequency of the information feedback within one hop

(7)

We also define the vector of the dissemi-nation factors over the network. Let represent the timeit takes for the information to be disseminated over hops

(8)

Since the network information requires time to pass through thevarious hops, we have . We set . Because theinformation is conveyed hop by hop, also depends onthe per-hop information feedback frequency . We define

as the information feedback frequency when the infor-mation is conveyed over hops in the following way:

(9)

where is defined as . Since the network conditions areassumed to be unchanged within the service interval , we

1The time interval is not the time fraction for transmitting the informationfeedback in a service interval, but rather the time between two subsequent in-formation feedbacks (which includes time for transmitting the video packet, theinformation feedback and also the protocol overheads).

define the information horizon as the number of hops fromwhich the information feedback can be accurately disseminatedduring

(10)

In [7] and [18], the dissemination time for the information feed-back is proportional to the number of hops across which the in-formation feedback is traversed, i.e.,

, and if we assume that is an integer, the relationship be-

tween and becomes a linear function

(11)

We focus on the impact of different information horizons di-rectly on the video qualities of multiple users sharing the samemultihop wireless network. Note that can be convertedinto an information horizon based on (10), as long as the infor-mation dissemination factors (i.e., vector) are given. Thus, forsimplicity, in the remainder part of the paper, we denote the in-

formation horizon by simply . An example of

is shown in Fig. 2. The local information feedback in (6)for a larger information horizon becomes a vector

where represents a set of nodes in the th hop thatfeedback the information for the class traffic to the decisionnodes (e.g., node “n1” in the example in Fig. 2).

B. Impact of Various Information Horizons

With a larger information horizon, more accurate networkstatus can be obtained, which can be used to adapt the cross-layer transmission strategies at various layers. A larger infor-mation horizon ensures that the information can be obtained ina timely manner and network status can be estimated more accu-rately. For example, a better routing decision can be determinedto avoid the congested regions in the network. This decreasesthe packet loss probability for each class, thus increasingthe for the important classes and improving the receivedvideo qualities. However, the penalty of the overhead is seldom

jointly considered in the prior works. Letrepresent the expected transmission time for a video packet inclass at node to the next hop with the information feed-back of horizon . Based on the geometric assumption, wecan write (12), as shown at the bottom of the next page, which iscalculated as an average transmission time over all the possible

relays in the next hop. denotes the time

overhead introduced by the various protocols [16] including thetime of waiting for the MAC acknowledgements etc., and alsothe information feedback. Consequently, a larger informationhorizon can induce larger overheads for the packet transmissiontime, and hence increases the end-to-end delay , which canlead to higher packet losses as they can miss their deadline[see (2)]. In this paper, we assume that the time overhead is a


known function of the information horizon and we will discussthis in more detail in Section VII.

In general, the information horizon might be different for var-ious users or classes and also can vary per node, depending onits location, congestion level, etc. Thus, a scalable informationfeedback can be implemented (e.g., the information horizon candepend on class and node ). For instance, to reduce theoverhead associated with the information feedback, some lessimportant classes can have smaller horizons. However, for sim-plicity, the information horizon is assumed to be the same forall classes (users) in the rest of the paper. The topic of imple-menting the scalable information feedback and the analysis ofits impact form a topic of our future research.

C. Distributed Cross-Layer Adaptation Based on theInformation Feedback With Larger Information Horizon

Instead of performing the exhaustive search for the dis-tributed optimization in (6), we present the following iterativecross-layer adaptation to solve the multi-user video streamingproblem. Based on the information feedback, the goal of thedistributed cross-layer adaptation is to determine an optimalpacket present in the queue (from ) to be transmittedthrough the optimal relay (from ) in the nexthop with the optimal retransmission limit (from ).

1) To determine a packet of a specific class for transmission,the packet scheduling policy in the queue of theintermediate node is optimized to first transmit the videopackets with larger , since they have a higher impact on theoverall video quality. With a larger information horizon, suchpacket scheduling can be improved as we will discuss it inSection IV.

2) To solve the routing problem, we deploy a priority queuingapproach based on the information feedback and apply dynamicrouting policies similar to the Bellman-Ford routing algorithm[10]. We exploit the in the local informationfeedback . The selection of is based on the

value that minimizes the end-to-end packetloss probability for the transmitted packet. We will discussthe routing problem in detail in Section VI.

3) At the MAC layer, we choose the appropriate retransmissionlimit per packet based on the suchthat its delay constraint is satisfied. Based on our prior results[11] in one-hop network, the optimal retransmission strategyis to send the highest priority packet until it is successfullyreceived by the next relay or until its delay deadline expires.Specifically, let represent the current delay incurred bya particular packet at the current node . The maximumretransmission limit for the packet of class over the link

from to is determined based on the delay deadline(where is the floor operation)

(13)

With a larger information horizon, the retransmission limit canbe improved as we will discuss it in Section V.

4) Then, we measure the SINR and estimate the correspondingand for each class at the node and

feed back this information to the nodes in the previous hops

within the information horizon .

IV. INFORMATION FEEDBACK DRIVEN SCHEDULING OF

PRIORITIZED VIDEO PACKETS

At each intermediate node , in order to optimize thescheduling of the various video packets, we determine the risk

( ) that the packets of classwill miss their delay deadline, based on the probability that theestimated received time at the destination is after their delaydeadlines. Higher probabilities of packet loss over the network(due to interference, congestion, nodes leaving etc.) will leadto higher risks of packets missing their delay deadlines. Basedon this risk, the scheduling of the various packets of the dif-ferent classes can be determined to ensure a maximized systemquality.

To compute the risk estimation for a packet, we need to con-sider both the delay deadlines as well as the expected delay

in the information feedback conveyed

from the intermediate nodes within the information horizon .The video packets at an intermediate node can be divided intothree categories:

• packets that will certainly be dropped (“dropped” packets);• packets that have very high probability to be dropped (“al-

most-dropped” packets);• packets that have low probability to be dropped (“seldom-

dropped” packets).“Dropped” packets are video packets with a current cumula-tive delay exceeding their delay deadline ( ).These packets will be dropped at the current node and hence,there is no need to compute their risk. The “almost-dropped”packets have not yet exceeded their delay deadline (

), but their current cumulative delay plus the expected delayto reach the destination does exceed their delay deadline, i.e.,

. We set the risks for these “al-most-dropped” packets to be 0, as they have a very high proba-bility of being dropped and hence, they will unnecessarily wasteresources that could be used for the successful transmissionof “seldom-dropped” packets. The remaining video packets are

(12)


“seldom-dropped” packets. Their current cumulative delay plusthe expected delay from the current node to the destination islower than the delay deadline, i.e., .Hence, these packets have a high probability of arriving at thedestination on time and their scheduling needs to be optimizedto maximize the video quality across the various users. Next,we discuss how to estimate the risk for these seldom-droppedpackets.

A. Risk Estimation Based on Priority Queuing Analysis

The risk estimation for the seldom dropped packets is deter-mined based on the priority queuing analysis, by using the ap-proximation of the waiting time tail distribution. Let rep-resent the queue waiting time for class at intermediate node

. The waiting time tail distribution can be approximated as[13], [14]

(14)

where is the measured average input rate andis the average service time of class at the intermediate node

. The expected average queue waiting time of the priorityqueue is

(15)

Equation (15) is determined based on the Mean Value Anal-ysis (MVA) of a preemptive-priority M/G/1 queue [10]. Untilnow, we do not consider the interference incurred in wire-less multihop networks (orthogonal transmission channelsare available for adjacent wireless links), the average ser-vice time is the average packet transmission time

in (12). If the influence of interference is

considered, the average service time can be approx-imated using a virtual queue analysis similar to the “serviceon vacation” concept in queuing theory [10], [15]. Using (14),the proposed risk estimation2 for the packets in class canbe computed as (16), shown at the bottom of the page, where

represents the expectedtime remaining after a packet reaches its destination. We candetermine the probability that the waiting time plus apre-determined time duration , which is a general vari-able for risk estimation, exceeds the expected time left ,and thus, that the packet will be lost. The time durationcan be viewed as an extension of the waiting time for the packet.Larger values lead to higher risks. An example of therisk estimation is given in Section IV-B. Note that the accuracyof computing the expected time left increases with

a larger information horizon. Thus, the

also depends on the information horizon and can be better

estimated given a larger .

B. Information Feedback Driven Scheduling

In a priority queue, the packet scheduler at an intermediatenode transmits first the most important packets (i.e., the packetswith the largest ). Each packet is transmitted until the packetis successfully received by the next hop node or until its deadlineexpires. Assume that there are total video packets at the inter-mediate node . Let the application layer packet scheduling

, where represents the sched-uling order for the video packet . The basic pri-ority scheduling can be written as

(17)

where is the number3of packets of theclass that are transmitted during a period of timeusing a specific packet scheduling . The notation

2The higher risk packets should be sent earlier, since they are with high proba-bility to exceed their deadlines. However, we do not want to waste our resourceson those almost-drop packets, hence the risk estimation for these packets are setto zero.

3Packet loss is considered in this number due to the delay constraint that dropspackets.

ifif

if

if

(16)


Fig. 3. System map for the IFDS packet scheduling.

indicates that the packet is not scheduled due to its deadlineexpiration.

A packet could be dropped in the future hops, as its deadline isexceeded at these hops, and the transmission time of this packetis wasted. This may results in the loss of other packets thatwould have arrived on time at their destination. Thus, enabledby the information feedback, an intermediate node gathers thenetwork status and makes a scheduling decision. Instead of al-ways transmitting the most important packet in the queue, someother video packets of the different users that are less importantbut have a higher packet loss probability (risk) can be sent first.Based on this, we propose a novel Information Feedback Drivenpacket Scheduling (IFDS). The system map of the IFDS sched-uling at an intermediate node is illustrated in Fig. 3. The riskis estimated using the information feedback andthe waiting time distribution [see (16)].

For the IFDS scheduling, the video packets ordered inare transmitted for a pre-determined period of time

. The IFDS scheduling is determined as (18),shown at the bottom of the page. As opposed to the priorityqueuing scheduling ((17)), the risk of losing a certain class

is considered jointly with the packetquality impact. The scheduler sends the packets in the order thatmaximizes the output video quality weighted bywithin the time interval . Since different traffic classeshave different packet transmission times [see(12)], the number of packets being transmitted per class

depends on which packets are sent

(scheduling decision). However, theremains constant and is independent of the scheduling decisionwithin . Recall that with a larger information horizon

, the risk is estimated more accurately because the node is

able to obtain more accurate information from nodes which arecloser to the destination. Hence, the packet scheduling policy

is more accurate and adaptive to the network changesthan the priority scheduling strategy of (17). Finally, the IFDSscheduling has the following constraint:

(19)

where the notation represents that packet is scheduledbefore packet . If belongs to user , the is a classdependent threshold, which can be defined as

(20)

Equation (20) provides a threshold for a particular class, whichis the quality impact value of the next important class of thesame user. The reason for the constraint in (19) is to avoidsending an unimportant class with high risk (i.e., for the classesof the same user, packets with higher must be sent first). Thisis important since the less important classes depend on the moreimportant classes of the same user and hence, their distortionwill be significantly impacted if the higher priority packets arelost [11] [see (2)].

An example of the risk estimation at an intermediate node

with fixed is given in Fig. 4 for a case of two users andfour classes with the quality impact parameters

. User 2 (with classes and ) has a smaller ex-pected time left than user 1 (having classes and ).Note that when , forall the classes, because they miss their deadlines after waitingfor . Let us now adopt the IFDS packet scheduling algo-rithm, and set the between and .From Fig. 4, we can observe that and

. Hence, the packets of class canwait for without significantly increasing the packet loss,while the packets of class that are less important ( )are transmitted.

From the example, we see that the setting of affectsthe risk estimation and hence the scheduling decision. Note thatif we set larger than the maximum delay deadline of allthe users, the risk will be 1 for all the seldom-dropped packets,and thus the information feedback driven scheduling will onlydepend on . If is set too small, the risk estimations

(18)


Fig. 4. Risk estimation versus time interval for two users.

will not affect the original priority decision. Thus, we define alower and an upper bound of the

(21)

since the risk estimations are large enough to take effect withinthis interval. For the example in Fig. 4

.

V. INFORMATION FEEDBACK DRIVEN MAC LAYER

RETRANSMISSION STRATEGY

For protection over an error-prone wireless link, a retransmis-sionschemeat theMAClayerisadopted.In[11], itwasshownthatfor thescalablevideocoderssuchas[23], thevideopacketsshouldbe retransmitted by the MAC until they are received without erroror their deadline expires in order to maximize the received videoquality. However, if a packet approaches its delay deadline, therisk that it will not reach its destination increases. Hence, simi-larly to the application layer scheduling strategies discussed inthe previous section, we propose a MAC layer information feed-

back driven retransmission strategy that explic-itly considers the risk of losing a packet based on the availableinformation feedback .

Let be an integer variable that represents the numberof retransmissions for a packet. If the transmission of the packetrepeatedly fails, the retransmission should last only until anotherclass of video packets starts to have a higher impact in terms ofoverall video quality. In both scheduling policies in the previoussection, the scheduler will send packets of class having alarger value [see (18)]. Therefore, the informationfeedback driven retransmission limit becomes

(22)

which states that the retransmission limit is the maximumnumber of retries such that the transmitting packet (of class

) has a greater than other packets in the queue.Due to the scheduling constraint in (19), we only need to checkthe classes that have a quality impact value larger than thethreshold in (22). Note that the information feedbackdriven retransmission limit is always smaller than the retrans-mission limit in (13) ( ), since whena packet approaches the deadline, it will first belong to the“almost-dropped” packets class ( ),for which . Thus, another class of packets willbe transmitted, thereby terminating the retransmission of thecurrent packet. Consequently, a packet retransmission will firstreach the information feedback driven retransmission limit

before the delay deadline. Thus, other packets thathave a better chance to reach the destinations could be sentearlier.

VI. ADAPTIVE DYNAMIC ROUTING BASED ON INFORMATION

FEEDBACK

A. Self-Learning Dynamic Routing With 1-Hop InformationFeedback

In this section, we propose a dynamic routing policy for therelay selecting parameters . The decision is basedon the information feedback of the expected delay from the cur-rent node to the destination, , for each classof traffic (class ; see Fig. 2). The decisions are made by thefollowing policy:

(23)

are normalized coefficients to ensure that the summa-tion equals to one

(24)

with and being constants, which are determined similarlyto the balking arrival probability in queuing theory [15]. Thevalue of is set depending on the arrival rate according to [15].The term weights the average delay suchthat the routing policy favors paths leading to a significant lowerdelay to the destination. Recall that represents a set ofnodes in the th hop that feedback the information

. We set for the nodes whoseinformation feedback is not received, indicating that nodeis not connected to node using the overlay infrastructure [8].We refer to this relay selecting policy as self-learning policy,since the decision of will influence the future infor-mation feedback. Recall that , hence therelay selecting parameters provide a routing description acrossthe network with multipath capability.

The expected delay to the destination of each class is periodi-cally updated at each node using the information feedback fromthe next hop. If the current node is node at the th hop, the


Fig. 5. Self-learning dynamic routing algorithm for large information horizon.

expectation of delay to the destination of each class is as fol-lows:

(25)

where is provided from the information feed-back of the nodes of the next hop, and the relay selecting param-eter is calculated based on (23). is theaverage queuing delay at the current queue, which can be eithercalculated from (15), or measured from the average queue size.Fig. 5 gives the algorithm of the proposed self-learning policy.The self-learning policy will dynamically adapt the relay selec-tion to minimize the delay through the network.

This method is inspired by the Bellman-Ford shortest path(delay) routing algorithm [10] that minimizes the end-to-enddelay across the network. Our routing algorithm reduces to theBellman-Ford algorithm when to the node

that feedbacks the smallest . The dif-ferences are that our distributed routing policy enables multi-path capability, and the delay of class is influenced by onlythe same or higher priority traffic. Since the packet lossesresult from the violation of the delay constraint, the minimumend-to-end delay provided by our routing policies leads tothe minimum end-to-end packet loss [see (2)]. Combined withthe optimal packet scheduling and the retransmission limit (tosatisfy the delay constraint), the self-learning policy can conse-quently maximize the of the traffic class with higherquality impact [see (6)].

B. Dynamic Routing With Larger Information Horizon

The self-learning policy is the shortest delay routing policy

with one hop information horizon ( ). However, if thenetwork conditions or topologies change, this informationwill be fed back with a certain delay. A larger informationhorizon ensures that any network changes in the horizon canbe timely learned and the cross-layer transmission strategiescan quickly adapt to the latest network conditions. With the

information feedback parameters , if somesignificant changes happen in the network at hop (such asnode failure, link status change etc.), the corresponding relayselecting parameters are recalculated based onthe new information feedback . Note that theexpected waiting time and the expected delay toward desti-nation ( from (15), and from(25)) depend on and can then be recalculated

recursively. Define as a vector of relay selectingparameters of the nodes between the changing

hop and the hop (if , otherwise therecalculation ended at the source nodes)

(26)

In Fig. 5, we provide the corresponding self-learning dynamicrouting algorithms for the increased information horizon.

C. Channel-Aware Self-Learning Routing Policy

In this section, we provide another dynamic routing policycalled channel-aware self-learning policy that not only deter-mines the relay selecting parameters based on the informationfeedback parameters (to avoid congestedarea), but also depends on the channel conditions of transmit-ting packets to different relays. Denote the operationas the smallest choices from a set . Assume is theset of the intermediate nodes that are selected as the relays forclass at the hop

(27)Note that if , then the algorithm is again equivalent to theBellman-Ford shortest delay algorithm, since the only relay thatwill be selected is the one with the smallest . If

, we determine the relay selecting parameters to minimizethe second moment of the service time of the class at node

if

if(28)

Note that the second moment of the service time is

(29)Using the same soft minimum (probabilistic) approach in theprevious two sections, we rewrite (28) as

if

if(30)


are normalized coefficients to make sure that thesummation equals to one

(31)

This routing policy is different from the self-learning routingpolicy that minimizes the end-to-end cumulative queuing delay.The channel-aware self-learning routing policy distributes thetraffic through the links with better channel conditions that di-rectly affect the , and the influence ofchannel conditions can be emphasized by using a larger

.The channel-aware self-learning routing policy has a prop-

erty to reduce the uncertainty of the packet loss probability.Note that the packet loss probability has the following upperbound of (32), shown at the bottom of the page. The new routingpolicy selects the relay selecting parameters todecrease the upper bound of the packet loss probability. First,we reduce the term byenforcing only for nodes with a smaller

[see (27)]. Second, we try to decrease thesecond moment of the queue waiting time , which isrelated to the higher moments of the service time distribution,i.e., are selected to have a service time (transmis-sion time of a packet) with small variation. In order to decreasethose higher moments of the service time distribution, besideslimiting the number of nonzero relay selection parameters forthe next hop nodes, we also determine by mini-mizing to decrease the variation of the service time.

The channel-aware self-learning routing policy for a largerinformation horizon can also apply the same algorithm stated inFig. 5 by substituting (26) with (30).

D. Proposed Dynamic Priority Hybrid Routing Algorithm

The self-learning dynamic routing policy introduced inSection VI-A attempts to minimize the average end-to-end delaytoward the destination, while the channel-aware self-learningrouting policy introduced in Section VI-C attempts to decreasethe upper bound of the packet loss probability (resulting fromdeadline expiration). Usually, the important video packets(with larger quality impact ) will be sent first. To reducethe uncertainty of the packet loss rate is more important tothese packets, since they experience significantly small av-erage queuing delay (as well as the packet loss rate). Thechannel-aware self-learning routing policy is more suitablefor these packets. On the other hand, some other packets ofthe less important classes experience larger queuing delay in

the network, and the self-learning routing policy that ensuresshortest delay is essential for these packets. Consequently, witha pre-determined threshold , we choose the channel-awareself-learning routing for the important classes with ,and the self-learning dynamic routing for the rest of the classes.We refer to this hybrid routing policy as Dynamic PriorityHybrid (DPH) routing algorithm. Our DPH routing can alsoapply the algorithm in Fig. 5 for larger information horizons.

VII. OVERHEAD ANALYSIS FOR INFORMATION FEEDBACK

The information feedback can enable the cross-layer adapta-tion of video streaming over a multihop network. As the infor-mation horizon increases, the network status can be estimatedmore timely and accurately, and the cross-layer strategies can beimproved for the delay-sensitive applications. However, a largerinformation horizon also consumes more network resources forvideo transmission and results in an increased time overhead per

packet transmission, [see (12)]. Various infor-

mation feedback parameters have different transmission over-heads. In this paper, we take the three information feedback pa-rameters illustrated in Fig. 2 as examples.

Assuming a certain topology, let us perform a worst-caseanalysis to quantify the maximum information feedback.We assume that the information feedback overheads are

for the three information feedbackparameters, respectively. We assume that the average number ofnodes in one hop is , the number of total classes is , and we

set the information horizon as for all users (classes). TheSINR information is fed back from potential receivers to thetransmitters to enable the link adaptation as well as to facilitatethe polling control signaling. Thus, an information horizon ofonly 1 hop is sufficient for the adopted overlay infrastructure,and the overhead in terms of the information feedback unit is

. As for the other two information feedback parame-ters, the parameters are required across the whole informationhorizon and different for all the classes. An aggregation scheme

can be applied to reduce the repeated information (as ine.g., [8], [31]). The worst-case overheads in terms of the in-

formation feedback unit are and

, respectively.

and represents the functions of aggregated

information feedback over hops for these two infor-mation feedback parameters. In conclusion, the informationfeedback overhead increases with the information horizon.

VIII. SIMULATION RESULTS

To assess the importance of information feedback, we con-sider several multi-user video transmission scenarios. Two

(32)


TABLE IDESCRIPTION FOR THE FOUR CASES OF THE SIMULATION RESULTS ( t = 100 ms)

TABLE IISIMULATION RESULTS FOR IFDS SCHEDULING WITH VARIOUS INFORMATION HORIZONS AND DIFFERENT NETWORK EFFICIENCIES

Fig. 6. Simulation settings of a 6-hop overlay network with two video se-quences.

video sequences, “Mobile” and “Coastguard” (16 frames perGOP, frame rate of 30 Hz, CIF format) compressed using ascalable video codec [23] are sent from distinct sources totheir corresponding destinations through the multihop wirelessnetwork shown in Fig. 6. We consider four different scenarioswith various information horizons and information feedbackoverheads as stated in Table I. Each video sequence is dividedinto four classes ( , ). The quality impact parame-ters and the number of packets in one group of picturefor each class are shown in the left side of Table II.

In our simulation, we captured the packet-loss pattern underdifferent channel conditions (described in the paper by thelink SINR) using our wireless streaming test-bed [8]. In thisway, we can assess the efficiency of our system under realwireless channel conditions and link adaptation mechanismscurrently deployed in state-of-the-art 802.11a/g wireless cardswith 802.11e extension. Link adaptation selects one appro-priate physical-layer mode (modulation and channel coding)depending on the link condition, in order to continuouslymaximize the experienced goodput [8]. Hence, each link inour network settings shown in Fig. 6 is assigned with aneffective transmission rate measured from the test-bed. Theparameter represents the streaming efficiency of thenetwork. The various efficiency levels are represented byvarying the available time fraction for the contention-freeperiod in the polling-based MAC protocol, which inducesthe various available transmission rates for the video packetsover the links. In our event-driven simulation, these networkefficiency levels range from 300 Kbps to 500 Kbps. A larger

gives higher network efficiency. We set ,

(see Section III) and varies from 1 to 4 for thefour scenarios. The information feedback overheads are set as

for all the classes. Note thatthe time overhead is limited, i.e., 2.5% of the average packet

transmit time when , and .


A. IFDS Scheduling and Retransmission Limit

Note that the effect of the IFDS scheduling depends onmany factors, such as the network topology, applicationcharacteristics, network transmission efficiency, and conges-tion/interference conditions, etc. Here, we would like to assessthe importance of the risk consideration in resource-constrainednetworks. We set the application playback delay deadlines areset to 500 ms and 300 ms for the classes of the two videosequences respectively. The transmission rates of the links inthe first hop are, relatively higher than the subsequent links.Consequently, most of the packets of the various classes willbe queued at the specific intermediate nodes n1 and n2 (someof them will still be left in the source queues), and the effectof risk can be highlighted for two streams with different delaydeadlines.

We adopt the IFDS scheduling and the retransmission limitalgorithm in Sections IV and V for cases with larger informa-

tion horizons ( ). In scenario 1, we make the packet sched-uling first transmit the packets with the highest quality impactparameter until the transmission success or delay deadlineexpiration [i.e., (17)]. In scenario 2, the risk estimation is con-sidered jointly with the quality impact parameters using (18). Inscenarios 3 and 4, larger information horizons are used in (18)for risk estimation. However, with larger information horizon,the performance degrades due to larger information feedbackoverheads. The simulation results of the packet loss rate of eachclass at their destinations are shown in Table II under variousnetwork transmission efficiencies. Since the delay deadline ofthe “Coastguard” sequence is smaller, it has higher packet lossrate, especially in networks with low transmission efficiency.However, it is shown that as the information horizon increases,the IFDS scheduling sends more “Coastguard” packets to im-prove its video quality without degrading significantly the videoquality of the “Mobile” sequence.

To observe the impact of the various information horizons onthe overall video quality, the average Y-PSNR decoded at thedestinations of the two sequences are shown in Fig. 7. It showsthat the optimal choice of information horizon varies with thenetwork transmission efficiency. For networks with high trans-

mission efficiency, a larger information horizon ( ) makesthe IFDS scheduling more efficient, and improves the videoqualities. However, for a network with low transmission effi-ciency that is more congested, a shorter information horizon

( ) results in better performance since the limited networkresource can be focused on the video transmission (payload).

B. Adaptive Dynamic Routing

We compare our DPH routing algorithm with an on-demandmultipath routing algorithm AOMDV [29] that selects severalloop-free link-disjoint paths based on a well-known on-demandrouting algorithm AODV [26]. For the on-demand routing,each video source requires a route discovering period (dependson the round-trip delay to the destination; here, we assume 100ms) [29] before streaming the video. Although this on-demandrouting saves additional time overheads for packet transmis-sions, it sacrifices the tolerable delay when transmitting thedelay-sensitive video packets. In Table III, we also compare the

Fig. 7. Y-PSNR versus various information horizon cases under different net-work transmission efficiencies.

Fig. 8. Overlay network simulation with teo video streams over the wirelessmesh network.

TABLE IIICOMPARISON OF THE DPH ROUTING POLICY WITH THE ON-DEMAND

MULTIPATH-ALGORITHMS

proposed DPH routing with the on demand routing upon a moregeneral overlay structure over wireless mesh network shownin Fig. 8. The same two video sequences are streamed fromnode S1 to D1 and from node S2 to D2 with delay deadline 500ms, respectively. In summary, the simulation results show thatour dynamic routing policy outperforms existing on-demandrouting solutions by more than 2 dB in terms of the receivedvideo quality in low efficiency wireless networks.

IX. CONCLUSION

In this paper, we investigate the impact of information feed-back with different network horizons on the video quality ofmultiple users sharing the same multihop wireless network. Weillustrate how the various cross-layer strategies can be adaptedto take advantage of the available information feedback from alarger network horizon through the proposed information feed-back driven scheduling, retransmission limit and the dynamic


priority hybrid routing algorithm. Unlike the end-to-end feed-back that exists in today’s networking protocols (such as therate control in TCP), the information feedback is performed in adistributed (per hop) fashion that explicitly considers the instan-taneous delays, which is essential for supporting delay-sensitivemultimedia applications. We investigate the tradeoff betweenthe increased transmission overhead and the benefit of largerinformation horizons leading to an improved prediction of net-work conditions. The results show that in a network with highertransmission efficiency, a larger information horizon can leadto an improved performance in terms of video quality, whichleads to more than 2 dB improvement in video quality as a re-sult of balancing the effect of different delay deadline amongusers. However, with lower transmission efficiency, smaller in-formation horizon performs better by ensuring limited overheadof the information feedback.

REFERENCES

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

[2] 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., pp. 59–65, Aug. 2005.

[3] B. Awerbuch and T. Leighton, “Improved approximation algorithmsfor the multi-commodity flow problem and local competitive routing indynamic networks,” in Proc. 26th ACM Symp. Theory of Computing,May 1994.

[4] S. Toumpis and A. J. Goldsmith, “Capacity regions for wireless Ad hocnetworks,” IEEE Trans. Wireless Commun., vol. 2, no. 4, pp. 736–748,Jul. 2003.

[5] A. Butala and L. Tong, “Cross-layer design for medium access con-trol in CDMA Ad hoc networks,” EURASIP J. on Applied Signal Pro-cessing, to be published.

[6] 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. Select. Areas Commun., vol. 23, pp. 136–150, Jan. 2005.

[7] Y. Andreopoulos, N. Mastronarde, and M. van der Schaar, “Cross-layeroptimized video streaming over wireless multi-hop mesh networks,”IEEE J. Select. Areas Commun., to be published.

[8] D. Krishnaswamy and J. Vicente, “Scalable adaptive wireless networksfor multimedia in the proactive enterprise,” Intel Technol. J..

[9] J. R. Evans and E. Minieka, Optimization Algorithms for Networks andGraphs. New York: Marcel Dekker, 1993.

[10] D. Bertsekas and R. Gallager, Data Networks. Upper Saddle River,NJ: Prentice-Hall, 1987.

[11] M. van der Schaar and D. S. Turaga, “Cross-layer packetization andretransmission strategies for delay-sensitive wireless multimedia trans-mission,” IEEE Trans. Multimedia, vol. 9, pp. 185–197, Jan. 2007.

[12] D. Qiao, S. Choi, and K. G. Shin, “Goodput analysis and link adaptationfor IEEE 802.11a wireless LAN,” IEEE Trans. Mobile Comput., vol. 1,no. 4, pp. 278–292, Oct.-Dec. 2002.

[13] T. Jiang, C. K. Tham, and C. C. Ko, “An approximation for waitingtime tail probabilities in multiclass systems,” IEEE Commun. Lett., vol.5, pp. 175–177, Apr. 2001.

[14] J. Abate, G. L. Choudhury, and W. Whitt, “Exponential approxima-tions for tail probabilities in queues I: Waiting times,” Oper. Res., vol.43, no. 5, pp. 885–901, 1995.

[15] L. Kleinrock, Queuing Systems Volume I: Theory. New York: Wiley,1975.

[16] Draft Supplement to Part 11: Wireless Medium Access Control (MAC)and Physical Layer (PHY) Specifications: Medium Access Control(MAC) Enhancements for Quality of Service (QoS) 2003, IEEE802.11e/D5.0.

[17] S. M. Faccin, “Mesh WLAN networks: Concept and system design,”IEEE Wireless Commun. Mag., pp. 10–17, Apr. 2006.

[18] 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.

[19] M. Waldvogel and R. Rinaldi, “Efficient topology-aware overlaynetwork,” ACM SIGCOMM Comput. Comm. Rev., vol. 33, no. 1, pp.101–106, Jan. 2003.

[20] J. Jannotti, “Network-layer support for overlay networks,” in Proc.IEEE Conf. Open Architectures and Network Programming, Jun. 2002.

[21] M. van der Schaar, Y. Andreopoulos, and Z. Hu, “Optimized scal-able video streaming over IEEE 802.11a/e HCCA wireless networksunder delay constraints,” IEEE Trans. Mobile Comput., vol. 5, no. 6,pp. 755–768, Jun. 2006.

[22] A. Ortega and K. Ramchandran, “Rate-distortion methods for imageand video compression,” IEEE Signal Process. Mag., vol. 15, no. 6,pp. 23–50, Nov. 1998.

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

[24] H.-P. Shiang and M. van der Schaar, “Multi-user video straming overmulti-hop wireless networks: A cross-layer priority queuing approach,”in Proc. IEEE Conf. Intelligent Information Hiding and MultimediaSignal Processing (IIH-MSP 2006), Dec. 2006, pp. 255–258.

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

[26] C. E. Perkins and E. M. Royer, “Ad hoc on-demand distance vectorrouting,” in Proc. 2nd IEEE Workshop on Mobile Computing Systemsand Applications, Feb. 1999, pp. 90–100.

[27] P. A. Chou and M. Zhourong, “Rate-distortion optimized streaming ofpacketized media,” IEEE Trans. Multimedia, vol. 8, pp. 390–404, Apr.2006.

[28] D. Jurca and P. Frossard, “Packet selection and scheduling for multi-path video streaming,” IEEE Trans. Multimedia, vol. 9, Apr. 2007.

[29] M. K. Marina and S. R. Das, “Ad hoc on-demand multi-path distancevector routing (AOMDV),” in Proc. Int. Conf. Network Protocols(ICNP), 2001, pp. 14–23.

[30] A. G. Konheim, “A queuing analysis of two ARQ protocols,” IEEETrans. Commun., vol. COM-28, Jul. 1980.

[31] B. Krishnamachari, D. Estrin, and S. Wicker, “The impact of data ag-gregation in wireless sensor networks,” in Proc. IEEE Int. Conf. Dis-tributed Computing Systems Workshop, 2002, pp. 575–578.

Hsien-Po Shiang received the B.S. and M.S. degree in electrical engineeringfrom National Taiwan University, Taipei, Taiwan, R.O.C., in 2000 and 2002,respectively. He is currently pursuing the Ph.D. degree in the Electrical Engi-neering Department, University of California at Los Angeles.

During his graduate study, he was with Intel Corp., Folsom, CA, in 2006, andresearched the overlay network infrastructure over wireless mesh networks. Hisresearch interests are the cross-layer optimizations/adaptations for multimediatransmission over wireless mesh networks, and the dynamic routing based oncollaborative information exchange for delay-sensitive applications.

Mihaela van der Schaar (SM’04) is currently an Assistant Professor in theElectrical Engineering Department at the University of California, Los Angeles(UCLA). Since 1999, she was an active participant to the ISO MPEG standardto which she made more than 50 contributions and for which she received threeISO recognition awards.

Dr. van der Schaar received the NSF CAREER Award in 2004, IBM FacultyAward in 2005, Okawa Foundation Award in 2006, Best IEEE TRANSACTIONS

ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY Paper Award in 2005,and Most Cited Paper Award from EURASIP Journal Signal Processing:Image Communication between the years 2004–2006. She holds 28 grantedU.S. patents. She is currently an associate editor of IEEE TRANSACTIONS ON

CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, IEEE SIGNAL PROCESSING

LETTERS and IEEE Signal Processing e-Newsletter. She is also the editor(with Phil Chou) of the book Multimedia over IP and Wireless Networks:Compression, Networking, and Systems.

Informationally Decentralized Video Streaming Over Multihop

Documents