Admission Control and QoS for Continuous MediaDisplays in MANETs

Shahram GhandeharizadehDepartment of Computer ScienceUniversity of Southern California

Los Angeles, CA 90089, USAEmail: shahram@usc.edu

Shyam KapadiaDepartment of Computer ScienceUniversity of Southern CaliforniaLos Angeles, CA 90089, USA

Email: kapadia@usc.edu

Bhaskar KrishnamachariDepartment of Computer Science

Department of Electrical EngineeringUniversity of Southern CaliforniaLos Angeles, CA 90089, USA

Email: bkrishna@usc.edu

Abstract— We consider continuous media delivery over a mo-bile ad-hoc network of vehicles equipped with car-to-car peer-to-peer (C2P2) devices. While the provision of high-bandwidth con-tinuous media content with tight QoS requirements is challengingeven in static networks, it is considerably more challengingwhen the network topology is dynamic due to node mobility.In this paper, we develop a unified client-centric, distributedadmission control framework for such a C2P2 network. Underthis framework, we develop several admission control strategies,namely, server count-based admission (SC), server bandwidth-based admission (SB), path bandwidth-based admission (PB),and finally a mobility-based admission (MAD [4]) policy, whichcan be further enhanced using a second-level sampling-basedadmission policy. We also develop QoS utility models to quantifythe performance of these policies. These policies are then eval-uated through extensive experimental simulations. Our resultsshow conclusively that traditional admission control strategiesdo not work, and that MAD can provide orders of magnitudeimprovement when compared with no admission control.

I. INTRODUCTION

During the past few years, automobile manufacturers havebeen marketing and selling vehicles equipped with entertain-ment systems. These systems typically consist of a DVDplayer, a fold-down screen, a video game console, and wirelessheadphones. In its present form, storage and content are tiedtogether. This limits the number of available titles to thoseDVDs and CDs in the vehicle. We envision a separationof storage and content where content is staged on demandacross the available storage for previewing. This would providepassengers access to a large repository of titles. In this vision,vehicles are equipped with car-to-car, peer-to-peer (C2P2)devices which form a mobile ad-hoc network (MANET).Each C2P2 is equipped with abundant amount of storage, aprocessor, and a wireless networking card. A C2P2 mightintegrate into the navigation system of a vehicle and itsexisting network for delivery of data to the on-board fold-down screen or wireless headphones. Different C2P2 devicesmight store different clips and exchange these clips with oneanother to support on-demand delivery of continuous media.

The principle characteristic of continuous media such asvideo and audio is their high sustained bit rate requirement. Ifa system delivers a clip at a rate lower than its pre-specifiedrate without special precautions (e.g., pre-fetching), the usermight observe frequent disruptions and delays with video and

random noises with audio. These artifacts are collectivelytermed hiccups. A hiccup-free display is the ideal quality ofservice (QoS) provided to an end user. A second QoS criterionis the observed startup latency, defined as the delay observedfrom when a user references a clip to the onset of display.

All C2P2 devices may replicate popular titles, but they mustcollaborate in order to provide users with a large choice ofcontent. Each C2P2 may contribute a fraction of its storage ina peer-to-peer manner to be occupied by the system-assignedclips. These clips might be referenced by users of those mobileC2P2 devices that are network reachable (including the localuser). This study focuses on the QoS metrics when a userreferences a clip that is available on one or more remoteC2P2s. A C2P2 must offer its passengers only those clipsthat it can download and play in a hiccup-free manner withina reasonable amount of time (e.g. by satisfying a maximumtolerable startup latency constraint).

Techniques that address these challenges are impacted bythe characteristics of the MANET, placement and deliveryscheduling of data, and admission control policies in supportof a hiccup-free display. Let us consider each in turn. Mobility,a key MANET characteristic, is the primary challenge becauseit dictates the life-time of paths between a producer and aconsumer of data. When the display of data is overlappedwith its delivery, a path repair may incur delays that resultin data starvation and hiccup. In addition, topology changesimpact the availability of both data [6] and bandwidth. Theavailability of data is also impacted by its placement across thenodes and its degree of replication [7]. One may replicate dataat the granularity of either a clip [3] or a block [5]. The systemmay replicate the first few blocks of a clip more frequently.This enables the display of a remote clip to consume the clip’sfirst few blocks from either local storage or C2P2 devices thatare a few hops away, minimizing startup latency.

Techniques to schedule delivery of data impact the QoSobserved by active displays. As we shall demonstrate (seeAppendix), content replication across multiple servers andblock-switching techniques provide a significant degree ofprotection against the frequent disruptions and topologicalchanges caused by the mobile environment, by dynamicallyproviding connections to proximate servers.

Finally, given the high data-rate requirements of continuous

media displays and limited bandwidth resources, the systemmust be configured with an admission control policy tohandle multiple simultaneous clip downloads. In particular,the admission control policy must strive to do a “good job”of rejecting requests for download that are unlikely to besatisfied within a reasonable amount of time. Section IIIformalizes and quantifies this qualitative notion by providingsome utility models for QoS that take into account the numberof rejected requests, and the number of admitted requests thatare successful and the number of admitted requests that failto meet a specified startup latency constraint.

Compared to traditional admission control in static en-vironments, there are some unique challenges in the C2P2MANET environment where we have a variable topologynetwork. For one, the admission control must be performed ina distributed manner as there is no central coordination pointfor the network. Further, it is hard to estimate the resourcesthat will be available for a clip download (since the topologywill change during the duration of the download).

The key contribution of this paper is the design and analysisof such simple admission control policies using C2P2/MANETsimulation studies. Even though these techniques are sim-ple they capture significant components that would be partof making an appropriate admission control decision. Thepolicies consider components such as number of availableservers, the available bandwidth at each server, the averagetime for which the server is in the range of the client and thebottleneck bandwidth along the path (traditional approach usedin wired networks). A more powerful and efficient admissioncontrol scheme can be designed by appropriately weighingeach of these components and then combining them. Thisstudy serves as a first step toward this goal by evaluatingpolicies that consider each of the components individually tomake the admission control decision and then observing theperformance in terms of some utility models.

Recent studies addressing QoS in MANETs are as follows.A survey of QoS in MANETs is provided by [14], [11],[17]. These studies either build on top of the existing routingprotocols [18], [12] like DSR,AODV etc or integrate theQoS metrics as part of the routing protocol [10], [16], [19].There have been studies that support both QoS and multicastrouting [1], [13]. Admission control has been a well knownproblem dealt with in wired networks [9]. However, forwireless ad hoc networks not much work has been done. In thispaper we have used the conventional definition of admissioncontrol with simple policies to study the number of rejected,satisfied and satisfied requests in a MANET (C2P2 network).

The rest of this paper is organized as follows.The im-portance of data placement and retrieval scheduling policiesfor single client display scenarios is demonstrated in theAppendix. Next, Section II formulates the admission controlproblem, and presents a general framework and several specificpolicies to address this challenge. Section III evaluates thesepolicies using a simulation study on the basis of utility-based QoS models. Our key findings are that the use ofintelligent admission control policies can significantly reduce

the number of admitted unsatisfied requests. Moreover, amobility-prediction based admission policy that we propose,termed MAD, can provide an order of magnitude improvementover a naive approach that admits all requests. Our conclusionsand future research directions are contained in Section IV.

II. ADMISSION CONTROL

Section II-A formalizes admission control as the processof admitting those requests that are able to download a fix-sized file within a specified duration of time. Section II-B develops a framework for the admission control policies.Finally, Section II-C presents four alternative policies for thisframework.

A. Introduction and Problem Statement

While it is possible to overlap the delivery and displayof continuous media clips using buffering, this strategy runsthe risk of incurring hiccups if network resources becomestemporarily unavailable during delivery. In this work, we willfocus our attention on the delivery of clips that will be playedonly after the entire clip is downloaded. We will assume thateach request specifies the file size of the clip to be downloadedand a maximum download time (i.e. startup latency).

We formulate the admission control problem as follows: Arequest to display a clip

�comes with a specification that all

blocks of�

be downloaded to the requesting C2P2 (denotedas C2P2 � ) within ��� time units. Note that ��� (the downloadtime or the startup latency) can be significantly smaller thanthe playback 1 time of the clip

�. For example, an audio

clip�

with a display time of 12 minutes might specify � �as 1 minute. This means that all blocks of this clip must bematerialized in one minute in order for a request referencing�

to be admitted to the system.We shall consider two flavors of admission control: in-

stantaneous and delayed. In instantaneous admission control,the decision to either reject or admit a request must bemade almost immediately after a request is issued. In delayedadmission control, the admission decision must be made in �time units where ������ . An admitted request is termed afailed request if the system fails to materialize all blocks ofclip�

in � � time units.Thus, assuming requests are issued to a system, an

admission control policy yield three classes of requests. First,a fixed number of admitted and rejected requests, termed �and � , respectively. Note � ������ . Second, some of theadmitted requests might fail, denoted as ��� , ������� . And,finally, some of the admitted requests are serviced successfully,denoted as ��� , where �� ���������� . An admission controlpolicy should strive to maximize both � and � � (ideally, to amaximum of ). By maximizing the total number of admittedrequests ( � ), a larger percentage of requests are processed bythe C2P2 MANET. By maximizing � � , a larger percentage ofadmitted requests are processed successfully. In Section III,we will develop QoS models that define a policy’s utility as aweighted combination of these raw metrics.

1The terms playback and display are used interchangeably in this paper

B. An Admission Control Framework

We now propose a framework for solving the admissioncontrol problem assuming an ad-hoc network (e.g. no central-ized base stations in a cellular wireless network). We focuson distributed, client-centric, admission control policies thatenable the C2P2 client device to either admit or reject arequest. For the rest of this paper, C2P2 � denotes a devicewhose user has requested the display of a title.

In order to make an intelligent admission decision, theadmission control component of C2P2 � must first obtaininformation about the state of the network and availabilityof resources. This information is then weighed against thespecifications of the request in order to determine if the requestcan be admitted. The admission control policy may considerseveral factors such as the number of servers, the availabilityof residual bandwidth at servers, the bottleneck bandwidth onthe network path to the server, and whether it is possible topredict from the mobility pattern of the cars that they might beable to collaborate to deliver the requested clip. In this paper,we design and evaluate policies that consider each of the abovementioned factor individually. A meta-policy may combinethese policies by assigning each an appropriate weight.

In a mobile ad-hoc network such as C2P2, knowledgeof the current bandwidth on the path between a client andserver is insufficient to determine if a given request can besatisfied. This is because the available bandwidth changesdynamically and might be either greater or lower in thefuture (due to unpredictable mobility-induced changes in thenetwork structure and network traffic). In order to contend withthe dynamic nature of a C2P2 network, flexible but simpleadmission control policies must be designed that considerdifferent kinds of information about the current state of thenetwork, and try to predict/estimate based on this informationwhether the required resources for a given request can beprovided. In our admission control framework, we develop twokinds of metrics: a Request Metric ( ��� ), and an InformationMetric (

� � ). The request metric ��� (which is normalized tobe a number between 0 and 100) is a measure of the amount ofresources required for a given request. Hence it depends on theQoS parameter of interest. In particular, we define � � as theratio of the requested bandwidth over the nominal maximumwireless link bandwidth � ����� as follows:

�� ��� ���� ������� �����

������� (1)

where ���� "!$# �&% is the size of the clip in Mb (MB), and� ����� is also defined in Mbps (MBps).

While ��� is not policy-dependent, the information metric� � is, and represents quantitatively the information obtainedby the client about the current and future resources availablein the network. The information metric is normalized to be anumber between 0 and 100.

� � is monotonic in the estimatedresource availability — the higher the value of

� � , the morelikely it should be that a given request can be satisfied. Both� � and ��� have no units. Now the admission control decision

is simply a matter of comparing the two metrics:� � and

��� , and deciding whether the given� � predicts sufficient

availability of resources for the request with metric �'� . Asimple yet flexible form for the admission control policy issimply to take the difference between

� � and ��� and comparewith a threshold ( . When

� �*) ���,+-( , a policy admits thecorresponding request. Otherwise the request is rejected.

Note that based on our definitions, the threshold can po-tentially take on values from ) ����� (which corresponds toallowing all requests) to ����� (which corresponds to rejectingall requests). Thus if ( is too low, there is a danger of greaterunsatisfied requests due to bandwidth contention. On the otherhand, if it is too high, the number of satisfied requests willbe low because too few requests are accepted. The optimalchoice of threshold may be an intermediate value and couldbe scenario-dependent as we shall see in our experiments.Of course, the optimal choice of threshold is also policy-dependent, since

� � is defined differently for each policy.When a request is issued by a client, in instantaneous

mode, a policy must either accept or reject requests usingthe threshold computation. With the delayed approach, theadmission control policy may analyze the rate of data flowfor � time units, where � � � � . A request is admitted if:(1) a policy admits this request initially, and (2) the averageobserved bandwidth during �/. ��� ���� ���� � . This approach isadvantageous because it recognizes the dynamic environmentalchanges may either affect delivery of

�in a positive or a

negative manner. A positive change is one where a serverbecomes network reachable due to mobility. The reverse ofthis is a negative change where a server is no longer networkreachable because of mobility. Obviously, as � approaches��� , the admission control is provided with more time to makea decision, reducing the number of admitted requests that fail(minimizing ��� ). However a large � may be quite undesirablefor several reasons. First, it might delay a user longer thannecessary prior to rejecting this user. Second, it might wastenetwork bandwidth by allocating resources to a request thatcannot be satisfied.

In the above scenario, a rejected request based on an initialinstantaneous decision might also be delayed � time unitsto enable the system to see if changes in the environmentduring � time units may facilitate admission of this request.In this case, the client tries to service the rejected requestand observes the amount of available bandwidth during �time units. Assume 0 units of data arrive during this period,� �-0 � size(X). After � time units, it invokes the admissioncontrol policy once again with size(X) )10 . If the request isadmitted then it proceeds to service this request. Otherwise, therequest is rejected. This approach suffers from the followinglimitation. When bandwidth is scarce, a request that should berejected might compete for the available bandwidth with otheractive displays, causing admitted requests to fail (reducing� � ). We do not evaluate this approach in our simulations.

Content replication and switched proximate server selectionare important for robust delivery of high-rate content (seeAppendix). In all our admission control policies, the client

always chooses the closest server from � candidate servers.When C2P2 � ’s path to a server breaks at time ��� , C2P2 �might have downloaded 0 bytes of a clip

�, � � 0 �

� �� !$# �&% . The bandwidth required to download the remainderof a clip is

�� ���� � ������ � � ��� . If this bandwidth exceeds the wirelessnetwork bandwidth, the request is discarded as a failed request.Otherwise, C2P2 � identifies another reachable server andbegins downloading the remainder of the clip from that server.

C. Admission Control Policies

This section details four policies for our proposed frame-work:

1. Server Count (SC): The server count-based admissioncontrol policy counts the number of C2P2 devices that arewithin � hops of C2P2 � , denoted as

. Next, it identifies

and counts the number of devices that contain the referencedclip X, denoted by � . The

� � value for SC is the density ofthese servers: � � � � ���"� (2)

The main strength of this policy is its simplicity — it mightbe implemented using the following probe based approach.C2P2 � sends a controlled flood within � hops. When a nodereceives a probe, it replies to C2P2 � with its identity andwhether it contains clip

�. Based on these replies, C2P2 �

evaluates

and � . C2P2 � may wait for � time units priorto rendering a decision. The value of � might be dictated bythe duration of time required for 2 � network transmissions:� transmissions to reach nodes that are � hops alwaysand, another � transmissions for their reply to arrive atC2P2 � . This policy suffers from several weaknesses. First, ininstantaneous mode, this policy ignores bandwidth limitationsand fails to detect scenarios where all � severs might bebusy processing other requests. In delayed mode, this policy isprovided with some measure of available bandwidth. Second,this policy ignores mobility where the � servers might bemoving away from C2P2 � . Once again, delayed mode ofadmission control provides this policy with some measure ofthis parameter.

2. Server Bandwidth (SB): This policy considers availablebandwidth of each server, requiring each probe to return theavailable bandwidth of each server � , ��� # � � % , containing thereferenced clip. Note that the server of choice for C2P2 � maychange over the duration of the clip download ( � � ). Sinceeach server � � is a potential candidate and has some probabilityof being chosen over �� the

� � metric considers the averageavailable bandwidth across all the servers � � . SB defines

� �as:

� � ������� ��� # � � % ����� �

� ����� (3)

where � ����� is the maximum network bandwidth availableto a C2P2 device. By using

in the denominator, this policy

considers the fraction of nodes that may act as servers. Similarto SC, SB is also relatively simple to implement using a localprobe.

3. Path Bandwidth (PB): This policy is similar to that usedon traditional admission control and bandwidth reservation inwired networks, and considers the bottleneck bandwidth onthe path between the client and each server within � hops.This policy only considers the average of all the instantaneousbottleneck bandwidth along the shortest path from the clientto each server ( � � ). These estimates will rapidly become staleas the nodes move and the paths between the C2P2 � and theservers change. Formally, this policy considers the availablebandwidth of most utilized node, BW( ����� �� ), that lies in themulti-hop path between each server and the client C2P2 � :

� � ������ ��� # � ��� �� % �������

�����"� (4)

To implement this policy, a probe accumulates the availablebandwidth of nodes on its path back from a server to C2P2 � .C2P2 � identifies the intermediate node with least availablebandwidth as � ��� �� . This policy does not consider dynamicevolution of the network where � ��� �� changes over time. Instatic scenarios, PB improves upon SB because it considersscenarios where the servers have sufficient bandwidth but theintermediate hops are data-starved. However as we shall see,this is not necessarily a good strategy in MANETs.

Note that in all the policies SC, SB and PB the probesconstitute the control traffic which is negligible in comparisonto the amount of data that is being exchanged between theclient and the servers.

4. Mobility Prediction (MAD): The mobility-based ad-mission control (MAD) policy tries to leverage the mobilitypattern in the network by estimating the duration of time aserver � � will be in the radio range of the requesting clientC2P2 � . This depends on both the direction and speed of � �and C2P2 � , C2P2 � sends out a probe message with a certaintime to live to obtain the list of proximate servers � � . Assuming� # � �������! "� � % denotes the duration of time � � and C2P2 � areexpected to be in radio range, for MAD:

� � ��#��� � # � � �����! "� � % � �

� ���"� (5)

Here,

is the total number of nodes reachable fromthe client via the probe. Note that both the numerator anddenominator are in units of time.

This policy only considers the time that a server is withinradio range of C2P2 � . This is conservative because it does notconsider the time the servers are reachable via multiple hops.Also this policy does not consider the available bandwidthat each server. We considered a hybrid policy that combinesMAD with SB that takes into account both the mobility of theservers and the available bandwidth at each server. However,we have not presented the results for such a policy here.

III. PERFORMANCE EVALUATION

In this section, we quantify the performance trade-off asso-ciated with the alternative policies. We conducted many exper-iments with different parameter settings. In order to summarize

Value of a Value of a Value of aModels rejected request, successfully admitted failed admitted��� (M) request, ����� (M) request, ����� (M)

Economy 0 1 0Standard 0 1 -1Premium -0.5 1 -5

TABLE I

THREE UTILITY MODELS TO QUANTIFY QUALITY OF SERVICE WITH ALTERNATIVE POLICIES.

the lessons learnt, we developed a utility models to summarizethe QoS observed with each policy in one number. Thesemodels are presented in Section III-A. Section III-B presentsour experimental environment that considers both stationaryand mobile servers. Section III-C demonstrates superiority ofMAD to an environment without admission control. Finally,Section III-D demonstrates that MAD typically outperformsSC, SB, and PB. This section also includes an in-depthanalysis of MAD with different thresholds and system loads.

A. Utility models for QoS

This section describes three utility models used to quantifythe QoS provided by the different admission control policies.

Recall that an admission control policy divides the totalnumber of requests into three groups: � rejected requests,��� requests that are admitted but fail (due to unsatisfiedrequirements), and ��� requests that are admitted and succeedin providing the clip within the specified deadline. Differentusers may value each of these differently, as per their utilitymodel. For the purpose of evaluating our experimental resultswe consider three distinct utility models that are representativeof different levels of Quality of Service that can be providedin a C2P2 ad hoc network. In each model

�, we define the

joint utility as a weighted sum:

� ����# � % ������� � # � % � � ����� � # � % � � (6)

where a weight � � # � %indicates the value of component � in

model�

.Table I shows the weights for three QoS models employed

to evaluate the experimental results. The economy model doesnot penalize a policy for either rejected or failed requestsand only rewards accepted requests. The Standard model isindifferent to rejections, but penalizes failed admitted requestsas much as it rewards successfully served admitted requests.The Premium model resembles a high QoS environment. Itpenalizes rejected requests (albeit slightly), rewards admittedrequests that are satisfied, and highly penalizes admittedrequests that fail. The high penalty for a failed admitted requestreflects our intuition that users are likely to be annoyed iftheir C2P2 device initiates a download, spends a considerableamount of time (maybe �� time units) only to discover that itcannot display a clip because an incomplete file is downloaded.

While these utility models can be applied on a per-userbasis and different QoS levels can be integrated within thesame system in practice, for the purposes of analyzing ourexperimental result and comparing the different admission

control policies, we shall apply each QoS utility model tothe system as a whole for the duration of the experiment.

B. Experimental Setup

The results presented in this section are based on a newsimulator for C2P2 networks written using C � programminglanguage2. The simulator models road stretches and cars thatnavigate these stretches. A car might be configured witha C2P2 device that provides a fixed amount of networkbandwidth. Each C2P2 device implements the DSR [8] routingpolicy. However, we believe that a the choice of the routingprotocol does not impact the trends in the observed results. Aproactive protocol like DSDV would have shown similar re-sults. We analyzed different scenarios consisting of a differentnumber of road-stretches, different number of C2P2-equippedcars, mobility patterns, and request specifications.

Our basic experiment consists of thirteen bi-directionalroad-stretches, numbered from 0 to 12, see Figure 1. Threestationary servers (these could be parked vehicles, accesspoints, or base stations) are located on the following road-stretches: 1, 6, and 11, numbered ��� , � � , and ��� , respectively.The radio-range of each C2P2 spans 3 road-stretches: itscurrent road stretch and its two adjacent road stretches. Thus,a C2P2 on either road-stretch 0 or 2 is in the radio range of� � which is on road-stretch 1. Note that there exists four roadstretches that are dark because they are not 1-hop reachableby a server. However, given sufficient number of clients andtransitive routing of packets across these clients, these darkroad stretches may lit-up as they become network reachableto either one or two servers. We analyzed configurationsconsisting of 13 possible client C2P2 nodes (in addition tothe 3 server nodes), with the simultaneous active displaysranging from 1 to 13 in our experiments. Initially, all carsare assigned to the road-stretches such that they are evenlyspaced across the road-stretches. For example, with 13 cars,one car is assigned to each road-stretch. The initial directionof each car is chosen randomly (i.e. to move left or right).The speed of each car is fixed at 5 meters per second. Once acar reaches the end of either road stretch 0 or 12, it switches

2We initially considered using the ns-2 simulator that is widely used inthe MANET community for these simulation experiments, but found thatsignificant programming extensions to ns-2 would be required to implementthe kinds of online per-client dynamic admission-control policies we areinterested in evaluating. The simulator we have written is designed toincorporate such policies with ease and provides for much faster simulations.We plan to make the simulator available in the near future for use by othersinvestigating admission control.

0 1 2 3 4 5 6 7 8 9 10 11 12

s1s0 s2

Fig. 1. Thirteen road stretches with 3 stationary servers.

directions and moves toward the opposite end. Within eachconfiguration, the cars that are requesting clips (correspondingto the number of active displays) are chosen randomly andthey re-issue requests periodically every 60 seconds. A totalof 100 requests are issued in each experiment. This meansthat with 10 active displays, each client issues 10 requests onaverage. All three servers are assumed to contain a replica ofthe clips being requested. Rejected requests do not consumeany network bandwidth and are terminated immediately. Oncea request is admitted to the system, it downloads a clip for 60seconds. The maximum network bandwidth of a C2P2 deviceis assumed to be 10 Mbps. The referenced clip is a media clipwith a bandwidth requirement of 340 Kbps. The display timeof each clip is fixed at 12 minutes. We require the clip (size30MB) to be downloaded in 60 seconds, requiring a downloadbandwidth requirement of 4 Mbps.

To analyze the impact of server mobility, we modify thisbasic experimental setup to include scenarios where all serversare also mobile. We also analyzed the impact of load in twoways: first, by varying the number of simultaneous clients issu-ing requests, termed active displays. Second, by changing thedownload bandwidth requirement to 8 Mbps, corresponding toa clip size 60MB and a media playback time of 24 minutesat 340 Kbps rate.

C. A Case for Admission Control

We start with experimental results that justify the use ofintelligent admission control policies in an ad-hoc network ofC2P2 devices. Figure 2 shows utility of MAD (Instantaneousmode of operation) with alternative models when comparedwith an environment that does not utilize an admission controlpolicy. Figurer 3 shows the number of rejected, satisfied,and unsatisfied requests for both MAD and the no admissioncontrol case. (As we shall show in Section III-D, MADis generally a superior admission control policy.) Withoutadmission control, the simulator admits a request upon itsarrival independent of server availability. Figure 2 shows theperformance of MAD for two different thresholds: -20 and -40. The x-axis of this figure shows the number of concurrentclients that initiate the display of a clip, termed number ofactive displays. This controls the load in the environment andis increased from 1 to 13. The y-axis of this figure is theutility of each model. We present results for all models shownin Table I. Each presented data point is an average of resultsobtained from 10 experiments utilizing different seeds. (Therandom seed impacts the order and identity of clients issuingrequests.)

With 3 stationary servers and a maximum wireless band-width of 10 Mbps per C2P2 device, our experimental environ-ment supports a total bandwidth of 30 Mbps. All active clientsissue their requests at the same time. A total of 100 requestsare issued by all clients in this environment. This implies thatwith two or fewer number of active displays, if there wereeither no dark regions or unpredictability due to mobility,without admission control, a total of 100 requests would besatisfied. Figure 3.b shows that only one half of requests areserved successfully with 2 or fewer active displays. This isbecause of mobility and dark regions that cause an active clientto starve for data, failing to download a clip

�in � � time

units. This is reflected in the number of unsatisfied requestsadmitted to the system, see Figure 3.c.

A primary observation from Figures 2 and 3 is that anenvironment with no admission control is clearly inferior toInstantaneous-MAD with ( =-40. With the Premium model,MAD shows a much better performance compared to the otherpolicies. With the Standard utility model, MAD remains supe-rior. With the Economy model, however, no admission controloutperforms MAD when ( =-20. This threshold renders MADtoo conservative, forcing it to admit too few requests. Whileall these requests are processed successfully, see Figure 3.cwhere ��� =0 with ( =-20, MAD does reject requests that canbe processes successfully. With a more relaxed threshold (-40),these requests are admitted into the system, enabling MAD tooutperform the environment with no-admission control whenusing the Economy model.

Note that utility of MAD with ( =-20 is a constant positivewith the Economy, Standard, and Premium models. When ( =-40, MAD’s utility drops as a function of load because thenumber of unsatisfied requests increases with this threshold,see Figure 3.c.

D. Detailed Experimental Results

Tables II, III, and IV show utility of each model in bothInstantaneous and Delayed mode. Each table summarizesresults for a specific model: Economy, Standard, and Premium.A table reports the maximum observed utility for each policyand the corresponding threshold at which this maximum isrealized. We employ a set notation to show threshold valuesbecause the same maximum might have been observed forseveral ( values.

Two different system loads are presented: First, a lightsystem load where 2 random clients issue requests at a time.Even if both reference different clips from the same server,the available network bandwidth (10 Mbps) accommodatesboth requests, allocating a total of 8 Mbps. Second, a high

2 Active Displays 10 Active Displays

Instantaneous Delayed, � =15 Sec Instantaneous Delayed, � =15 Sec

Max Threshold Max Threshold Max Threshold Max Thresholdutility � utility � utility � utility �

Admit-All 43.00 N/A 42.00 N/A 22.00 N/A 18.00 N/AReject-All 0 N/A 0 N/A 0 N/A 0 N/A

SC 45.00 �������� 43.00 �������� 30.00 ������� 22.00 �������SB 45.00 �������� 43.00 �������� 27.00 �������� 19.00 ��������PB 45.00 �������� 43.00 �������� 27.00 �������� 19.00 ��������

MAD 45.00 �������� 43.00 �������� 29.00 �������� 22.00 �������TABLE II

A COMPARSION OF ALTERNATIVE ADMISSION CONTROL POLICIES WITH THE ECONOMY MODEL, ������������� � = 4 MBPS, AND 16 C2P2 DEVICES: 3

STATIONARY SERVERS AND 13 MOBILE CANDIDATE CLIENTS.

2 Active Displays 10 Active Displays

Instantaneous Delayed, � =15 Sec Instantaneous Delayed, � =15 Sec

Max Threshold Max Threshold Max Threshold Max Thresholdutility � utility � utility � utility �

Admit-All -14.00 N/A 32.00 N/A -56.00 N/A 13.00 N/AReject-All 0 N/A 0 N/A 0 N/A 0 N/A

SC 11.00 ������� 32.00 � -101,-80,-60,-40 1.00 ���� 18.00 �������SB 13.00 ������� 32.00 � -101,-80,-60,-40 1.00 ���� 13.00 � -101,-80,-60 PB 12.00 ������� 32.00 � -101,-80,-60,-40 1.00 ���� 13.00 � -101,-80,-60

MAD 18.00 ������� 33.00 �������� 21.00 ������� 21.00 �������TABLE III

A COMPARSION OF ALTERNATIVE ADMISSION CONTROL POLICIES WITH THE STANDARD MODEL, ������������� � = 4 MBPS, AND 16 C2P2 DEVICES: 3

STATIONARY SERVERS AND 13 MOBILE CANDIDATE CLIENTS.

2 Active Displays 10 Active Displays

Instantaneous Delayed, � =15 Sec Instantaneous Delayed, � =15 Sec

Max Threshold Max Threshold Max Threshold Max Thresholdutility � utility � utility � utility �

Admit-All -242.00 N/A -12.80 N/A -368.00 N/A -14.70 N/AReject-All -10 N/A -10 N/A -10 N/A -10 N/A

SC -8.90 ���� -8.90 ���� -8.90 ���� -5.40 �������SB -8.90 ���� -8.90 ���� -8.90 ���� -8.90 ����PB -8.90 ���� -8.90 ���� -8.90 ���� -8.90 ����

MAD 9.80 ������� 9.80 ������� 9.30 ������� 9.30 �������TABLE IV

A COMPARSION OF ALTERNATIVE ADMISSION CONTROL POLICIES WITH THE PREMIUM MODEL, ������������� � = 4 MBPS, AND 16 C2P2 DEVICES: 3

STATIONARY SERVERS AND 13 MOBILE CANDIDATE CLIENTS.

system load where 10 random clients issue requests simul-taneously, requiring 40 Mbps and exhausting the availablebandwidth. (With 3 servers, a maximum bandwidth of 30Mbps is supported by the system assuming an even distributionof active displays across the clients.) With each load, a totalof 100 requests are issued to the system. The utility shownwith the Economy model, see Table II, is the number ofsatisfied requests. A low system load results in a highernumber of satisfied requests with all policies when comparedwith a high load. This value is maximized by increasing thenumber of admitted requests, explaining why Instantaneousmode outperforms the Delayed mode. With the Standard utilitymodel assigning a -1 to each failed admitted request, Delayedmode of operation becomes superior to Instantaneous, seeTable III. This holds true across all policies. Note that SB and

PB are no different than Admit-all in this case. This is reflectedby the fact that a threshold of -101 (admits all requests) is oneof the thresholds that maximizes utility of SB and PB with theStandard model.

With the Premium model assigning a -5 to each failedadmitted requests and a -0.1 to each rejected request, MADoutperforms all other policies, see Table IV. Note that theDelayed mode continues to improve an Admit-all policy.Moreover, a policy that rejects all requests provides a com-parable utility to SC, SB, and PB.

Tables II, III, and IV show the superiority of MAD with thePremium and Standard models. With these models, a Delayedmode of operation that samples available bandwidth for 15seconds further enhances all policies (including the naiveAdmit-all policy). In the rest of this section, we analyze thecharacteristics of MAD as a function of different thresholds

1 2 3 4 5 6 7 8 9 10 11 12 13−400

−350

−300

−250

−200

−150

−100

−50

0

50

Number of Active Displays

Premium Utitlity

MAD, Threshold = −20

MAD, Threshold = −40

No Admission Control

2.a) Premium

1 2 3 4 5 6 7 8 9 10 11 12 13−70

−60

−50

−40

−30

−20

−10

0

10

20

30

Number of Active Displays

Standard Utitlity

MAD, Threshold = −20

MAD, Threshold = −40

No Admission Control

2.b) Standard

1 2 3 4 5 6 7 8 9 10 11 12 130

10

20

30

40

50

Number of Active Displays

Economy Utitlity

MAD, Threshold = −20

No Admission Control

MAD, Threshold = −40

2.c) Economy

Fig. 2. A comparison of Instantaneous-MAD with an environment thatemploys no admission control. The environment consists of 16 C2P2 devices:3 stationary servers and 13 mobile clients.

with mobile servers, Instantaneous versus Delayed mode ofoperation, and different loads (both number of active displaysand bandwidth required during �� ).

Figure 4 shows the utility of MAD with alternative modelsfor both stationary and mobile servers as a function of alter-native threshold ( ( ) settings. This figure shows the following.First, the utility of all the models improves when servers aremobile. Second, the utility of all models becomes a fixedconstant when ( + � � . Third, the observed trends whencomparing the Delayed mode with the Instantaneous mode ina mobile server environment are similar to the discussion ofstationary servers. We now describe the first two observationsin turn. (See the discussion of Tables II, III, and IV for anexplanation of the third observation.) To explain the first, notethat there are no permanent dark regions with mobile servers.

1 2 3 4 5 6 7 8 9 10 11 12 130

10

20

30

40

50

60

70

80

90

100

Number of Active Displays

Number of Rejected Requests

MAD, Threshlod = −20

MAD, Threshlod = −40

No Admission Control

3.a) Rejected

1 2 3 4 5 6 7 8 9 10 11 12 130

10

20

30

40

50

60

70

80

90

100

Number of Active Displays

Number of Satisfied Requests

MAD, Threshold = −40

No Admission Control

MAD, Threshold = −20

3.b) Successful Admit

1 2 3 4 5 6 7 8 9 10 11 12 130

10

20

30

40

50

60

70

80

90

100

Number of Active Displays

Number of Unsatisfied Requests

No Admission Control

MAD, Threshold = −40

MAD, Threshold = −20

3.c) Failed Admit

Fig. 3. Rejected ( � ), successfully admitted ( ��� ), and failed admitted ( ��� )requests with different number of active displays and 16 C2P2 devices: 3stationary servers, and 13 mobile clients.

With permanent dark regions (stationary servers), an activelydisplaying client might travel in one of these regions whereno bandwidth is available for a sustained period of time. Uponexiting such an area, this client attempts to download theremaining portion of a clip at a higher rate. When this rateexceeds 10 Mbps (maximum network bandwidth), this requestis discarded as a failed request because it is impossible todownload the remaining portion of the clip in the remainingtime. With transient dark regions as is the case with mobileservers, the likelihood of such scenarios is minimized.

In Figure 4 (and Figure 5), MAD rejects all requests with(-+ � � . This causes the Standard and Economy models toconverge to zero, while the Premium model converges to -10 because each rejected request has a value of -0.1 (seeTable I) and there are 100 requests. This explains the second

−101 −80 −60 −40 −20 0 20 40 60 80 100−60

−50

−40

−30

−20

−10

0

10

20

30

40

50

THRESHOLD

UT

ILIT

Y

Premium StationaryStandard StationaryEconomy StationaryPremium MobileStandard MobileEconomy Mobile

4.a) Instantaneous

−101 −80 −60 −40 −20 0 20 40 60 80 100−60

−50

−40

−30

−20

−10

0

10

20

30

40

50

THRESHOLD

UT

ILIT

Y

Premium StationaryStandard StationaryEconomy StationaryPremium MobileStandard MobileEconomy Mobile

4.b) Delayed

Fig. 4. Analysis of Instantaneous and Delayed MAD with mobile andstationary servers. The environment consists of 16 C2P2 devices: 3 servers,and 13 mobile clients. With mobile, servers and clients move at the samespeed. With stationary, servers are parked in road stretches 1, 6, and 11.

observation.Figure 5 compares Delayed MAD with different system

loads. While Figure 5.a shows two different number of simul-taneous displays (2 and 10), Figure 5.b shows two differentbandwidth requirements during � � (4 and 8 Mbps). Bothfigures show a reduced utility for all models with a highersystem load. Moreover, the threshold that maximizes the utilityof a model changes for a given system load. For example,the maximum utility with the Premium model is observedat a different threshold; at -60 with 8 Mbps and -20 with 4Mbps, See Figure 5.b. This is also observed with the Standardand Economy models. This again emphasizes that for a givenutility model, each policy has an ideal threshold ( � � � ��� whichprovides the best performance.

IV. CONCLUSIONS AND FUTURE RESEARCH

DIRECTIONS

The primary contribution of this paper is a mobility basedadmission control policy (MAD) for delivery of continuous

−101 −80 −60 −40 −20 0 20 40 60 80 100−60

−50

−40

−30

−20

−10

0

10

20

30

40

50

THRESHOLD

UT

ILIT

Y

Premium With 2 Active DisplaysStandard With 2 Active DisplaysEconomy With 2 Active DisplaysPremium With 10 Active DisplaysStandard With 10 Active DisplaysEconomy With 10 Active Displays

5.a) 4 Mbps Media Type

−101 −80 −60 −40 −20 0 20 40 60 80 100−60

−50

−40

−30

−20

−10

0

10

20

30

40

50

THRESHOLD

UT

ILIT

Y

Premium With 4MbpS MediaStandard With 4MbpS MediaEconomy With 4MbpS MediaPremium With 8MbpS MediaStandard With 8MbpS MediaEconomy With 8MbpS Media

5.b) 2 Active Displays

Fig. 5. A comparison of Delayed MAD ( � = 15 Sec) with different loadparameters and 16 C2P2 devices: 3 mobile servers, and 13 mobile clients.

media in an ad-hoc network of C2P2 devices. Our experimen-tal results demonstrate the following. First, with continuousmedia, an environment that employs admission control issuperior to one without admission control. Second, MADoutperforms traditional admission control policies (for wirednetworks) when the environment is penalized for failing toservice admitted requests. Third, with a delayed approach,under heavy load conditions, the performance of all policiesincluding the naive admit-all approach improves considerably.

We are currently extending this study in several ways. First,we are developing techniques that enable our framework todetermine the appropriate threshold value ( ( ) for use withMAD. Based on the experimental results of Section III-D, these techniques estimate system load and the expectedamount of work imposed by a display to determine ( . Second,we are investigating the use of a deadline driven data deliverytechnique as an extension of DSR to further maximize thenumber of admitted requests that are satisfied with MAD. Thediscussions of Section III assumed the data is delivered from aserver to a client at a rate of

��� ���� ���� � . With the deadline driven

approach, each block is tagged with a lifetime. This enables aserver to push data to a client at a faster rate than

��� � � ���� � whenthe ad-hoc network is idle. Moreover, deadlines might expeditedelivery of those blocks that are needed more urgently. Finally,we are investigating the design and evaluation of a meta-policythat combines SC, SB, and PB with MAD. This policy assignsa weight to each component in order to consider both availablebandwidth and mobility.

V. ACKNOWLEDGMENTS

This research was supported in part by NSF grant IIS-0307908, National Library of Medicine LM07061-01, and anunrestricted cash gift from Microsoft Research.

REFERENCES

[1] B. Bellur, R. Ogier, and F. Templin. Topology broadcast based onreverse-path forwarding. Internet-Draft Version 01, IETF, March 2001.Work in progress.

[2] E. Cohen and S. Shenker. Replication Strategies in Unstructured Peer-to-Peer Networks. In Proceedings of the ACM SIGCOMM, August 2002.

[3] S. Ghandeharizadeh and T. Helmi. An Evaluation of Alternative Contin-uous Media Replication Techniques in Wireless Peer-to-Peer Networks.In Third International ACM Workshop on Data Engineering for Wire-less and Mobile Access (MobiDE, in conjunction with MobiCom’03),September 2003.

[4] S. Ghandeharizadeh, T. Helmi, S. Kapadia, and B. Krishnamachari. ACase for a Mobility based Admission Control Policy. In In Proceedingsof the 10th International Conference on Distributed Multimedia Systems,September 2004.

[5] S. Ghandeharizadeh, B. Krishnamachari, and S. Song. Placement ofContinuous Media in Wireless Peer-to-Peer Networks. IEEE Transac-tions on Multimedia, April 2004.

[6] Shahram Ghandeharizadeh, Shyam Kapadia, and Bhaskar Krishna-machari. PAVAN: a policy framework for content availabilty in vehicularad-hoc networks. In VANET ’04: Proceedings of the first ACM workshopon Vehicular ad hoc networks, pages 57–65. ACM Press, 2004.

[7] Shahram Ghandeharizadeh, Shyam Kapadia, and Bhaskar Krishna-machari. Comparison of Replication Strategies for Content Availabilityin C2P2 Networks. In To Appear in MDM ’05: Proceedings of the 6thInternational Conference on Mobile Data Management. ACM Press,2005.

[8] D. B Johnson and D. A Maltz. Dynamic source routing in ad hocwireless networks. In Imielinski and Korth, editors, Mobile Computing,volume 353. Kluwer Academic Publishers, 1996.

[9] E. Knightly and N. Shroff. Admission control for statistical qos: Theoryand practice. IEEE Network, vol. 13, no. 2, pp. 20-29, 1999.

[10] Seoung-Bum Lee, Gahng-Seop Ahn, Xiaowei Zhang, and Andrew T.Campbell. INSIGNIA: An IP-based quality of service framework formobile ad hoc networks. Journal of Parallel and Distributed Computing,60(4):374–406, 2000.

[11] P. Mohapatra, J. Li, and C. Gui. Qos in mobile ad hoc networks. IEEEWireless Communications, pp. 44-52, June 2003.

[12] Elena Pagani and Gian Paolo Rossi. A framework for the admission con-trol of qos multicast traffic in mobile ad hoc networks. In Proceedingsof the 4th ACM international workshop on Wireless mobile multimedia,pages 2–11. ACM Press, 2001.

[13] C. Perkins, E. Royer, and S. Das. Ad hoc on demand distance vector(aodv) routing. Internet-Draft Version 07, IETF, November 2000. Workin progress.

[14] D. Perkins and H. Hughes. A survey on quality of service supportin wireless ad hoc networks. Journal of Wireless Communicationsand Mobile Computing (WCMC), Speical Issue on Mobile Ad HocNetworking Research, Trends and Application, 2(5) pp. 503-513, 2002.

[15] J. D. Salehi, Z. Zhang, J. F. Kurose, and D. Towsley. SupportingStored Video: Reducing Rate Variability and End-to-End ResourceRequirements through Optimal Smoothing. In Proceedings of the 1996ACM Sigmetrics Conference, May 1996.

[16] Prasun Sinha, Raghupathy Sivakumar, and Vaduvur Bharghavan.CEDAR: a core-extraction distributed ad hoc routing algorithm. InINFOCOM (1), pages 202–209, 1999.

[17] K. Wu and J. Harms. Qos support in mobile ad hoc networks. CrossingBoundaries - an interdisciplinary journal, Vol 1, No 1, 2001.

[18] H. Xiao, W. Seah, A. Lo, and K. Chua. A flexible quality of servicemodel for mobile ad-hoc networks. In IEEE VTC, 2002.

[19] C. Zhu and M. Corson. Qos routing for mobile ad hoc networks. InINFOCOM, June 2001.

APPENDIX

Before we move on to address the challenges of admissioncontrol for multiple simultaneous displays, we first need tounderstand the basic data placement and delivery schedulingstrategies that are needed to provide satisfactory performanceeven in a single display scenario. This section focuses on thedisplay of a single clip in a C2P2 network. It shows thatcontent replication and block-level switched retrieval of datafrom proximate servers provides low-latency delivery of highdata-rate content.

The basic principle behind replication and block-switchedretrieval is as follows.

The clip is replicated on � servers, and divided into iden-tifiable segments known as blocks (of size � ). The client isassumed to be able to determine the identity and location ofthese servers through a resource discovery protocol. (See [2]for a description of techniques to identify nodes with relevantdata in an unaddressable wired network.) When the retrievalof a clip is initiated, the client requests the first block fromthe nearest server. As each successive block is retrieved, thisprocess is repeated, with the client always requesting theinformation from the nearest server. There are therefore twokey parameters of interest: the block size � , and the numberof servers � .

Intuitively, as the number of servers is increased and theblock size is decreased, there is a greater diversity of choices(higher likelihood that at least one server is always near theclient in the network) and greater adaptation to mobility (dueto increased frequency of switching), enabling an improvementin the quality of continuous media retrieval. This requiresgreater overhead as the frequency of switching increases.However, this approach depends on the block size and speed ofthe cars which in turn determines the frequency of switching.Note that if the mobility characteristics of the environment leadto frequent changes in the proximate server when a decision isto be made per block then the benefits of choosing the closestserver may become lower the cost incurred due to the addedoverhead of extremely frequent switches. In other cases, whenthe proximate server does not change on a per block basis thebenefits obtained are substantial and exceed the overhead inthe server selection process. (Typically much lesser overheadis incurred in probing to find out if the current server is the”closest” than switching to a new server, and if this switchesare not that frequent then compared to the duration of datatransfer this overhead also becomes negligible.) We validatethis through simulation results but have ignored the switchingoverhead.

Figure 6 shows sample runs of data delivery using block-switched retrieval tested via ns-2 simulation of a 20-nodeMANET with random waypoint mobility.

0 50 100 150 200 2500

2

4

6

8

10

12

14x 10

4 (a) Block size 1MB

Time (s)

Pac

ket S

eque

nce

Num

ber

server switches

Packet Arrival Time

Playback Time

path failures Startup Latency

6.a block size = 1 MB

0 50 100 150 2000

2

4

6

8

10

12

14x 10

4 (b) Block size 1KB

Pac

ket S

eque

nce

Num

ber

Packet Arrival Time

server switches

Playback Time

Time (s)

path failure Startup Latency

6.b block size = 1 KB

Fig. 6. Simulation results of packet arrival times for a samplerun involving block-based staggered non-redundant retrieval from 8servers for two different block sizes; maximum node velocity 30meters/sec; delivery rate 2 Mbps and playback rate 4 Mbps.

Other simulation parameters are: area is 200x200m, link-bandwidth is 10 Mega bits per second (Mbps), clip-displaytime is 120 Seconds, and the underlying routing protocol isDynamic Source Routing (DSR) [8]. The vertical lines inFigure 6.a represent times when there is a server switch.There are two sloping curves. The top one shows the packetarrival times, with some small gaps that correspond to pathbreaks due to mobility. The second curve is a line whichrepresents the rate of data playback as a function of time3.For illustration, we have chosen a playback rate of 4 Mbpsin this figure, while the delivery rate is 2 Mbps. The pointwhere the playback time curve intersects the x-axis represents

3We focus on clips with a Constant Bit Rate (CBR) display requirement.Hiccup-free delivery of Variable Bit Rate (VBR) encoded clips can beconceptualized as a sequence of CBR delivery schedules [15].

0 2 4 6 8 10 120

5

10

15

20

25

30

35

40

45

50

No of servers

Startup latency (seconds)

1KB

1MB

7.a Startup latency

0 2 4 6 8 10 121.2

1.25

1.3

1.35

1.4

1.45

1.5

1.55

1.6x 10

5

No of servers

Total no of packet transmissions

1MB

1KB

7.b Total transmitted packets

Fig. 7. Startup latency and total bandwidth usage (in terms ofpackets � hops) for hiccup-free delivery as a function of the number ofservers for low and high block sizes (max. velocity 30 m/s, deliveryand playback rate 4 Mbps) for a single-client scenario.

the minimum startup latency for hiccup-free display. Whenthere are greater numbers of path failures, the startup latencydeteriorates. When a smaller block size (such as 1KB) is used,the client refreshes its server connection frequently and thereare fewer intra-block path failures in this scenario and as aresult the corresponding startup latency can be reduced4.

Figure 7 shows how the startup latency and the total in-network bandwidth usage varies as a function of the numberof servers for different block sizes.

The first observation is that the startup latency is reducedsignificantly as the number of servers (i.e. the degree of

4Naturally, choosing a smaller block-size also increases the number ofroute setup messages required by the underlying DSR protocol, but thisadditional overhead is relatively insignificant compared to the large bandwidthapplication data.

replication) is increased. We can see reductions of more than20 times in some cases, by increasing the number of serversfrom 1 to 12. Figure 7 also shows that the startup latency canbe reduced by up to 12 times with a 1 KB block size whencompared with a 1 MB block size. These results show an orderof magnitude QoS improvements in the startup latency metricand significant reduction in bandwidth usage with sufficientreplication and block-switched retrievals.

Having addressed the strategies needed for robust single-display downloads, we now turn to the challenges inherentin the download of multiple simultaneous displays. Given thelimited bandwidth of wireless channels, the high data-ratesfor continuous media delivery, and the dynamics of the C2P2MANET topology, intelligent admission control is a must.