Adaptive QoS Management for Collaboration over Heterogeneous Networks

Prathima Agrawal1, Pravin Bhandarkar2 and Manish Parashar2

Telecordia Technologies,445 South Street, NJ 07960-38

Telephone: (973)829-3158; Email: pagrawal@research.telecordia.com2

Department of Electrical and Computer Engineering & CAIP, Rutgers University,

94 Brett Road, Piscataway, NJ 088

Telephone: (732) 445-5388; Fax: (732) 445-0593; Email: {parashar,pravinb}@caip.rutgers.edu1

Abstract – A key requirement for enabling collaborationbetween distributed (wired and wireless) clients in a networkedenvironment, with heterogeneous and dynamic capabilities, is anunderlying information and QoS management protocol that canadapt on-the-fly based on the local system and client state, so as

to maintain effective information sharing. This paper presentsthe design, implementation and evaluation of such an adaptiveQoS management framework. The framework enablesindividual clients to locally interpret and transform informationto match local resources, client interests and system/networkstate. The framework is built on an innovative semanticinteraction mechanism that provides each client with direct andimmediate access to all information defined by its needs andcapabilities, without having to maintain and update globalrosters. It interfaces with the Simple Network ManagementProtocol (SNMP) to determine the state of the network bydirectly querying network elements.

Index Terms: Adaptive QoS management, Heterogeneousdistributed collaboration, SNMP, Information coordiation.

I. INTRODUCTION

Collaboration can be defined as interaction andinformation interchange between people working atphysically disparate locations working in dynamicheterogeneous environments with the purpose ofaccomplishing mutually beneficial activity. The ubiquityof network connectivity and recent advances in computingand networking technology provide the potential ofenabling collaboration and computer-mediatedinformation sharing and decision making in many criticalapplication domains including Medical Telediagnosis,Crisis Management, and Electronic Trading. Forexample, consider an interaction between paramedicsrushing a patient from a distant location in an ambulance,and collaborating with doctors at the hospital on thepatient's EKG. In this scenario, the two locations arephysically separated and the paramedics are continuouslyin motion. Furthermore, the paramedics typically have alower processing capability, lower bandwidth, and a lossywireless connection, while the doctors at the hospital haverelatively high end processing capability; i.e. there isheterogeneity in the collaboration environment. Finally, aspecialist's opinion may be needed, requiring her to jointhe ongoing session (possibly from her home); i.e. thecollaboration environment can be dynamic.

The essential requirement for enabling the abovecollaboration scenario is providing each client with theability to have direct and immediate access to allinformation defined by its needs, interests, resources andcapabilities. Furthermore, client and networkheterogeneity require the ability to locally transforminformation on-the-fly so that it matches the client’s localcapabilities and resources, while maintaining the semanticcontent of the information for effective sharing.Consequently, supporting collaboration betweendistributed clients in a networked environment, withheterogeneous and dynamic capabilities requires anadaptive framework for QoS management framework thatcan react to changing client and system/network statesand modify the local information processing. Theobjective of this paper is to present the design,implementation and evaluation of such a framework forenabling seamless collaboration in distributed wired andwireless environments. The framework is founded on aninnovative semantic information coordination model thatapplies the “pull” knowledge management model todistributed information management. It uses an extensionagent based on the Simple Network ManagementProtocol (SNMP) to determine the state of the network bydirectly querying network elements.

The rest of this paper is organized as follows. SectionII outlines a typical architecture of a network-basedcollaboration framework and presents related research inthis area. Section III presents the semantic interaction andinformation coordination protocol that forms the basis ofour collaboration framework. Section IV describes thearchitecture and operation of the adaptive QoSframework. Section V presents an experimentalevaluation of the framework and demonstrates its abilityto adapt to changing systems and client states. Section VIpresents our conclusions and outlines future work.

II. ARCHITECTURAL OVERVIEW OF ACOLLABORATION FRAMEWORK

User InterfaceConcurrencyInformationEventControlTransformerReplicationMiddlewareTransport protocolNetwork FabricFigure 1 - Typical Collaboration Framework

Figure 1 shows a block diagram of a typicalcollaboration framework. The overall architecture can beeither centralized or distributed. Centralized server basedcollaboration architectures, where session management isperformed by a single central server, provide tightlycontrolled interactions. In these architectures, it is easierto manage event concurrency and maintain eventsordering. However these architectures are not scalableand cannot readily adapt to changing client interests andcapabilities. The distributed peer based collaborationarchitectures have loosely coupled interactions, and aremore scalable, but need special synchronizationmanagement, concurrency management, and eventordering.

The user interface of a collaboration framework is ametaphor for the global virtual space through which all

the clients of a collaborative session interact.

The middleware layer is the most important layer ofthe framework. Key issues addressed by this layerinclude:

• Event Replication: Event replication is the process

of efficiently and transparently transmitting eventsgenerated by one client’s action to all other clients inthe collaboration session, and reproducing theoriginal action on the remote clients.

• Concurrency Control: Concurrency Control is the

process of arbitration and consistency maintenancewhen multiple client's concurrently manipulate thesame set of shared objects in the collaborationsession.

• Network Adaptation: To enable heterogeneity

management it is essential to monitor the state of thenetwork and adapt applications to varying networkconditions. Simple Network Management Protocol

(SNMP) can be used to determine current networkconditions by directly interacting with the networkelements [21].

• Heterogeneity Management: Heterogeneity

management is the process of enabling effectiveinteraction and information sharing among a group ofclients with different resources and capabilities.Client interests and capabilities can also be dynamic.For example, in a wireless network, the networkcapability may change rapidly due to error rate orload on a wireless link.

The transport layer mechanisms form thecommunication architecture in a collaboration session.The underlying communication can leverage the Internetby the use of the Internet Protocol (IP) to exchangeinformation. High volume data to be sent to a group ofusers poses a unique problem for collaboration.Transport layer protocols such as TCP can be used forsystems where the communication is point-to-pointbetween a small number of clients. TCP can ensure highreliability of data delivery but is not scalable. UDP basedmulticast schemes are more scalable but do not guaranteereliability. Protocols like Real-Time Transport Protocol(RTP) [10] and Real Time Control Protocol (RTCP)together can enable transfer of real-time multimedia data.RTP is used to identify the synchronization source,transfer media data and sequencing support andsynchronizing. RTCP identifies the participant, gets thedata content, quality of service information, and requestsretransmission.A. Related Work

Meeting Quality of Service (QoS) guarantees indistributed and heterogeneous systems are an end-to-endissue, and entail interaction between the various softwaremodules to define “contracts” to provide QoS. Thisrequires mapping QoS parameters from the uppermost tothe lowermost layer for correct interpretation of userrequests. There are a number of research projects tryingto address QoS for distributed multimedia systems.Primary areas of research include specification of QoSparameters, level of service based on contracts betweenuser and network, soft state versus hard state, and QoSmapping at the various layers of software to manageheterogeneous demands [3]. This section outlines thearchitectures of important research frameworks forenabling collaboration. Note that none of theseframeworks provide adaptive QoS capabilities to addressheterogeneity and dynamism in the clients interests and/orsystem resources and capabilities.

The Habanero [9] framework developed at NCSA,UIUC, uses a combination of a “central arbitrator” and“central router” to replicate events amongst collaborating

clients. Application events are trapped by the “distributedarbitrator”, who marshals them and passes them to thecentral arbitrator. The central arbitrator then uses thecentral router to multicast them to other clients. Allevents are managed, replicated and communicatedthrough a centralized server that houses the centralarbitrator and central router. The distributed

collaborative framework at University of North Carolina [12] uses an approach based on bus-agents. In thisframework a central server (a Java remote object),located at a well-known resource within a given Internetdomain, maintains a central registry and allows users andother bus agents to obtain remote references to active busagents. Bus agents wishing to allow initiation ofcommunication by another bus agent or process, mustregister with the central registry. Info-Bus technology [1]provides a push-pull model for the transfer of events, butis limited to a single machine and does not supportdistributed sessions. The Tango web collaboratoryproject at NPAC, Syracuse University [8] leverages Javatechnology to enable collaborative applets to be obtainedvery easily using web browsers. The JETS project [20]also uses applet-based collaboration. Clients initiate asession by downloading Java applets from a centralizedserver. The server maintains a repository of applicationobjects, and replicates various client events using Javaobject serialization. The server also keeps track of all theclients in a session and can relay information about usersjoining or leaving a session. The communication betweenthe entities is performed using multicasts. TheInfospheres project at CalTech [15] aims at building aframework composed of two units: dapplets which aremultithreaded, communicating objects, and sessionswhich are composed of groups of dapplets. A dapplethandles a message by possibly changing its state andsending messages on its output port. Dapplets arecomposed together, in parallel, to form a distributedsession.

III. SEMANTIC INFORMATION MANAGEMENT FOR

HETEROGENEOUS COLLABORATION

A key requirement for real-time systems forcollaboration and knowledge sharing for distributed(multimedia) applications (collaborative medicine,strategic battle planning and aerospace simulations) is thedefinition of an effective and efficient model forinformation coordination and replication in real-time. Afurther requirement for heterogeneous clients is the abilityto locally interpret the events to reflect the interests,capabilities and resources of each client. The localcapability and local resource availability can bedetermined using different network centric approaches.Traditional distributed information management

approaches are based on global naming services, whereall communications use unique names assigned to clients.In such a system, every application client that enters asession must register itself with the naming server,explicitly stating its interests. The server then assignscapabilities to the entering clients and informs existingclients about the new client’s interests. Existing clientscan now forward relevant information from the existingcollaboration session to the entering client. Clearly, thedynamics of such a collaborative framework is limited bythe rate at which the network can synchronize distributingnames, interests and capabilities.

Figure 2 - Semantic Interaction Model

Semantic information management (SIM) [18] is anew approach for information coordination andreplication to support real-time collaboration amongstheterogeneous, distributed and dynamic applicationclients. This approach implements the “pull” distributedinteraction model using semantically enhanced events andstate-based [2] communication techniques. In thisscheme, each client maintains a profile that defines itscurrent state, its interests and its capabilities. Allinteractions in this scheme are then addressed to profilesrather than explicit names. Consequently, the group ofinteracting clients is determined only at run-time. Aclient’s profile may also encapsulate network/systemstate. In this formulation, clients can freely join or leavesessions, and are liberated from static sites and complextracking protocols, allowing them to migrate freely andutilize available resources.

State-based interaction techniques are the applicationof semantic content-based resolution techniques, used bythe naming service, applied directly to the run-timeinteraction between clients. The overall approach issummarized in Figure 2. Each client in a collaborationsession will locally export a “profile”. A client’s profileis a mutable set of attributes that specify its type,state,interests, current network state and capabilities. Profilesare maintained and modified by clients to reflect theircurrent interests. All communications between thecollaborating clients are now defined as state-basedmulticast messages where a message is semanticallyenhanced to include a sender-specified “semantic-selector” in addition to the message body. The semantic-selector is a prepositional expression over all possibleattributes and specifies the profile(s) of clients that areto receive the message. Thus the conventional notion ofa static client or client group name is subsumed by theselector which descriptively names dynamic sets of clientsof arbitrary cardinality (conventional names of clients orclients groups are non-descriptive and statically bound.).State-based messages are received by semanticallyinterpreting message selectors in terms of the clientprofiles.

Figure 3 - Semantic Interpretation

Figure 3 illustrates the semantic interpretation process.The semantic selector describes the attributes of theincoming stream as color video, with MPEG2compression and 1 MB data. Client 1's profile (Profile 1)matches this incoming selector and hence the message isaccepted. Client 2 (Profile 2) on the other hand is onlyinterested in B/W video with no encoding and so themessage is rejected. Client 3 ( Profile 3) is interested incolor video with JPEG encoding and has the capability totransform MPEG2 to JPEG. It thus accepts the messagewith a transformation.

Heterogeneity management in collaboration enablesshared information to be intelligently transformed so thatit matches the client’s local capabilities and resources,and yet maintains semantic contents for effective sharing.SIM manages client heterogeneity by information

reduction and transformation, using informationtransformation modules. For example, a client with ahearing disability can require speech information to betransformed into text. This can be achieved byappropriately setting the client’s profile expression.Similarly, a mobile client collaborating on a video streammay to conserve battery power[19] by adapting theamount of video information received. Clients controltheir own profiles and can add, delete or modify attributesin the profile at any time. The modification of a profile isan atomic action with respect to its matching with aselector. The only global knowledge upon which such aninteraction depends is the program-specific meanings ofthe attributes. This implies that all clients haveinformation about possible application objects that canparticipate in a session and their states; informationinherently known to all clients

IV. AN ADAPTIVE QOS FRAMEWORK

HETEROGENEOUS DISTRIBUTED COLLABORATION

Figure 4 – Adaptive QoS Framework for Collaboration

Figure 4 presents a schematic overview of the adaptiveQoS framework for heterogeneous distributedcollaboration. The essential components of thisarchitecture are the user and application interface, theinteraction protocol, the adaptive network managementmodule, and the multicast communication substrate.These components are described below.A. Multicast Communication Substrate

The communication substrate consists of a pair ofmulticast channels: the first is used for control messageswhile the second is used for large data (image) messages.Maintaining two separate channels prevents control

messages from being blocked by larger data messages.The multicast channel pair is created when a newcollaboration session is created. Clients logging on to anongoing session subscribe the existing multicast channels.Note that multicast communications use UDP packets fortransmission and do not provide any reliability [7][22].Consequently, we built an additional layer on top of thecommunication substrate to ensure reliable and orderedtransmission of information packets. Control messagesare typically very small and can be transmitted in a single(1 Kbyte) packet. These messages are delivered withminimum or no losses. Data messages containinginformation such as images, are typically high volumeand require transmission of many packets. Reliable andordered delivery of these packets is critical for successfulreplication of the data at a collaborating remote client.We use a simple scheme to achieve this. Each outgoingpacket is tagged with a sequence number beforetransmission on the multicast channel. The sequencenumbers of incoming packets are checked at the remotesite. This enables us to detect out-of-order arrival ofpackets or packet drop owing to congestion. In thecurrent framework, high-volume data traffic primarilyconsists of images that are using our collaborative imageviewer. As these images are segmented beforetransmission, we drop out-of-order packets and use thegood packets to construct the image (possibly at a lowerresolution). This allows us to avoid elaborate flow controlschemes requiring retransmission and the associatedoverheads, which is critical as late arrival can render thedata meaningless.

B. Inference Engine, System State Interface and ClientProfile

The ability to adapt the quality of service based onnetwork and system state is jointly provided by threecomponents, viz. the client profile, the system stateinterface and the inference engine. The inference engineserves as a policy database and encodes policies forinformation transformations. It uses the client profilegenerated by current client and system/network state. Theformer encodes the client’s interests, preferences andcapabilities, while the latter encodes available resourcesand their current state. The profile is dynamic andchanges locally to reflect changes in the client or systemstate. Users can specify individual system and applicationparameters that will make up the local system state, aswell as constraints on these parameters. These user-entered rules define a QoS “contract” that needs to besatisfied by the inference engine. The inference engineuses its policy database to locally interpret incoming“semantic selectors”. Interpretation determines whetherthe incoming information is of interest to the client or not,

whether the client is capable of receiving the information,and what transformations (if any) need to be applied tothe information to satisfy the client’s profile. Forexample, the resolution of an incoming image may bereduced to match the client’s resources or the availablenetwork bandwidth. Similarly, incoming textualinformation can be transformed into speech if the profilespecifies that the client has chosen speech as the preferredmodality.

C. Communication Module

The communication module is responsible for (a)associatively multicasting messages on thecommunication media, and (b) interpreting incomingmessages, corresponding to remote events, for relevanceand translating them into local events. The sendercomponent multicasts semantic messages over thecommunication media. Events are thus transmittedasynchronously to all the other interested clients, withoutany knowledge of the current membership of this group,or the interests or capabilities of other clients. Thereceiver component monitors incoming messages andextracts their semantic selectors, which are used tolocally process the message.D. Information Transformer

The information transformer component maintains asuite of media-specific information abstraction modules.Information abstraction aims at intelligently transforminginformation while maintaining its semantic content.Examples of information transformers include image-to-text, image-to-speech, text-to-speech, and speech-to-textconversions. Such a translation is critical for enablingcollaboration across heterogeneous clients andinterconnects with large variations in capabilities. Theinformation transformer library is designed to beextendible so that new modules and media types can beeasily incorporated. Our current implementation uses animage transformation module capable of progressivedescription of visual information [6][11]. The moduleuses robust segmentation of the image to extract arealistic sketch of the main features. This sketchpreserves the essential information required for effectivecollaboration, and requires up to 2000 times lesser datathan the original. Furthermore, a verbal description canbe tagged to this sketch and can be used to enable clientswith minimal capabilities (e.g. a client on a wirelessconnection) to be effective participants in a collaborationsession. Image detail is hierarchically added to the sketch.This hierarchical representation of the visual informationalong with a verbal sketch thus facilitates thecollaboration among clients with heterogeneouscapabilities. Each of the users may access the samevisual information but at different resolutions or using

different modalities. In the framework, the inferenceengine uses the profile to locally decide on an acceptableresolution for the incoming data that meets therequirements and constraints specified in the QoScontract. The resolution threshold is used to determine thenumber of image segments (i.e. the number of imagepackets) to be received.E. System State Interface

The system state interface is a generic component thatencapsulates the state of the system. This includes CPUload, memory available, network bandwidth, latency andjitter. The network state interface is a specificimplementation of this generic component. Thesystem/network state interface is implemented using theSimple Network Management Protocol (SNMP) [21],which enables it to determine the state of networkelements and hosts. It uses the IP address of the networkelement, the community string, and the object identifier(OID) of the parameters of interest (bandwidth, CPUload, page-faults, etc.) to directly query the SNMP MIB(management information base). In our Windows NTbased implementation we use Windows NT SNMP [13].The system/network state interface consists of twocomponents: the manager component that runs on themanagement station, and the agent component that runson the network element to be monitored. Routers andswitches have standard agents to monitor the localparameters through instrumentation routines. To monitorNT hosts, we built a specialized extension agent that runson each host and continuously obtains management data.This data is then forwarded to management stations uponrequest.

1) SNMP manager

The manager is an SNMP-based network managementapplication running under the Microsoft Windows NT4,using WinSNMP API [23].

2) Extension Agent using NT SNMP

Managed devices such as hosts and routers containmonitoring and possibly control instrumentation. The NTextension agent represents the managers access to thisinstrumentation (via a MIB), filtered by the SNMPsecurity mechanisms. The manager communicates withthe NT agent via SNMP to monitor and control managedhosts. The agent is based on the Microsoft SNMPExtension API which provides the basic functionality forcommunicating with the SNMP service[5].F. User and Applications Interface

A clients local user interface (UI) represents a globalvirtual space where all the clients in the collaborationsession interact. Locally generated events aretransparently transformed into event in the virtual spaceand are visible to all clients interacting in this space.

Similarly, remote events are transparently made visible tothe local client. Our current implementation of the UI hasthree modules, viz., chat area, image viewer andwhiteboard, as shown in Figure 5.

The chat area is built using Java AWT and itskeyListener interface. Keyboard events in the chat areaare intercepted by the UI, and on construction of asentence (marked by a carriage return), the sentence ismulticast using the control multicast channel.Transmitting sentences rather than individual charactersreduces the amount of communication and lowersprocessing overhead. Furthermore, corrections can bemade locally and do not have to be transmitted. Thewhiteboard implements a collaborative workspace inwhich users interact by writing or drawing. Ourimplementation of the whiteboard uses the semanticinteraction model to enable selective filtering events basedon interests. For example, if a user is currently interestedin a particular type (color, line type, etc.) of drawings,only those whiteboard events that are associated with thetype of interest are processed; all other whiteboard eventsare automatically filtered out at the communicationinterface. The image viewer module enables users tocollaboratively view images. When the local client loadsan image, the image data is encoded and multicast toother users in the collaboration session. The image vieweruses progressive transmission using context based imagecompression.

Chat Area

Image Viewer

WhiteBoard

Figure 5 - Collaborative User Interface

The UI is coupled to the adaptive framework using theapplication interface. This component is responsible forlocally orchestrating an application client’s collaborationsession. It monitors all local objects of interest to theclient and encodes their state as entries in the client's state

repository. Similarly, when a remote instance of theobject changes state, the change is received by thecommunication module and forwarded to the applicationinterface, which in turn updates the client's session. Theimplementation of the user and application interface isJava based and enables users to leverage the elaborateJava delegation-event model [17].G. Framework Operation

User InterfaceInformation TransformerClient State + System StateInference EngineSystem State Interface (Manager)NT SNMP Extension AgentRemote Network Element AgentFigure 6 - Components for Adaptive QoS Management

The QoS framework uses policies and system/clientstate information to adapt information processing toheterogeneous and dynamic environments. Figure 6shows the interaction between the various components ofthe adaptive QoS management framework. The user setsthe local client preferences, which are encoded in theclient state. The user indicates network parameters thatdefine system state. Possible parameters include routerbandwidth, CPU load, page faults and the battery levelfor mobile system. These parameters are monitored bythe WinSNMP manager by querying local and remotenetwork agents. The client and system state together formthe local profile. The inference engine maintains policiesfor adaptation and information interpretation. Dependingon the current profile and the user encoded policies itdecides on the information transformation to be applied tothe incoming data. A typical sequence of events is shownin Figure 7 and described below.

1. The user logs on and subscribes to the data and

control multicast channels.

2. User interests are used to construct the client state.3. The user indicates the system/network parameters

that will define the system state.

4. Client profile is constructed as a combination of the

client and system state, and is updated whenever userinterests and/or system parameters change.

5. The system state manager obtains the network

element to be queried.

6. The communication module receives the incoming

event and notifies the manager.

7. The manager queries the network element and obtains

the SNMP parameter from the agent.

8. Client profile and the current condition of the

network element of interest is forwarded to theinference engine

9. Inference engine decides on the information

transformer to be applied.

User logon123Information TransformerClient StateSystem State94Client Profile8InferenceEngine5Communication6 ModuleSystem State Interface8Agent Component7Figure 7 - Sequence of Events

V. EXPERIMENTAL EVALUATION OF THE SIM

ARCHITECTURE

To demonstrate the adaptability of the collaborationframework with dynamically changing system conditions,SNMP parameters on the host and the image viewerperformance were monitored. The experimental set upconsisted of three NT work stations. The objective of theexperiment was to determine the behavior of the imageviewer software by plotting the compression ratios, thenumber of packets received and the quality of image (bitsper pixel) versus the SNMP parameters obtained from thelocal hosts. The two SNMP parameters monitored werethe number of page faults and the CPU load.

A. The Image Viewer Parameters versus the PageFaults

Figure 8 shows the adaptability of the framework tochanging conditions on the host. These measurementsindicate that as the number of the page faults increasesthe number of image packets that can be processed at thelocal host decreases. The numbers of packets vary from 1to 16 in powers of 2 corresponding to page faults varyingfrom 30 to 100. The exact number of packets isdynamically determined by the inference engine based oncurrent system state. It can be seen that as the number ofpackets decreases (due to increased page faults) thecompression ratio increases indicating that lesser data isavailable. The bits per pixel (BPP) is an indicator of thequality of the image, this parameter decreases as thenumber of packets decrease. The image viewer

component adapts to provide images of varying BPP

ranging between 2.1 to 0.1 and compression ratios from

3.6 to 131.

No of Packets vs Page Faultsstek20ca1516P 10fo 58No of Packets.o0421N30507090100Page FaultsCompression Ratio vs Page FaultsetaR150 n131oi100sse65.5CompressionRatiorp50moC03.65.416.330507090100Page FaultsBits per pixels vs Page Faultsle3xiP r22.11.9eP1.4Bits per pixel s1tiB00.30.130507090100Page Faults Figure 8 - Image Viewer Parameters versus Page Faults

A. The Image Viewer Parameters versus the CPU LoadFigure 9 shows the BPP, compression ratios of theimages and the number of packets versus the CPU Load.The number of packets and the BPP decreases as theCPU load increases and the compression ratiocorrespondingly increases. The CPU load varies from 30to 100 percents results in a drop in the number of imagepackets accepted from 16 to zero. The exact number ofpackets is again dynamically determined by the inferenceengine based on current system state. The correspondingvariation in the BPP is between 14.3 and 0.7 and thecompression ratio vary between 1.6 and 32.7.

From the above two experiments it can be clearly seethat it possible to obtain a wide range of compressionratios and quality of images to adapt different network/system state parameters.

VI. SUMMARY AND CONCLUSIONS

This paper presented the design, implementation andevaluation of an adaptive QoS management framework tosupport collaboration among distributed heterogeneous

(wired and wireless) clients. The framework uses thecurrent state (resources, capabilities and limitations) ofthe system/network, and the current interests of the client,

to construct a client profile. This profile is used inconjunction with a policy driven inference engine tomanage and adapt application QoS. Adaptation policiesas well as mechanisms are programmable. Theframework architecture is implemented using lightweightJAVA components and uses SNMP (WINSNMP and NTSNMP) to obtain system/network state. The framework isexperimentally evaluated using a collaborative imageviewer application. The evaluation clearly demonstratesthe adaptability of the system over a wide range ofcompression ratios and image qualities corresponding towidely varying network conditions.

No of Packets vs CPU Loadste20kc16a15P f10O8No of Packets .5o4N02030507090100CPU Load (%)Compression Ratio vs CPU Load no100100is80soeit60CompressionrpaR40m32.7Ratioo20C01.610.916.330507090100CPU Load (%)Bits/Pixel vs CPULoad le15x14.3iP r10ePBits per pixel s5tiB02.21.40.7030507090100 CPULoad (%)Figure 9 - Image viewer parameters versus CPU load

VII. REFERENCES:

[1] “The Info Bus Specification,” available at

http://java.sun.com/beans/infobus/ibspec.html.

[2] B. Bayerdorffer, “Distributed Programming with Associative Broadcast,”

Proceedings of the 27th Annual Hawaii International Conference onSystem Sciences. Volume 2: Software Technology (HICSS94-2), Wailea,HW, USA, pp.353-362, 1994.

[3] Campbell. A., et al., “A review of Qos Architectures,” ACM Multimedia

Systems Journal, 1996.

[4] Common Object Request Broker Architecture - http://www.omg.org.

[5] Dave Perkins et al., “Understanding SNMP MIB's,” Prentice Hall, Nov.

1996.

[6] E. Lamboray, “Progressive Transmission for Very Low Bit-rate Context-based Image Compression,” Technical Report, Rutgers University, June1997.

[7] Eliot Rusty Harold, “Java Network Programming,” O'Reilly Publishers.[8] Geoffrey Fox et al., “Tango - A Collaborative Environment for the World

Wide Web,” Technical Report, NPAC Syracuse University, Syracuse NY.[9] Habanero - http://havefun.ncsa.uiuc.edu/habanero/Whitepapers/ecscw-habanero.html.

[10] Henning Schulzrinne, “Real Time Transport Protocol,” MCNC 2nd Packet

Video Workshop, vol. 2, (Research Triangle Park, North Carolina), Dec.1992.

[11] J. M. Shapiro, “Embedded image coding using zerotrees of wavelet

coefficients,” IEEE Transactions on signal processing, San Fransico,CA,Pagers 657-660,1992.

[12] J. Munson, “Collaboration Bus Infrastructure: Bus Agents,”

http://www.cs.unc.edu/~munson/DARPA/busagent.html.

[13] James Murray, “Windows NT SNMP,” O'Reilly publishers, January

1998.

[14] Jim Farley, “Java Distributed Computing,” O'Reilly Publications.

[15] K. Mani Chandy, Joseph Kiniry, Adam Rifkin, and Daniel Zimmerman,

“A Framework for Structured Distributed Object Computing,” TechnicalReport, California Institute of Technology.

[16] Korth, Abraham Silberschatz, “Database Systems Concepts,”

2nd edition, 1991.

[17] M.Campione, Kathy Walrath, “The Java Tutorial,”

http://java.sun.com/docs/books/tutorial/.

[18] P. Bhandarkar, M. Parashar, “Semantic Communication for Distributed

Information Co-ordination,” IEEE conference on Information technology,1998.

[19] P. Agrawal et al., “Battery Power Sensitive Video Processing in Wireless

Networks,” PIMRC'98.

[20] S. Shervin, J.C. Oliveira and N.D. Georganas, “Applet-Based

Telecollaboration: A Network-centric Approach,” IEEE Multimedia,Spring/Summer 1998.

[21] SNMP Documentation, http://www.snmpinfo.com.

[22] W. Richard Stevens, “TCP/IP Illustrated,” Addision Wesley Publication.[23] WinSnmp Documentation, http://www.winsnmp.com.