518 ieee transactions on mobile computing, vol. 5, no. 5 ...dsm/pubs/01610594.pdf · information...

Information Collection Services for QoS-AwareMobile Applications

Qi Han, Member, IEEE, and Nalini Venkatasubramanian, Member, IEEE

Abstract—Efficient resource provisioning that allows for cost-effective enforcement of application QoS relies on accurate system state

information. However, maintaining accurate information about available system resources is complex and expensive, especially in

mobile environments where system conditions are highly dynamic. Resource provisioning mechanisms for such dynamic environments

must therefore be able to tolerate imprecision in system state while ensuring adequate QoS to the end-user. In this paper, we address

the information collection problem for QoS-based services in mobile environments. Specifically, we propose a family of information

collection policies that vary in the granularity at which system state information is represented and maintained. We empirically evaluate

the impact of these policies on the performance of diverse resource provisioning strategies. We generally observe that resource

provisioning benefits significantly from the customized information collection mechanisms that take advantage of user mobility

information. Furthermore, our performance results indicate that effective utilization of coarse-grained user mobility information renders

better system performance than using fine-grained user mobility information. Using results from our empirical studies, we derive a set of

rules that supports seamless integration of information collection and resource provisioning mechanisms for mobile environments.

These results have been incorporated into an integrated middleware framework AutoSeC (Automatic Service Composition) to provide

support for dynamic service brokering that ensures effective utilization of system resources over wireless networks.

Index Terms—Mobile environments, distributed applications, multimedia applications, middleware.

�

1 INTRODUCTION

THE proliferation of mobile devices and wireless networkshave increased the scope of applications that will execute

in future pervasive environments. Increasingly, distributedapplications (e.g., mobile multimedia) have diverse Qualityof Service (QoS) requirements (such as transmission rate,delay and delay jitter, etc.) that require guaranteed levels ofresource availability (such as network bandwidth and serverCPU, etc.) from the underlying systems infrastructure.Resource provisioning algorithms, such as QoS based net-work routing and server selection, utilize information aboutcurrent system status to ensure that applications meet theirQoS requirements. The accuracy of system state informationcan play a significant role in the efficacy of the QoSmanagement techniques. For example, resource provisioningalgorithms may select a network path for a flow/connectionbased on current resource availability, reserve the chosenpath, and, subsequently, admit the request. In this process,imprecise system state information can lead to two types offailures. A routing failure may occur when a feasible pathcannot be found for the new connection; a setup failure mayoccur when a seemingly feasible path is selected thatultimately does not have enough resources for the newconnection. Neither failure is desirable; in particular, setupfailures incur extra overhead to reserve resources that maynever be used along the path. Maintaining accurate system/network status information can therefore help make right

decisions, ensure the desired application QoS, and, conse-quently, provide better user experience for distributedapplications.

This paper addresses challenges in collecting andmaintaining accurate information about system state formobile applications. Maintaining accurate informationabout available system resources is complicated by dyna-mically changing network, server and client status, andchanging QoS profiles of applications. Gathering accurateinformation becomes even more challenging in a mobileenvironment due to two main reasons. First, mobile devicesroam between access points that connect them to wirednetworks, and the constant user mobility causes significantvariation in the resource availability [1], [2] on variousnetwork links. Second, hand-held devices typically havehighly limited storage, computing resources. The resourceavailability can be substantially affected by the computationand communication profiles of the applications executingon the device—this implies that capturing device limita-tions and changing device status as a part of the systemimage is necessary.

In this paper, we consider the scenario where mobile andstationary hosts coexist in a mixed wired and infrastruc-ture-based (cellular) wireless networks. In the cellularinfrastructure, the coverage area is divided into a numberof cells with one base station per cell. All the wirelessdevices residing in a cell communicate with other entities inthe distributed system via the base station. The basestations, in turn, are connected to the wire line network.Our work aims at ensuring the desired QoS levels forapplications executing on mobile devices, despite frequentchanges in system state. This is achieved by effectivelymonitoring system status in the wired and wireless net-works and allocating resources accordingly. Fig. 1 shows ahigh level system overview. Satisfying user QoS implies

518 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 5, NO. 5, MAY 2006

. Q. Han is with the Department of Mathematical and Computer Sciences,Colorado School of Mines, Golden, CO 80401. E-mail: [email protected].

. N. Venkatasubramanian is with the Department of Computer Science,University of California, Irvine, CA 92697. E-mail: [email protected].

Manuscript received 8 Dec. 2003; revised 1 Aug. 2004; accepted 17 Jan. 2005;published online 15 Mar. 2006.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-0209-1203.

1536-1233/06/$20.00 � 2006 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

accurate maintenance of system state via frequent monitor-ing; this, in turn, introduces significant network trafficresulting in poor utilization of underlying computation,communication, and storage resources. The challenge isthen to obtain sufficiently accurate state information toreduce the cost of collection while meeting user QoS needs.

1.1 Our Approaches

To address the competing goals of accuracy and cost-effectiveness, we exploit the fact that users are able to toleratesystem state imprecision at a certain level. Our strategy is totailor the accuracy and fidelity of the captured data to ensurethat applications can execute at desired levels of quality whileminimizing resource consumption. The questions that needto be answered are: 1) What information needs to be gathered,2) how frequently the information should be collected, and3) at what granularity the information should be maintained?We aim to improve the performance of QoS-based resourceprovisioning by providing a reasonably accurate picture of theunderlying systems obtained and maintained by the pro-posed information collection services. The relevant informationincludes system status information, such as network para-meters, server and client characteristics, etc., as well asmobility-related information, such as current locations ofmobile clients, power levels of mobile devices, and groupmobility patterns, etc. Collecting mobility information at afine-grained level (i.e., for each individual mobile client)would facilitate adaptive resource provisioning; however, itmay also incur significant overhead. We therefore explore thebenefit of using coarse-grained mobility information (i.e., foreach cell). We use the coarse-grained information to adjustinformation collection parameters, which indirectly im-proves the performance of resource provisioning. Fig. 2illustrates the approaches studied in this paper.

1.2 Contributions of This Paper

Recent studies evaluated the impact of stale link state on theperformance of QoS routing [3]; results indicated that therouting algorithm, network topology, and link-state updatepolicy have significant impact on the probability ofsuccessfully routing new requests, as well as the overheadof distributing network load metrics. Along similar lines,this paper aims to explore the impact of system statusinformation collection policies on the efficacy of QoS-basedresource provisioning for mobile applications. Specifically,this paper makes the following contributions:

. We discuss the issues involved in collecting systemstatus information and present three general collec-tion policies. Furthermore, we explore how systemstatus collection policies can benefit from mobilityrelated information. We show that collecting coarse-grained mobility information introduces significantlyless overhead as compared to collecting fine-grainedmobility information. We then present a family ofstrategies where coarse-grained mobility informationis used to adjust the granularity of the system statusinformation and the degree of dynamics with whichcurrent state information is obtained and updated.

. We empirically evaluate the impact of the informa-tion collection strategies on diverse resource provi-sioning techniques under varying mobility status,network/server conditions, and different applica-tion workloads. Through the extensive performancestudies, we identify the most effective informationcollection strategy for different scenarios and derivea set of service composition rules.

. We propose a prototype middleware framework,AutoSeC (Automatic Service Composition), currentlybeing developed at UC Irvine, that can dynamicallyselect an appropriate combination of informationcollection and resource provisioning policies accord-ing to current system conditions and user require-ments. AutoSeC is designed to support heterogeneousapplication workloads (network or computationintensive) and manage resources in heterogeneousenvironments with both fixed and mobile deviceswhile introducing minimal management overhead.We also present the preliminary results and discussthe integration of AutoSeC into existing wirelessnetwork infrastructures.

The rest of this paper is organized as follows: In Section 2,we discuss strategies for system status collection; inSection 3, we develop information collection policies thatare driven by user mobility. We evaluate the impact ofinformation collection algorithms on the performance ofQoS-based resource provisioning and discuss the results inSection 4. Section 5 presents AutoSeC, a middlewareframework for dynamic service brokering in wired/wire-less environments and discusses implementation issues andthe integration of AutoSeC into existing wireless networkinfrastructure. We present related work in Section 6 andconclude with future directions in Section 7.

2 SYSTEM STATUS INFORMATION COLLECTION

Information used in QoS-based resource provisioningincludes network parameters (e.g., available link bandwidth,

HAN AND VENKATASUBRAMANIAN: INFORMATION COLLECTION SERVICES FOR QOS-AWARE MOBILE APPLICATIONS 519

Fig. 1. A high level system overview.

Fig. 2. Approaches presented in this paper.

link delay), server characteristics (e.g., current server load,server resource availabilities), client characteristics (e.g.,resource constraints, power levels, locations), and content-specific attributes (e.g., number of replicas of video file, theirlocations), etc. Traditionally, much of this information isgathered and maintained independently; for instance, net-work topology can be maintained by routing informationexchange, a replica map can be obtained from a distributeddomain name service, network management software keepstrack of topology and link parameters, load balancingservices maintain server load patterns, and content manage-ment and replication services keep track of data placementand distribution. Integrating the above information into acommon repository (database) has several advantages. First,effective end-to-end QoS provisioning algorithms canexploit information from various levels for better systemutilization [4]. Second, keeping track of dynamic changes inserver, network, and content availability can be decoupledfrom policies for resource provisioning. This provides for aclean separation of concerns in system design. Third, usingwell-defined and uniform representations for the informa-tion allows easier manageability of data. Fourth, knowledgeof cross-layer information allows for flexible and efficientdata collection. Such knowledge allows us to tailor theaccuracy of the data in the database based on applicationneeds, collection overhead, and connectivity conditions, e.g.,we may relax collection parameters when user QoS needs arenot stringent, which will reduce collection overhead in analready congested network.

Dynamic changes in system status must be capturedrapidly with low overhead to allow applications to adapt tothe changing conditions. The need for flexible informationcollection is further aggravated in mobile environmentswhere the degree of dynamicity can vary significantly overtime in different parts of the underlying system. The moreaccurate the database, the easier it is to satisfy more user

requests in a timely fashion, but a higher overhead isinvolved in maintaining the database. An important issuehere is to cost-effectively maintain imprecise system stateinformation without degrading the performance of resourceprovisioning. In this section, we elaborate on variousstrategies for information representation and collection thataddress the cost-precision tradeoff.

Information representation (Fig. 3) and collection strate-gies are often intertwined. Any parameter in the databasecan be represented using either a single instantaneous value(i.e., the last measured value) or by a range-basedrepresentation that approximates the value of a parameterby using an interval with an upper (U) and a lower (L)bound. The range size may remain static or changedynamically. Corresponding collection policies determinewhen and how often to sample the network components forcurrent status information and whether to update thedatabase with the collected sample. Sampling periods maybe fixed or may vary over time. We classify policies forinformation collection as shown in Fig. 4.

System Snapshot-Based Information Collection (SS). Inthis policy [5], system status information is stored in thedatabase as an absolute value based on periodicallysampling network and server nodes. The sampling fre-quency solely determines the accuracy of the information.To prevent information from being outdated, the samplingfrequency should be very high; however, this would entailhigh overhead. Hence, a sampling period needs to balancethe tradeoff between information freshness and informationacquisition cost.

Static Interval-Based Information Collection (SI). Inthis policy [6], we define a fixed interval B to partition thecapacity of the collected information into a fixed number(say n) of equal size classes:

ð0; BÞ; ðB; 2BÞ; ð2B; 3BÞ; . . . ; ððn� 2ÞB; ðn� 1ÞBÞ:

The residue capacity information is indicated by thecorresponding class indices 0; 1; 2; . . . ðn� 1Þ. A probe isinitiated at each sampling interval to obtain currentinformation from the managed entities. If the obtainedvalue is out of the range indicated by current index, thedatabase is updated with another index, otherwise, noupdate is needed.

Dynamic Range-Based Information Collection. In thisapproach, the database holds the monitored informationusing a range with an upper bound U and a lower bound L.The central value of the range (ðU þ LÞ=2) is returned to


Fig. 3. Various data representations.

Fig. 4. General information collection policies studied.

represent current status. The range may be modifieddynamically based on the sampled information.

The collection process follows the state transition shownin Fig. 5: Regular Probing (RP), Steadiness Confirmation(SC), Transient Noise Filtering (NF), and Range Adjusting(RA). Initially, the system is in the RP state. At everysampling interval, the monitoring module sends out probesto collect current status information. If the sampled valuefalls within the current range, the system transitions to theSC state. In the SC state, the range size is tightened toprovide higher accuracy if steadiness of state is confirmed;otherwise, the system goes back to the RP state and waits formore input. However, if the sampled value in the RP statefalls outside the current range, the system goes to theNF state. In practice, the transition to the NF state can becaused by several events—measurement error, a transientburst, or a significant load level change. In order to assist theunderlying model to adapt to a confirmed change and filterout high frequency traffic components, the monitoringmodule sends Np additional probes in period Tprobe, whereTprobe ¼ T=Np with T as the original sampling period. If asignificant change is confirmed, the system proceeds to theRA state to expand the range; otherwise, it goes back to theRP state.

We have explored two dynamic range-based collectionalgorithms: 1) A Throttle-based approach (TR) which usesperiodic sampling and adjusts ranges by looking at theaverage values of recent historical samples, and 2) a timeseries (moving average) based policy (MA) [7], which appliesa prediction model using moving average to adjust bothrange size and sampling period dynamically. The two policiesdiffer in how and when range adjustment is performed. Ourperformance studies show that TR outperforms MA for QoS-based resource provisioning [8], so we restrict our discussionson dynamic range-based approaches to TR. TR uses theaverage value of samples in the previous monitoring windowto decide whether a range adjustment is needed. If deemednecessary, the range may be increased or decreased exponen-tially using a prespecified throttle factor.

Range tightening in TR. In TR, a history table is updatedwith the sampled value vðsÞ and its time stamp tðsÞ. Thehistory table is searched in inverse time order to find out thetime period tp during which the sampled values fall intocurrent range. If tp � Tsteady, where Tsteady is a giventhreshold parameter, then we anticipate more steadinessfor the link in the future and the current range Ra is

tightened. To tighten the range, we first predict a futurevalue vn, averaged from samples taken within thespecified sliding window Twindow: vn ¼ AVGðvðsiÞÞ, wheretðsiÞ > tðcurrentÞ � Twindow. Next, the range is tightenedaccording to a predefined throttle factor THr to get a newrange Rn: Rn ¼ Ra � THr. Finally, the new range Rn isadjusted such that the average value vn lies in the middle ofthe new range: UðRnÞ ¼ minðvn þ 0:5 �Ra; capacityðlinkÞÞ,LðRnÞ ¼ maxðvn � 0:5 �Ra; 0Þ. If tp < Tsteady, the processgoes back to the RP state and waits for more input.

Range relaxation in TR. To confirm significant changes,additional probes are sent out in the NF state. We averagethese additional sampled values to determine if it fallswithin the current range Ra. If yes, the noise is transient andcan be ignored; otherwise, it is a significant change and therange Ra is relaxed as follows: UðRnÞ ¼ maxðvðsiÞ; UðRaÞÞ,LðRnÞ ¼ minðvðsiÞ; LðRaÞÞ, where i ¼ 1; 2; ::; Nr and tðsiÞ >tðcurrentÞ � Twindow:

3 CUSTOMIZING INFORMATION COLLECTION FOR

MOBILE APPLICATIONS

In mobile environments, an accurate picture of the under-lying system must incorporate not only the informationabout network components and services, but also informa-tion about the mobile clients. To accommodate potentialclient mobility patterns, information collected may includeintermittent connectivity patterns, location, residual energylevel, etc. QoS-based resource provisioning can be enhancedby adapting to these constantly changing factors. Forinstance, accurate information about the residual energylevel of mobile clients can be used to decide how therequested content should be delivered (e.g., burst sizes andschedules) and information about client connectivity can beused to provide effective synchronization and caching.Similarly, accurate information about current and futurelocations of mobile clients can be exploited to providereservations for smooth handoff. We are specificallyinterested in location information since this has implica-tions on which links in the wired network will be used toserve the mobile user. In this section, we discuss differentapproaches to collecting location information for mobileapplications and then present techniques for improvedsystem status collection using location information.

3.1 Approaches for Collecting Location Information

We explore two ways using which location information canbe collected. Fine-grained approaches maintain currentlocation of each individual mobile client [9], [10], whilecoarse-grained collection captures information at an aggre-gated level for multiple clients. In particular, we use thenotion of client aggregation (i.e., mobile client population ineach cell at a certain time instant) as a coarse measurementfor location information.

Fine-grained location information management hasgained a lot of attention from researchers over the lastfew years. It typically involves three issues [9], [10]: locationupdate strategies which decide when mobile users shouldinform the network about their current locations, pagingstrategies which decide when the base station should sendout queries to search for the mobile user, and locationinformation maintenance architectures which decide how to


Fig. 5. State transition diagram for dynamic range-based informationcollection.

store and disseminate the location information. In addition,user mobility patterns (such as [11]) are studied concur-rently in order to better capture current user location.

Our strategy is to ensure resource availability for mobileapplications by providing reasonably accurate system statusinformation. We refrain from using fine-grained locationinformation to determine overall resource availability.Perturbation of residual resources caused by a singlemobile user is almost negligible; furthermore, keeping trackof individual user mobility may entail significant overhead,since we need to probe each mobile host separately andconstantly. Hence, gathering individual user location doesnot benefit our overall goal of efficient system statuscollection.

However, the movement of a large number of users maylead to nonuniform distribution of mobile users across cells.We conjecture that using coarse-grained mobility informa-tion to adjust collection parameters will be sufficient tosupport cost-effective resource provisioning. In addition,collection using coarse-grained mobility information isindependent of individual mobility models—this relievesus from the inaccuracies introduced in modeling orpredicting individual user mobility. One measure ofmacro-level changes in mobile settings is the client aggrega-tion status (i.e., number of users in a cell). This has thepotential to significantly affect resource availability in thenetwork/system [1], [2]. Client aggregation status can beobtained from cellular access points (i.e., base stations) thatmanage the communication of the mobile hosts residingwithin that cell. The base stations can apply a simplisticstrategy (e.g., an update from the mobile host is triggeredwhen handoff occurs) to maintain the total population ofmobile hosts or more complex prediction-based approaches.

Fig. 6 shows the results of a simple experiment thatcompares the overhead of collecting coarse-grained (aggre-gate) and fine-grained (individual) location information.Without loss of generality, we applied the system snapshot-based collection policy to gather mobility information. Weset the sampling period to be 1 second and count thenumber of total sampling messages as the metric forcollection overhead. In our experiment, the coverage areaconsists of 10 cells. In order to obtain coarse-grainedmobility information (i.e., number of users in a cell), weprobe each base station every second. Since the number ofbase stations remain unchanged irrespective of the changes

in the number of mobile hosts, the sampling cost involvedremains 10 per second; however, in the case of collectingfine-grained mobility information, we need to sample allthe mobile hosts and the number of sampling messagesincreases linearly as the number of mobile hosts increases.

In the absence of this information from base stations(which requires tight coordination with the service provi-der), it may be possible to derive/predict aggregate mobilitystatus from individual mobility models. Prediction ofaggregate status requires some knowledge about the dis-tribution of mobile hosts and their mobility patterns in aregion. For example, the incremental mobility model [11] isoften used to capture the mobility pattern of users in ametropolitan area where the mobility of most users isrestricted by the layout of streets and freeways. In thismodel, mobile hosts are distributed and can move randomlyin a closed coverage area with the size of ðXmax; YmaxÞ. Thelocation of the mobile host ðx; yÞ and its velocity vector (bothspeed and direction) incrementally change as time pro-gresses. Change of the location is determined based onprevious velocity and direction; change of velocity isuniformly distributed between the minimum and maximumacceleration/deceleration of the host.

We now describe the derivation of aggregate mobilityfrom the incremental individual mobility model; similarderivation can be obtained for other mobility models. Theclosed coverage area is divided into a number of nonover-lapping equal sized regions with the size ðXregion; YregionÞ.Each region has a collection point (e.g., base station) thatserves as the wired network access point for all the mobile

hosts in its controlled region. If we let Xdim ¼ d Xmax

Xregione and

Ydim ¼ d YmaxYregione, then there are Nregion ¼ Xdim � Ydim regions in

the area. Mobile hosts either move or stop in this area, thepopulation/aggregation in each region changes all the timeand can be captured. At a certain time t, a mobile host

located at ðxðtÞ; yðtÞÞ is in region b xðtÞXregion

c þXdim � b yðtÞYregionc. At

time t, the aggregation of region i (AiðtÞ) is the number ofmobile hosts located in region i, i.e., the cardinality of theset of the mobile hosts who are in this region.

AiðtÞ¼ jxjðtÞXregion

� �� ¼ i mod Xdim;yjðtÞYregion

� �¼ i

Xdim

� ��

��: ð1Þ

Once the coarse location information is obtained (eitherfrom base stations or via model-based predictions), wewould like to utilize it to enhance system status collection.

3.2 Mobility-Aware System Status InformationCollection

Given that sampling frequency and range size are thetwo most important factors that impact the efficiency of thecollection process, in this section, we present a family ofcollection strategies (Fig. 7) that uses client aggregationstatus to drive the adjustment of sampling frequency andrange size. These policies are not only independent of theindividual mobility model adopted, but are also expected toprovide reasonable levels of data accuracy with signifi-cantly less overhead, enabling efficient QoS support. Sinceour empirical results (presented later) show that dynamicrange-based collection policies outperform the instanta-neous value-based or static interval-based approaches, we


Fig. 6. Overhead comparison of collecting coarse-grained versus fine-grained information: The overhead is the number of probing and updatemessages normalized over the number of messages used for coarse-grained mobility information collection.

restrict our discussion to customizations for the dynamicrange based policy.

We consider all the components in the system thatprovide information of interest to users as informationsources, while all the components that use the availableinformation to make corresponding decisions are referredto as information consumers. The consumers can be middle-ware services such as QoS-based resource provisioningalgorithms or user applications such as location-awareinformation delivery. Consumers may have varying dataaccuracy needs that are different from database values,which will cause consumer-initiated triggers and updates.Similarly, information sources may be intelligent enough tonotify the collection process of sudden changes in theirstates, which will cause source-initiated triggers and updates.

Fig. 7 illustrates the various aggregation-based informa-tion collection techniques proposed in this paper. In the PureAggregation-based (Pure-ABIC) approach, the collectionprocess adjusts sampling frequency and range size based onmobile host aggregation status and current resource utiliza-tion. The Opt-ABIC [12] approach extends the Pure-ABICapproach to accommodate heterogeneity of both informa-tion source and information consumer. In addition, wepropose a Cost Driven Aggregation-based (CDIC) approach[13], where collection cost is analyzed to derive an optimalrange size and to explore the conditions under whichmaintenance overheads may be further reduced.

Pure Aggregation-Based Information Collection (Pure-ABIC). Pure-ABIC utilizes the combination of mobile hostaggregation status and resource (server/link) utilization todynamically adjust the collection parameters such assampling frequency SF and range size R. It is a relativelysimple policy that assumes uniform consumer needs andhomogeneous source behavior. We classify aggregation andresource utilization levels into three categories—high,medium, and low. The threshold utilization values thatdemarcate the boundaries between high, medium, and loware implementation-specific and dependent on the trafficcharacteristics and system capacities. It is difficult todetermine an optimal threshold since system conditions(such as resource availability and traffic patterns) arevarying dynamically. In our simulation, we consider theutilization level higher than 0.9 as “High” and that lowerthan 0.4 as “Low.” These parameters are chosen based onour previous experimental studies [13], where we observethat when the utilization level goes above 0.9 or below 0.4,the system is more sensitive to sampling frequency andinformation granularity. In the case of AL (AggregationLevel), assuming that Nmh is the total number of mobilehosts and Nregion is the total number of regions, we use Nmh

Nregion

and Nmh

8 as two boundary values. The choice of theseparameters is obtained by observing the traffic variation andmobility patterns under the chosen mobility model in thesimulation. Using different parameter settings and/or othermobility models could yield different threshold values.

When resource utilization is high or low, the aggrega-tion levels have little impact, since most of the requestsare either rejected or accepted. The choice of a small SFand a small R is desirable. When resource utilization ismedium, aggregation status comes into play: We increaseSF and decrease R when the aggregation level increases.In other words, SF ðtÞ ¼ minfð1þ �Þ � SF ðt� 1Þ; SFmaxgand RðtÞ ¼ minfð1þ �Þ �Rðt� 1Þ; Rmaxg, where � is set tobe �H ¼ 0:2, �M ¼ 0:4, and �L ¼ 0:8, respectively, forhigh, medium, and low aggregation levels. Similarly, � isset to be �H ¼ 0:8, �M ¼ 0:4, and �L ¼ 0:2, respectively,for high, medium, and low aggregation levels.

Optimized Aggregation-Based Information Collection(Opt-ABIC). In Pure-ABIC, the adjustment of informationcollection parameters is performed by a local collectionmodule. This localized decision making reduces commu-nication overhead among different components in thesystem; however, when information sources and consumersparticipate in determining collection parameters, the systemis expected to accommodate to unexpected changes insystem load and consumer specific accuracy needs. Thedesign of a more generalized Aggregation Based Informa-tion Collection (Opt-ABIC) grew out of this observation.

In addition to coarse-grained adjustment of collectionparameters driven by aggregate mobility status as in Pure-ABIC, Opt-ABIC (Fig. 8) utilizes feedback from the sourcesand the consumers for further customization of thecollection process. Every source maintains the current exactvalue and the approximation is held in the database. Thereare three reasons that may cause the exact source value todeviate from the stored interval: 1) resource reservation,2) resource release, and 3) changes in application load.Application load is dynamic and is often caused byunmediated requests (those without QoS needs) that areaccepted or completed. These best-effort requests are notaccounted for by the resource provisioning process sinceadmission control is not required. When the current valuechanges unexpectedly and falls outside of the current range,a source-initiated trigger is issued to the collection module forpotential range size relaxation. Every so often, admissiondecisions cannot be made based on the approximateinformation in the database. In this case, the resourceprovisioning process (which is the information consumer inour case) may request data at a certain level of accuracy,generating a consumer-initiated trigger. The collection process


Fig. 7. Evolution of information collection policies for mobile applications.

responds to these two types of triggers by determining if therange size should be adjusted. The adjustment process andsubsequent collection are similar to the dynamic range-based collection approach discussed in the previous section.

Cost Driven Aggregation-Based Information Collection(CDIC). In this approach, we analyze the cost factorsinvolved in a generic information collection process andincorporate the cost measures into Opt-ABIC. Careful studyof a generic information collection process reveals that thecommunication cost involved (Fig. 9) consists of:

. Regular sampling overhead Crs: This is introducedby probing messages issued by the resource provi-sioning module to sources.

. Regular update overhead Cru: This is introduced byupdate messages which report current source valueto the resource provisioning module and update thedatabase with the new value.

. Source/consumer-initiated trigger overhead Cst andCct: This is introduced by messages from thesources/consumers to the resource provisioningmodule that can potentially cause range expansionand tightening.

. Source/consumer-initiated database update over-head Csu and Ccu: This is introduced by the messagessent by resource provisioning module to updatedatabase with the new range.

Overhead incurred by regular sampling and update (Crsand Cru) is dependent upon the sampling frequency, whichis driven by mobile host aggregation status and actualresource utilization as shown in the previous section.Therefore, to minimize the overall communication cost,we need to minimize the overhead introduced by source/consumer-initiated trigger/update. We observe that bothtypes of triggers/updates (source and consumer-initiated)provide an opportunity for the collection module to adjustthe range size in the database. Our objective in selecting agood range size is to avoid the need for future updates and,consequently, minimize the communication costs. To avoidsource-initiated triggers and updates, the range size shouldbe large enough to incorporate changes in source values. Onthe other hand, to avoid consumer-initiated triggers andupdates, the range size should be small enough to satisfyconsumer accuracy constraints. Hence, it is not obvioushow to choose a range size that minimizes the possibility ofboth source and consumer-initiated triggers and updates.

In Opt-ABIC, one source/consumer-initiated trigger doesnot necessarily cause one source/consumer-initiated updatesince a significant change needs to be confirmed. Unlike

Opt-ABIC, CDIC triggers always result in updates. Ourobjective here is to minimize the weighted sum of consumer-initiated cost (Ccc ¼ Cct þ Ccu) and source-initiated cost(Csc ¼ Cst þ Csu). Assuming that Pcc, Psc represent thefrequency of consumer-initiated triggers/updates andsource-initiated triggers/updates, we can rewrite the costformula as � ¼ Ccc � Pcc þ Csc � Psc. Intuitively, a smallerrange size would lead to more frequent source-initiatedtriggers/updates and less consumer-initiated triggers/up-dates. Therefore, the range size needs to be adjustedcarefully in order to minimize the overall overhead.Previous studies [14] have shown that, in approximatecaching, for a given R, we have Psc ¼ Ksc=R

2, Pcc ¼ Kcc �R,where Ksc, Kcc are parameters that depend on the nature ofthe source and the frequency of consumer requests and thedistribution of consumer requirements. This result can beapplied in our context here. Hence, we can rewrite the costexpression as � ¼ Csc �Ksc=R

2 þ Ccc �Kcc �R.

Finding the value for R that minimizes this expression

can be reduced to finding the root of the derivative. Using

this approach, the optimal value for R is ð2 � CscCcc� Ksc

KccÞ1=3. Even

though consumer-initiated cost Ccc and source-initiated cost

Csc vary depending on system conditions, we can still

reasonably assume the ratio CscCcc

to be constant. This implies

that there is no need to choose values for each type of cost.

For example, consumer-initiated cost is 2 messages (i.e.,

1 message from information consumer to the collection

process; and 1 message from the collection process to the

database). Similarly, source-initiated cost is also 2 messages.

Hence, CscCcc¼ 1.

Setting the optimal range size based on the cost analysis,assume that source behavior and application workloadsremain stable. In practice, parameter values and applicationneeds may change dynamically at runtime. The overhead ofmonitoring these factors at runtime in order to set Kst, Ksu,Kct, and Kcu appropriately affects the system performancesignificantly. Therefore, we apply the following strategies tofurther decrease the communication cost without sacrificingoverall QoS:

Selective Triggering. The triggering process can be sig-nificantly optimized to suit the needs of the informationconsumer (the resource provisioning process in our case).The approach described below is especially suitable inmixed wired/wireless networks. Consumer initiated trig-gers are used when the consumer determines that thesupplied data is not accurate enough. Such consumerinitiated triggers are impractical in the case of link relatedinformation due to the following reasons: 1) QoS-basedrouting techniques select a path based on aggregate link


Fig. 8. State diagram of Opt-ABIC.

Fig. 9. Cost factors in Opt-ABIC.

characteristics, hence it is difficult to determine whethermore accurate information is needed for links on anindividual basis, and 2) since the number of links can belarge, customizing the collection process on a per-link basisis not a scalable proposition.

Furthermore, we observe that consumer initiated triggersmay be necessary only when the required resources for arequest closely match the available resources (if there areample available resources, the request is likely to beaccepted anyway). In such a case, we can assume that theserver or the network is overloaded.1 At this time, there is ahigh likelihood that incoming requests are rejected; hence,maintaining accurate database information will have littleuse. For the above reasons, we expect that turning offconsumer-initiated triggers and updates would not degradeQoS-provisioning performance greatly, but the collectionoverhead can be significantly decreased.

Lazy Sampling. Once consumer-initiated triggers andupdates are eliminated, the database modifications canonly be driven by the regular probing process or by source-initiated triggers. To further reduce communication over-head, we apply lazy sampling strategies that can decreasethe probing cost even more. The sampling frequency can bereduced in two cases: 1) If the number of source-initiatedtriggers in a given period is less than a predeterminedvalue, we can infer that the current approximation of thesource value is reasonable (i.e., the source value does notchange very dramatically), therefore, we can reduce thesampling frequency, and 2) if the range is relaxed to exceeda certain value (1=4 of the maxima as indicated in ourexperiments), it is likely that the range is large enough toaccommodate reasonable changes in the source value, hencewe can reduce the sampling frequency. In both cases, weuse the following equation to reduce the samplingfrequency: SFnew ¼ SFold

1þ� .A more focused study of this approach as applied to

network management appears elsewhere [13].

4 PERFORMANCE EVALUATION

Information collection is an underlying support service, soits effectiveness is reflected by improvements in overallsystem performance. In this paper, we use QoS-basedresource provisioning as a sample information consumerand evaluate the impact of the proposed informationcollection policies on QoS-aware mobile applications. Theinteraction between the information collection service andresource provisioning process is as follows. Resourceprovisioning policies consider current residual networkand server resources as well as mobile client aggregationstatus to ensure better overall performance. Given theserver and/or path assignment, the client proceeds to set upa connection along the assigned network path from the basestation where the mobile host resides to the server. Duringthe connection setup, the routers and the servers along thepath check their residual capacity and either admit theconnection by reserving resources or reject the request.When a client moves out of current region/cell, anotherpath/server that better serves the client is found andallocated accordingly. When the connection terminates, the

client sends termination requests and resources arereclaimed along the connection path.

Resource Provisioning Policies Studied. We evaluatethe following three representative resource provisioningpolicies that cover a range of mobile applications incomposition with various information collection techni-ques: 1) server selection (load based), 2) server selection(proximity based), and 3) combined path and serverselection (CPSS).

Server Selection (SVRS). As distributed informationservices like the World Wide Web become increasinglypopular on the Internet, scalability becomes a very seriousissue. Performance degradation can be caused by excessiveserver load due to the request of popular documents,wasted network bandwidth due to redundant documenttransfer, and excessive latency in delivering documents tothe client due to the potential for transfers over slow paths.A promising technique that addresses all of these problemsis service/document replication where the same data/service is partially replicated across distributed servers.However, when a service is replicated, clients must be ableto discover an optimal replica. Server selection schemesattempt to choose the best replica/server for a givenrequest. Although requests from all the mobile clientswithin the same cell go through the same base station,different servers may be chosen for different mobile clientsfor improved load balancing. We apply the followingtwo server selection policies for evaluation.

. Least Utilization Factor Policy (SVRS-UF). In thispolicy, the capacity of a server can be specified asfour parameters: CPU cycles, memory buffers, anddisk bandwidth as shown in [15]. In determining theserver utilization factor, the bottleneck resource (i.e.,one that is impacted the most) is conservatively usedto estimate the impact of the request on a server.This policy chooses the server with the minimalutilization factor. If there is a tie between the servers,the policy randomly picks one.

. Shortest Hop Policy (SVRS-HOP). This policychooses the nearest server to the mobile host interms of the number of hops. If there is a tie betweenthe servers, the policy chooses the least loaded oneas in SVRS-UF described above.

Combined Path and Server Selection (CPSS). Serverselection algorithms discussed above are often used todirect user requests to the optimal server based on chosenmetrics (such as proximity or load) when data is replicatedacross multiple servers. These mechanisms often treat thenetwork path leading from the client to the server as static.While this is useful for computation-intensive applications,interactive applications such as mobile multimedia mustguarantee the availability of network resources as well.QoS-based routing techniques typically aim to select theoptimal path between a source-server pair and ignore thesituation where there may be multiple servers that can servethe same request. CPSS [16] is an integrated approachwhich allows load balancing not only between replicatedservers, but also among network links. This has thepotential to achieve higher system-wide utilization andallows more concurrent users. We have developed exten-sions for CPSS that account for mobile user aggregationstatus in a cell and adapt well to request locality (i.e., usersin a region may request similar information). Previous workhas studied a family of CPSS policies; in this paper, we


1. Typical request sizes and system configurations are such that anincoming request will occupy only a small fraction of the total systemcapacity (most servers and networks are designed to deal with hundreds ofrequests.)

choose the policy which selects the server and links thatmaximize the overall utilization of resources.

We now describe the simulation setting used forperformance studies in this paper, and then present theextensive simulation results and analysis, whose objective isto determine the most appropriate information collectionpolicy for a given resource provisioning mechanism basedon current system state and application workload.

4.1 Simulation Environment

Our simulator is a message driven multithreaded systemintended to study the dynamics of QoS sensitive networkenvironments. We use the cellular architecture in which eachcell has one base station as its access point to wired networks(Fig. 10). Eight hundred mobile hosts are distributed in20 cells. The movement-based approach [17] is used formobile hosts to perform location update. In other words, wecount the number of cell boundary crossings for each mobileuser and the location is updated when this counter exceeds apredefined threshold (five in our simulation). Note that ouraggregation mobility-based approach is in fact independentof underlying location management scheme used, since itdoes not need the knowledge of each mobile host location atany time instant. Therefore, we resort to a simple yetrelatively reasonable scheme, movement-based updates, forour simulation purpose. The choice of five as the threshold forcell boundary crossing is based on our observation of trafficvariation and mobility patterns under the chosen incrementalmobility model and corresponding parameters in ourexperiments. In addition, ideally, QoS-based resource provi-sioning in mobile environments integrates appropriatepaging schemes to determine the location of mobile hosts inadvance and incorporates predictive resource reservation.However, the focus of this paper is on information collectionservice itself instead of novel resource provisioning ap-proaches, hence, we abstracted out paging schemes. Thelocation management is purely triggered by location updaterequests from mobile hosts. Furthermore, we model twolevels of mobility: high mobility and low mobility. Specifi-cally, we set maximal velocity to be 65 mph, maximalacceleration/deceleration to be 0:9meter=sec2, and maximaldirection change per unit time to be 0:1745radian=sec. Withhigh mobility, the velocity change is twice that of lowmobility.

To better emulate the real network, the capacities ofnetwork links are selected from various widely used link

types from 1.5Mbps to 155Mbps, with the mean value being64Mbps. When defining the capacity of the server nodes, wecalibrate CPU units using a basic 64Kbit voice processingapplication, memory units to be 64Kbytes, and diskbandwidth units to be 8Kbytes/s. The server capacities arealso selected from popular models of multimedia servers andweb servers, with the mean CPU, memory, disk bandwidthto be 1,845, 6,871, and 5,770 calibrated units, respectively.

We model request arrival at the source nodes as aPoisson distribution, and the request holding time isexponentially distributed with a prespecified average value.The request holding time is the time for which therequested network and server resources, such as linkbandwidth, CPU, buffer, disk bandwidth, etc., are reserved.We predefine a set of client request templates to capturetypical multimedia connection request patterns in terms ofnetwork bandwidth, CPU, memory, disk bandwidth, andend-to-end delay. For each request generated, the requestedparameters are randomly selected from the set of requesttemplates, with the mean requested bandwidth being2.5Mbps, mean end-to-end-delay being 400ms, and CPU,memory, and disk bandwidth being 150, 374, and 271 cali-brated units, respectively.

Network traffic is classified into two categories: nonuni-form traffic (NUT) and uniform traffic (UT). To representnonuniform traffic, we designate some sets of candidatedestinations as being “hot” (i.e., serving popular videos, Websites, etc.), and they are selected by the clients more frequentlythan others. To reduce the effect of local and nearby requests,we choose three pairs of source-destination sets from thetopology. The requests arrive at these hot pairs, as foregroundtraffic, at a higher rate than other background traffic. In ournonuniform traffic pattern, we set the foreground arrival rateto be five times higher than the background rate and, inuniform traffic pattern, we set them to be equal to eachother. Specifically, we set the foreground arrival rate to be10 seconds, and the background rate to 50 seconds.

Application workload can be classified into two cate-gories based on the types of requests sent to network:1) Pure QoS requests: Servers such as VoD servers andcomputation bounded servers are designed for a specificpurpose and thus only accept requests with QoS require-ments, therefore, all the changes on the server load areexposed to resource provisioning; 2) Mixed requests:Servers like distance learning servers are general-purpose,i.e., they accept both QoS-based requests and best-effortrequests. Generally, there is no admission control for best-effort requests, which makes the server load change withoutthe notice of the resource provisioning process. For thissimulation, we use the trace [18] to represent the server loadchange (Fig. 11). Table 1 depicts all the different scenarioswe studied in this paper.

4.2 Performance Results

The following three metrics are used to compare theperformance of resource provisioning mechanism:

. Request success ratio: It is the ratio of the number ofsuccessful requests over the number of all requests.In traditional networks where no mobile clients exist,a request is considered successful if resources alongthe chosen path and server are successfully reservedfor it. In mobile environments, a successful request is acompleted request. Since mobile hosts constantlymove, dealing with request handoff is unavoidable.


Fig. 10. Simulation Topology: The base stations are connected as in atypical MCI ISP network. There exist a number of mobile devices, fixedhosts, and servers in each cell.

Incoming requests go through an admission controlprocess, where they may be accepted or rejected. Anadmitted request executing on a mobile host willrequire reprovisioning of resources (network links,servers, etc.) as it migrates through different regions.The intermediate reprovisioning may or may not besuccessful. Admitted requests may not complete dueto several reasons: there is no route with sufficientresources (path failure), locating mobile hosts fails(location failure), or the alternate rescheduling servermay not have sufficient resources if path to originalserver is not available.

. Information collection overhead: Basically, it is the totalcommunication overhead involved in the wholecollection process, including the cost incurred bysampling information sources and updating data-bases. Note that the collection overhead consists ofboth system status collection overhead and mobility-related information collection overhead. When fine-grained mobility information is maintained, the costinvolves all the overhead of communicating witheach individual mobile host; on the other hand, withcoarse-grained mobility, the cost refers to the over-head introduced by maintaining region or cell-basedmobility. For simplicity, we assume the collectionoverhead is proportional to the number of messagesexchanged among different system components, i.e.,it is the sum of the number of samplings, the numberof database updates, and the number of triggersfrom both sources and consumers.

. Overall performance efficiency: It is a metric defined asthe ratio of the number of successful requests to theinformation collection overhead.

4.2.1 Basic Performance Results

In the previous sections, we presented three differentapproaches to system status information collection: mobilityincognizant collection, collection using fine-grained mobi-lity information, and collection using coarse-grained mobi-lity information. We now compare the performance ofresource provisioning under these approaches. We use theThrottle-based (TR) algorithm as the representative formobility incognizant collection policies, since it can beeasily extended and customized to incorporate mobilityinformation. We use Pure-ABIC to represent the approachesthat use coarse-grained mobility information.

From Fig. 12, we observe that the completion ratio ofCPSS under the three approaches is close to each other;however, using fine-grained mobility information intro-duces significantly higher overhead, while using coarse-grained mobility information incurs the lowest overhead.This demonstrates the effectiveness of utilizing coarse-grained mobility information, which, in turn, exhibits thelowest overall efficiency. This result remains the same withSVRS-UF or SVRS-HOPS under other scenarios.

4.2.2 Performance Comparison of Underlying Collection

Policies

The system status collection policies (SS, SI, TR) presentedin Section 2 are the underlying policies based on which thecoarse-grained location information driven system statuscollection approach is customized. We conducted a series ofexperiments to determine the best combination of acollection policy (SS, SI, TR) and a resource provisioningpolicy (SVRS-UF, SVRS-HOP, CPSS) under varying applica-tion workload. For fair comparison, we compare thesecombinations using the parameter settings that exhibit thebest performance for each of them. Performance studiesindicate that the best overall efficiency is achieved by using10 seconds as the sampling period for snapshot-based (SS),static interval-based (SI), and throttle-based (TR) collectionpolicies. In our experiments, we observed that the perfor-mance comparison of different policy combinations undervarying scenarios is very similar, so we only show theresults of case M1 (high mobility, nonuniform traffic, pureQoS requests) here.

Fig. 13 shows the performance of SVRS-UF with thethree information collection policies under case M1. Thecompletion ratios do not show much difference. However, itcan be seen that SS introduces the highest overhead and TRincurs the lowest overhead. In terms of overall efficiency,TR outperforms the other policies.

We study the performance of CPSS under differentinformation collection policies and scenarios. Fig. 14 showsthe performance of CPSS with the three informationcollection policies under case M1.

The SS, SI, and TR algorithms exhibit similar completionratios to each other. In general, the overhead of all thestrategies increases with an increase of the number ofrequests because more dynamic change is expected withmore requests and the collection process performs moremonitoring. While SI and TR incur a similar amount ofoverhead, as expected, the SS algorithm introduces far moreoverhead than the other three due to frequent sampling anddatabase updates containing exact values. Hence, SS exhibitsthe lowest overall efficiency among all the three algorithms.The efficiencies of the dynamic range-based algorithm and


Fig. 11. Server load change pattern: It is a day’s worth of all HTTPrequests to the EPA WWW server located at Research Triangle Park,North Carolina.

TABLE 1Scenarios Studied

static range-based algorithm are very close to each other. Thereason is that, in mobile environments, we must account forhost mobility in addition to network and link changes. Not allrequests admitted can be completed and the range changemay not keep up with the movement of mobile hosts, hence,changing the range size does not always exhibit better results.

Among the three mobility-incognizant policies we studiedhere, TR outperforms the others in terms of overall efficiency;therefore, we use TR as underlying policy to supportintegration with coarse-grained mobility information.

4.2.3 Impact of Integrating Coarse-Grained Mobility

Information

In Section 3, we presented three different ways to customizeTR to take into account coarse-grained mobility informa-tion. We now use experiments to determine the bestcombination of a coarse-grained mobility informationdriven collection policy (Pure-ABIC, Opt-ABIC, CDIC)

and a resource provisioning policy (SVRS-UF, SVRS-HOP,CPSS) under varying application workload.

Fig. 15 shows the performance of SVRS-UF with thethree information collection policies under case M1. Thecompletion ratios of the three policies are similar to eachother. CDIC and Pure-ABIC have much lower overheadthan the others, which results in highest overall efficiency.Overall, Pure-ABIC performs the best for SVRS-UF andCDIC also performs comparably. CDIC has the advantagethat it gives more flexibility to information sources andconsumers to control the collection process.

We study the performance of CPSS under differentinformation collection policies and scenarios. Fig. 16 showsthe performance of CPSS with the three informationcollection policies under case M1. In general, the completionratio of the three policies is very close and they follow asimilar trend: With an increase of the number of requests inthe system, the completion ratio decreases. As the system


Fig. 12. Basic performance comparison.

Fig. 13. Performance of SVRS-UF with different underlying collection policies under Case M1 (high mobility, nonuniform traffic, pure QoS requests).

Fig. 14. Performance of CPSS with different underlying collection policies under Case M1 (high mobility, nonuniform traffic, pure QoS requests).

saturates (the number of requests reaches 500 in this case),the completion ratio reaches a minimum and then risesgradually due to completing requests.

The overheads of Pure-ABIC and CDIC are low and closeto each other. The reason for the low overhead of Pure-ABIC isthat coarsely selecting collection parameter such as samplingfrequency and range size based on the aggregate mobilitystatus and resource utilization status is good enough torender satisfactory completion ratio; adding source andconsumer-initiated triggers and updates just make thecollection process more complex and adds more overheadwithout increasing the completion ratio. Overall, Pure-ABICshows the highest efficiency, which is followed by CDIC; Opt-ABIC and CDIC exhibit much lower overall efficiency.

In summary, if balancing the tradeoff between systemperformance and the overhead is the goal, then Pure-ABICis the most desirable strategy collecting network and serverinformation in mobile environments; however, in scenarioswhere resources are very limited, CDIC is a good candidatebecause of its very low overhead; Opt-ABIC can always beused when more flexibility needs to be given to theinformation sources and consumers.

4.3 Deriving Automatic Service Composition Rules

Our evaluation indicates that both the accuracy andoverhead of information collection policies have a signifi-cant impact on the performance of resource provisioningprocess. Although snapshot-based collection can obtainvery accurate information, the huge overhead introducedby frequent sampling and database updates makes it a badchoice. In contrast, the throttle-based algorithm turns out to

be the best choice among the mobility-incognizant collec-tion algorithms. Although seemingly naive and ad hoc, itadapts very well to the constantly changing environment.We further observe that the family of aggregation-basedinformation collection policies works uniformly better thanthe mobility-incognizant policies. This finding shows theeffectiveness of utilizing mobility status in designingcollection policies.

In summary, Pure-ABIC works the best for both CPSSand Server Selection under all the scenarios studied ifoverall system efficiency is the major concern. Clientaggregation information is beneficial for collecting not onlylink resource availability status but also server resourceusage status. This is because, when user density increases ina certain region, potentially, they more frequently access theservers in that same region, which leads to more variationin server resource status. In addition, CDIC outperforms theother policies if maintenance overhead is the major metric.For example, users issuing computation-bounded requeststend to remain in a certain region for a while, i.e., they donot constantly change their locations. In this case, requestcompletion ratio is relatively stable and overhead intro-duced by collecting the information becomes the majorconcern, hence, CDIC should be used.

Using the results from empirical study, we derivegeneral heuristics to select appropriate composition ofinformation collection and resource provisioning policiesfor an incoming request. We classify the client requests into:1) Web requests that include normal HTTP, FTP, orTELNET sessions that do not present stringent and explicitQoS requirements, 2) multimedia requests that include


Fig. 15. Performance of SVRS-UF with different coarse-grained mobility information driven collection policies under Case M1 (high mobility,nonuniform traffic, pure QoS requests): Pure-ABIC shows similar completion ratios to the other policies, but performs the best overall due to itslowest overhead.

Fig. 16. Performance of CPSS with different coarse-grained mobility information driven collection policies under Case M1 (high mobility, nonuniformtraffic, pure QoS requests): Pure-ABIC exhibits the best overall efficiency which is followed by CDIC.

requests for audio/video streaming or for rendering3D graphics/images, and 3) computation requests thatinclude complex scientific computations and usually CPUbounded in nature. We also classify servers into four types:1) Web servers that support Web requests and alsomultimedia requests, 2) multimedia servers that deal withmultimedia requests, 3) computation servers that supporthigh speed computations, and 4) general purpose serversthat support all the different types of servers.

With Web requests, there is no need to reserve eitherserver or path resources, so we do not provide any resourceprovisioning or information collection policies for them. Formultimedia requests, CPSS should be used to find the bestpath to the best server since requested items are oftenreplicated on different servers and there are a number ofpaths available between the user and the servers. However,computation-bounded requests only need to reserve re-sources at the server since relatively smaller amount ofinformation is transmitted along the path. Performanceresults indicate that, given the specific resource provisioningpolicy (SVRS-UF, SVRS-HOP, or CPSS), there exists aparticular information collection policy that works the bestfor it under varying network/system conditions. This showsthat the adaptations of collection parameters in informationcollection policies are sufficient and policy switching amongcollection policies is not necessary. We propose the servicecomponent selection rules illustrated in Table 2.

5 AUTOSEC: DYNAMIC SERVICE BROKER

FRAMEWORK

The AutoSeC (Automatic Service Composition) frameworkprovides a user-transparent middleware infrastructure thatsupports automatic and adaptive selection of policies formanaging dynamic distributed environments. In particular,it supports selection of optimal combinations of informationcollection and resource provisioning services, thus shelter-ing the applications from the complexity of decision makingby hiding the diversity of the underlying provisioningservices. AutoSeC has five components (Fig. 17): informationsource, information repository, resource provisioning,information mediator, and dynamic service broker. Wedescribe these components in detail below.

The Information Source corresponds to different compo-nents in the system, such as the server, link, mobile host, orstationary host. The state information about sources in-cludes network parameters (such as residual link band-width, end-to-end delay on links, etc.), server parameters(such as CPU utilization, buffer capacity, disk bandwidth,etc.), stationary host parameters (such as client capacity,

connectivity, etc.), and mobile host parameters (such asmobile host location, connectivity, power level, etc.).

The Information Repository consists of distributed data-bases which hold system state information about informa-tion sources. In this paper, for simplicity, we assume acentralized database within which adaptive data represen-tation is supported. Each data item can be represented by aninstantaneous value or a range bounded by an upper and alower value. The information in the database is updatedbased on current network conditions and user requirementsusing different update policies.

The Resource Provisioning Module uses information held inthe database and feedback from the resource provisioningmodule to perform admission control and resource allocation.

The Information Mediator serves as the decision point forthe information collection. It listens to notifications fromsources or the resource provisioning module and invokessuitable actions so that the database maintains informationat a suitable level of accuracy. It consists of an informationcollection module and a network monitoring module. Theinformation collection module determines whether and how toupdate the database with the current system image. Thenetwork monitoring modules are distributed in the networkand each module monitors portions of the entire domainbased on sampling parameters supplied by the informationcollection module. The monitoring module issues probes tonetwork components to collect system state information.The probes consolidate the collected sample values andthen hand them to the monitoring module. The monitoringmodule returns the collected results to the informationcollection module that updates the database.

The Dynamic Service Broker implements the servicecomposition rules in Table 2. It selects the best combinationsof resource provisioning and information collection policiesto match current user requests and system conditions. Italso determines when to switch among different policies ifnecessary.

Prototype Implementation. Building AutoSeC is animportant step in understanding and controlling thecomplex dynamics of QoS-based resource provisioningunder varying levels of information precision. In ourprototype system, the information mediator and thedatabase reside on Sun Ultra5 workstations runningSolaris 2.7 and information sources execute on Pentium IIIPCs running Windows 2000 connected via a 10/100MbpsEthernet link. AutoSeC is implemented in Java to facilitateportability. Fig. 18 shows the prototype system architecture.

XML is used as a messaging format to communicatebetween components (sources, consumers, mediator, reposi-tory) in the distributed environment since it allows flexibledevelopment of user-defined document types. It provides arobust, nonproprietary, persistent, and verifiable file format


TABLE 2Sample Service Component Selection Rules

for the storage and transmission of text and data. In order foreach component to verify that the data it receives is valid, wedefine DTD’s (Document Type Definition) where the legalbuilding blocks of an XML document are specified. We collectserver and network related information in our prototypesystem. This information is obtained using Simple NetworkManagement Protocol (SNMP), which is readily available onmost servers and routers. Through SNMP, the mediator canobtain current status from the sources. In addition, traps canbe defined for sources so that sources can report seriousconditions to the mediator. AdventNet SNMP API is acomprehensive toolkit for rapid development of SNMP-based management applications, so we leverage it to monitorand track the performance of the sources. As illustrated, theSNMP messagesarewrapped inXML format to enable systeminteroperability and facilitate integration of non-SNMPprotocols. The status of information sources is stored in theinformation repository, which is implemented using Open-LDAP’s slapd LDAP (Lightweight Directory Access Protocol)server together with the Berkeley Sleepy Cat Database as thebackend. slapd is a stand-alone LDAP daemon that listens forLDAP connections on a port and responds to the LDAPoperations it receives over these connections. Fig. 19 showsthe LDAP schema used for our prototype.

A preliminary study was carried out to determine thecommunication latency between different system compo-nents. The source probing latency consists of the followingparts:

1. XML wrapping of probing request,2. transmission of the wrapped request from the

mediator to the source,3. unwrapping of probing request,

4. SNMP request,5. XML wrapping of the result,6. transmission of the wrapped result from the source

to the mediator, and7. unwrapping the result.

Table 3 indicates that, in our implementation, the timespent on XML parsing and SNMP calls is negligiblecompared to network delay. As the database is a vital partof the information collection framework, we conducted apreliminary study on the scalability of the database byincreasing the number of the sources in the system. Fig. 20indicates the variation in the time of database operations(addition, retrieval, and modification) as the number ofsources in the database increases. As can be seen, attribute


Fig. 17. The AutoSeC framework.

Fig. 18. AutoSeC prototype system architecture.

modification of a source in the database incurs significantlymore overhead than addition of a source or retrieval valuesof a source. This trend is fairly consistent.

Integrating AutoSeC into existing wireless infrastruc-tures. The proposed AutoSeC framework can be easilyintegrated into existing or next-generation wireless net-works. In particular, cellular handsets/mobile hosts andbase stations can be incorporated into the AutoSeCarchitecture as information sources to provide devicespecific parameters (by mobile hosts) and information onthe population of mobile hosts (by base stations). Typically,cellular networks have predefined access points to the wirednetwork (see GMSC in Fig. 21). Hence, all the informationrequests from AutoSeC are forwarded by the access point tospecific entity in the wireless network. We now illustrate atypical cellular network infrastructure (UMTS) and itsintegration into the AutoSeC data collection process.

UMTS [22] is one of the premium standards for cellularnetworks evolving from GSM. In this hierarchical system,coverage area is divided into a number of radio cells andeach cell has one base transceiver station (BTS). A mobilestation (MS) or a handset communicates with the BTS in itscell. A base station controller (BSC) manages several BTSs. Itreserves radio frequencies, handles the handover fromone BTS to another, and pages the MS when needed. SeveralBSCs in a geographical region are managed by a singleentity, the mobile services switching center (MSC). In addition,each MSC is associated with one home location register (HLR)and one visitor location register (VLR). HLR is a database thatstores all user relevant information, and VLR is a verydynamic database which stores all important informationneeded for the mobile stations currently in the region that isassociated to the MSC. MSCs are managed by a GatewayMSC (GMSC), which is the access point through which

external networks (e.g., wired networks, Wi-Fi networks)can integrate with the cellular infrastructure.

In the UMTS infrastructure, the GMSC serves as theconduit through which mobility related information isdelivered to the mediator. GMSC collects mobility relatedinformation from MSC and device-specific parameters fromMS. In order to manage cellular networks, SNMP agents aretypically deployed in MSC or MS. These SNMP agents formthe communication mechanism of the UMTS components(GMSC, MS) with AutoSeC.

The collected information can be utilized by bothtraditional and emerging algorithms. Consider, for exam-ple, the open-shortest-path-first (OSPF) routing protocol [5]currently used in the Internet. The routing is done in adynamic and adaptive fashion based on Dijkstra’s shortestpath algorithm. The term “shortest path” implies theoptimum path by considering multiple factors, such asgeographical distance, delay, throughput, or connectivity,etc. A router that detects a change in the networkimmediately multicasts the information to all other routersin the network so that all will have the same routing tableinformation. In addition, in the proposed QoS extensions toOSPF protocol [19], each switch maintains its own view ofthe network status, selects a path capable of meeting theQoS requirements of a given request, and establishes andmaintains the path by reserving the resources usingresource reservation protocol (RSVP). These examplesindicate that network connection parameters and band-width availability information are maintained at each routerand exchanged among routers either periodically or when achange in the network is detected. By maintaining thesystem status information in a common repository, we canimprove the scalability of the routing protocols in largescale distributed systems. Furthermore, our proposedarchitecture provides source-based routing algorithms aglobal view of system status. Standards, such as MPLS andRSVP, can be used together with the information in thedatabase to find a feasible path to the optimal server andreserve the resources along the path before actually routingthe request. This enables improved resource allocation andbetter QoS guarantees.

More recently, researchers have proposed numerousQoS-based resource allocation algorithms (in the areas ofreplicated service selection, QoS-based routing, load balan-cing services, and content distribution), which can benefitfrom the information collected. For example, dynamicserver selection policies have been studied in the context


Fig. 19. Schema diagram of AutoSeC.

TABLE 3Source Probing Latency

Fig. 20. Database access time.

of replicated Web data. One such effort [20] uses the notionof a Predicted Transfer Time (PTT) to reduce the responsetime. PTT is calculated using parameters such as availablebandwidth and round-trip delay. As another example [21],link parameters such as delay, available bandwidth, etc., isdescribed using probability distribution and used to findthe most probable path satisfying the requested bandwidthand end-to-end delay. The Combined Path and ServerSelection (CPSS) algorithm used in this paper utilizes theinformation from both server and network routers to assistthe selection of an optimal server and a path for betterresource utilization.

The development of AutoSeC framework and its inte-gration into existing wireless infrastructures has thepotential to improve system and network resource provi-sioning while ensuring application QoS.

6 RELATED WORK

In this section, we compare our work to related research innetwork status collection, data caching and QoS-basedresource provisioning.

System status collection for wired networks has been widelystudied in network management community. Existingcollection techniques may be classified as being passive oractive. Passive collection observes the existing traffic andinfers network performance from the collected measure-ments, e.g., Tcpdump, NetFlow. In contrast, active collectioninjects probes into the network and then derives networkperformance from the end-to-end measurements, e.g.,traceroute, netperf, netstat. Tools such as Pathchar [23], Packetpair technique [24], and Packet tailgating [25] measure linkand/or path capacity by inspecting the change in the latencyof the test packets sent out. Systems that use active probingfor the measurement of communication resources acrossInternet-wide networks include Network Weather Services(NWS) [26], topology-d [27], and Remos [28]. Similarly, thesnapshot-based collection policy (SS) has been used inrouting protocols such as OSPF [5] and QoS-based pathselection [29], where link state information (e.g., dynamicresidual bandwidth) is periodically exchanged among

routers. To decrease the overhead introduced by activeprobing, a straightforward strategy is to control the probingfrequency. Adaptively adjusting probing frequency has beenstudied using threshold or history measurements [30],application information [31], or bandwidth availability [32].An alternative strategy to reduce the overhead is to degradeinformation accuracy to a certain extent. For example, theStatic Interval-based policy (SI) [6] stores an interval ratherthan an instantaneous value for each data item. In addition toall the existing work in wired network monitoring, networkmonitoring for wireless networks has been focusing on mobiledevices’ location information management [10]. In thispaper, instead of focusing on mobile host location manage-ment, we are interested in the collection of system statusinformation where mobility status is accounted for improvedperformance. In addition, we studied the impact of SS and SIon the performance of QoS-based resource provisioning inmobile environments and compared the performance againstour proposed approaches.

Our information collection policies address the trade offbetween information accuracy and maintenance overhead,which bears similarity to adaptive data caching. In adaptivedata caching, one of the strategies is to use approximations toexact values in the cache in order to reduce cache main-tenance cost. Divergence caching considers setting theprecision of approximate values in a caching environment[33], where precision is inversely proportional to the numberof updates to the source value not reflected in the cachedapproximation, independent of the actual updates. TheTRAPP system automatically selects a combination of locallycached bounds and exact master data stored remotely todeliver a bounded answer to a query consisting of a rangethat is guaranteed to contain the precise answer [14], [34]. Inaddition, research on data streams [35] deals with contin-uous monitoring/querying of fast changing data in bothtraditional Internet environments and emerging ad hocsensor networks, etc. These projects include OpenCQ,Niagara, Telegraph, STREAM, and Aurora. Our work differsfrom the aforementioned research in the evaluation methodsof cache maintenance strategies. We use QoS-based resourceprovisioning as a motivating example and measure the


Fig. 21. Integration of AutoSeC and next-generation wireless networks UMTS.

effectiveness of proposed collection policies by comparingtheir impact on the performance of resource provisioning.

Network resource provisioning has been extensivelystudied in the context of QoS-based routing both in wirednetworks and mobile ad hoc networks. QoS routing is to finda network path to satisfy user quality requirements whileachieving system efficiency of resource utilization. Based onhow the network information is maintained and how thefeasible path is found, QoS routing can be classified intosource routing, distributed routing, and hierarchical routing[36]. Among a number of challenges faced by QoS routing,imprecise state information is one that is relevant to thispaper. Source routing algorithms are proposed [21], [37] todeal with imprecise state information. In their approaches,link parameters such as delay, available bandwidth, etc., canbe described using probability distributions and can be usedto find a most probable path satisfying the requestedbandwidth and end-to-end delay. In order to avoid theexpensive centralized computation at the source, a distrib-uted routing scheme [38] is proposed, which searchesmultiple paths in parallel for a satisfactory one. The stateinformation of intermediate nodes is collectively used toguide the routing messages along the most appropriate pathsin order to maximize the success probability. Further studies[3] show that the routing algorithm, network topology, andlink-state update policy have significant impacts on theprobability of successfully routing new requests, as well asthe overhead of distributing network load metrics. While theresearch mentioned above focuses on routing itself, CPSS(Combined Path and Server Selection) [16] is proposed toconsider a joint optimization in searching both a path and aserver for requests with QoS requirements. In mobileenvironments, the problem of imprecise state informationbecomes even more severe due to node mobility. A predictivelocation-based algorithm [39] is proposed to compute thefuture route in advance before existing routes break. In thispaper, we use QoS-based resource provisioning as aconsumer of the information provided by the collectionservices. We evaluate the impact of information collectionpolicies on QoS-based resource provisioning under differentsystem conditions and application workload.

7 CONCLUSIONS

A key concern of our work is to preserve the desired QoS ofapplications in mobile environments without introducingsignificant overhead. In this paper, we have indicated thatinstantaneous representation of information obtained viafrequent periodic sampling incurs significant overhead andis not necessary; on the contrary, adaptive dynamic range-based information collection mechanisms exhibit superiorperformance under most system conditions and user work-loads. We also observed that resource provisioning processbenefits significantly from the information collection me-chanisms that utilize the mobile user aggregation status. Wefurther showed that coarse-grained mobility information(i.e., aggregate user mobility status) is sufficient for effectiveresource provisioning, while fine-grained mobility informa-tion (i.e., individual user’s location information) introducesvery high overhead and is hence undesirable. The integratedmiddleware framework AutoSeC presented in this paper is aresult of incorporating the insights gained from studying theimpact of various information collection policies on theperformance of a family of resource provisioning mechan-isms. In general, the dynamic nature of applications such as

those of multimedia under varying network conditions,request traffic, etc., implies that information collectionpolicies must be dynamic and customizable. Therefore,intelligent policy composition mechanisms, such as thoseexplored in AutoSeC, must be developed to achieve effectiveutilization of resources while ensuring user QoS.

Adaptive information collection policies must not onlyaccommodate statistical fluctuations in network and serverloads and handle changes in user mobility status, but alsodeal with a variety of application workloads. The distrib-uted object paradigm is often used to facilitate the develop-ment of large scale distributed applications. However, thetraditional object messaging layer operates with limitedawareness of underlying system/network conditions,whereas current system/network monitoring tools operateat the network layer with little awareness of applicationlevel object communication requirements. We have exploredthe possibility, mechanisms, and benefits of filling the gapbetween object messaging and system monitoring. Weintroduced the connection abstraction as the mechanismfor these two layers to communicate and exchange informa-tion. Through this integration, object messaging can proac-tively adapt to changing system conditions; systemmonitoring policies and parameters can be optimized basedon interobject communication properties [40].

In this paper, a specific information consumer—QoS-based resource provisioning is used to evaluate theinformation collection policies presented here. In fact,information consumers can be extended to a large scope ofcontext-aware applications and our eventual goal is todevelop a fundamental middleware service for contextinformation collection. To achieve true context awareness,systems must produce reliable and real-time information inthe presence of uncertain, rapidly changing, and partiallytrue data from multiple heterogeneous sources. With theemergence of sensor networks, a large amount of contextinformation is captured by sensors. Distributed sensorenvironments are very different from traditional mobileenvironments: Sensor devices are expected to interact withthe physical world by allowing continuous monitoring andreaction to natural processes; hence, real-time communica-tion is crucial. In addition, sensor nodes are energyconstrained and sensor networks are subject to high faultrates. The concept of QoS for distributed sensor applicationsdiffers from the QoS concept for multimedia applicationsand can be defined by information delivery latency,information accuracy, and energy efficiency. Therefore,context information collection for distributed sensor envir-onments should address timeliness/accuracy/energy con-sumption trade offs while ensuring this particular QoS.Middleware techniques for adaptive context informationcollection, such as those described in this paper, are key toguaranteeing application QoS in highly dynamic hetero-geneous environments. This is a necessity to achieve the goalof true ubiquitous computing.

ACKNOWLEDGMENTS

The authors would like to thank Zhenghua Fu for usefuldiscussions that led to the framework described here andhis contributions to designing the simulation environments.This research was supported by the US Office of NavalResearch under MURI Research Contract N00014-02-1-0715and the US National Science Foundation under AwardNumbers 0331707 and 0331690. Any opinions, findings, and


conclusions or recommendations expressed in this paper

are those of the authors and do not necessarily reflect the

views of the research sponsors.

REFERENCES

[1] A. Acampora and M. Naghshineh, “Control and Quality-of-Service Provisioning in High-Speed Microcellular Networks,”IEEE Personal Comm. Magazine, vol. 1, no. 2, 1994.

[2] S.K. Das, R. Jayaram, and S.K. Sen, “An Optimistic Quality-of-Service Provisioning Scheme for Cellular Networks,” Proc. Int’lConf. Distributed Computing Systems, 1997.

[3] A. Shaikh, J. Rexford, and K.G. Shin, “Evaluating the Impact ofStale Link State on Quality-of-Service Routing,” IEEE/ACM Trans.Networking, vol. 9, no. 2, Apr. 2001.

[4] Z. Fu and N. Venkatasubramanian, “Combined Path and ServerSelection in Dynamic Multimedia Environments,” ACM Multi-media, 1999.

[5] “OSPF Version 2,” RFC 1247, July 1991.[6] G. Apostolopoulos, R. Guerin, S. Kamat, and S. Tripathi, “Quality

of Service Based Routing: A Performance Perspective,” Proc. ACMSIGCOMM, 1998.

[7] Z. Fu and N. Venkatasubramanian, “Adaptive Parameter Collec-tion in Dynamic Distributed Environments,” Proc. Int’l Conf.Distributed Computing Systems, 2001.

[8] Q. Han and N. Venkatasubramanian, “AutoSeC: An IntegratedMiddleware Framework for Dynamic Service Brokering,” IEEEDistributed Systems Online, vol. 2, no. 7, 2001.

[9] I. Akyildiz, J. McNair, J. Ho, H. Uzunalioglu, and W. Wang,“Mobility Management in Next-Generation Wireless Systems,”IEEE Proc. J., vol. 87, no. 8, Aug. 1999.

[10] V.W.-S. Wong and V.C. Leung, “Location Management for Next-Generation Personal Communications Networks,” IEEE Network,Sept./Oct. 2000.

[11] Z. Haas, “A New Routing Protocol for the ReconfigurableWireless Networks,” Proc. IEEE Int’l Conf. Universal PersonalComm., Oct. 1997.

[12] Q. Han and N. Venkatasubramanian, “Aggregation BasedInformation Collection for Mobile Environments,” J. High SpeedNetworks, vol. 11, nos. 3-4, 2002.

[13] Q. Han and N. Venkatasubramanian, “A Cost Driven Approach toInformation Collection for Mobile Environments,” Proc. IEEEMWCN, 2002.

[14] C. Olston, B.T. Loo, and J. Widom, “Adaptive Precision Setting forCached Approximate Values,” Proc. ACM SIGMOD, 2001.

[15] N. Venkatasubramanian and S. Ramanathan, “Load Managementfor Distributed Video Servers,” Proc. Int’l Conf. DistributedComputing Systems, May 1997.

[16] Z. Fu and N. Venkatasubramanian, “Directory Based CompositeRouting and Scheduling Policies for Dynamic MultimediaEnvironments,” Proc. IEEE Int’l Parallel & Distributed ProcessingSymp., 2001.

[17] A. Bar-Noy, I. Kessler, and M. Sidi, “Mobile Users: To Update orNot to Update?” ACM/Baltzer J. Wireless Networks, vol. 1, no. 2,July 1995.

[18] L. Bottomley, “EPA-HTTP Single WWW Server Rrace,” http://ita.ee.lbl.gov/html/contrib/EPA-HTTP.html, Nov. 2003.

[19] “QoS Routing Mechanisms and OSPF Extensions,” RFC 2676, Jan.1998.

[20] R. Carter and M.E. Crovella, “Dynamic Server Selection UsingBandwidth Probing in Wide-Area Networks,” Proc. IEEE InfoCom,1997.

[21] D. Lorenz and A. Orda, “QoS Routing in Networks with UncertainParameters,” Proc. IEEE InfoCom, 1998.

[22] J. Schiller, Mobile Communications, chapter 4. Addison-Wesley,2000.

[23] V. Jacobson, “Pathchar—A Tool to Infer Characteristics of InternetPaths,” ftp://ftp.ee.lbl.gov/pathchar/, 1997.

[24] S. Keshav, “Measuring Link Bandwidths Using a DeterministicModel of Packet Delay,” Proc. ACM SIGCOMM, 1991.

[25] K. Lai and M. Baker, “Measuring Link Bandwidths Using aDeterministic Model of Packet Delay,” Proc. ACM SIGCOMM,2000.

[26] R. Wolski, N. Spring, and C. Peterson, “Implementing a Perfor-mance Forecasting System for Metacomputing: The NetworkWeather Service,” Proc. IEEE/ACM Supercomputing Conf., 1997.

[27] K. Obraczka and G. Gheorghiu, “The Performance of a Service forNetwork-Aware Applications,” Proc. SIGMETRICS Symp. Paralleland Distributed Tools, 1998.

[28] T. Dewitt, T. Gross, B. Lowekamp, N. Miller, P. Steenkiste, and J.Subholk, “ReMoS: A Resource Monitoring System for NetworkAware Applications,” Technical Report CMU-CS-97-194, CarnegieMellon Univ., 1997.

[29] Q. Ma and P. Steenkiste, “On Path Selection for Traffic with Band-width Guarantees,” Proc. IEEE Int’l Conf. Network Protocols, 1997.

[30] P. Dini, G.V. Bochmann, T. Koch, and B. Kramer, “Agent BasedManagement of Distributed Systems with Variable PollingFrequency Policies,” Proc. IEEE/IFIP Int’l Symp. Integrated NetworkManagement, 1997.

[31] P. Moghe and M. Evangelista, “An Adaptive Polling Algorithm,”Proc. IEEE/IFIP Network Operations and Management Symp., 1998.

[32] K. Yoshihara, K. Sugiyama, H. Horiuchi, and S. Obana, “DynamicPolling Scheme Based on Time Varation of Network ManagementInformation Values,” Proc. IEEE/IFIP Int’l Symp. Integrated NetworkManagement, 1999.

[33] Y. Huang, R. Sloan, and O. Wolfson, “Divergence Caching inClient-Server Architectures,” Proc. IEEE Third Int’l Conf. Paralleland Distributed Information Systems, 1994.

[34] C. Olston and J. Widom, “Offering a Precision-PerformanceTradeoff for Aggregation Queries over Replicated Data,” Proc.Very Large Data Bases Conf., 2000.

[35] M. Garofalakis, J. Gehrke, and R. Rastogi, “Querying and MiningData Streams: You Only Get One Look,” Proc. Very Large DataBases Conf., 2002.

[36] S. Chen and K. Nahrstedt, “An Overview of Quality-of-ServiceRouting for the Next Generation High-Speed Networks: Problemsand Solutions,” IEEE Network Magazine, vol. 12, no. 6, Nov.-Dec.1998.

[37] R. Guerin and A. Orda, “QoS-Based Routing in Networks withInaccurate Information: Theory and Algorithms,” Proc. IEEEInfoCom, 1997.

[38] S. Chen and K. Nahrstedt, “Distributed QoS Routing withImprecise State Information,” Proc. IEEE Int’l Conf. ComputerComm. and Networks, 1998.

[39] S.H. Shah and K. Nahrstedt, “Predictive Location-Based QoSRouting in Mobile Ad Hoc Networks,” Proc. IEEE Int’l Conf.Comm., 2002.

[40] Q. Han, S. Gutierrez-Nolasco, and N. Venkatasubramanian,“Reflective Middleware for Integrating Network Monitoring withAdaptive Object Messaging,” IEEE Network, Jan. 2004.

Qi Han received the MS degree in computerscience from the Huazhong University ofScience and Technology, China, and the PhDdegree in computer science from the Universityof California, Irvine. She is an assistant profes-sor at the Colorado School of Mines. Herresearch interests include distributed systemsmiddleware, mobile and pervasive computing,and systems support for sensor applications.She is currently developing adaptive middleware

techniques for collecting various dynamic context data in heterogeneousenvironments to support context aware applications. She is a member ofthe IEEE and ACM.

Nalini Venkatasubramanian received the MSand PhD degrees in computer science from theUniversity of Illinois, Urbana-Champaign. She isan associate professor in the School of Informa-tion and Computer Science, University of Cali-fornia, Irvine. Her research interests includedistributed and parallel systems, middleware,mobile environments, multimedia systems/appli-cations and formal reasoning of distributedsystems. She is specifically interested in devel-

oping safe and flexible middleware technology for highly dynamicenvironments. She was a member of technical staff at Hewlett-PackardLaboratories in Palo Alto, California, for several years, where sheworked on large scale distributed systems and interactive multimediaapplications. She has also worked on various database managementsystems and on programming languages/compilers for high-perfor-mance machines. She is a member of the IEEE and ACM.


518 ieee transactions on mobile computing, vol. 5, no. 5 ...dsm/pubs/01610594.pdf · information...

Documents