Quality of Service in Heterogeneous Distributed Systems

Quality of Service in Heterogeneous Distributed Systems 1

Christophe Diot1 and Aruna Seneviratne2 INRIA Sophia Antipolis, F ar nc e, Ph.+33.93.65.77.56, e.mail: [email protected] 2 University of Technology, Sydney, Austr alia, Ph. +61.2.330.2441, e.mail: [email protected] du.au

Abstract The future distribute dsystems will be very heterogene ous. They will consist of end computing systems with widely varying capabilities, inter connected by networks which oer very dierent services. Delivering data with a guarante edlevel of servic e is a real problem that requir es both support at the network level integrated services networks and a QoS Quality of Servic emanagement envir onment. This paper describes a simple and exible approach to QoS management. This approach is based on resource reservation by the way of trac control and application dynamic adaptation. It aimsattemps to provide a simple ans ecient solution to the best possible quality of transmission considering the level of resource reserved and the current network load. Our appr oach is compared to other classic QoS archite ctur es.It is analysed, and evaluated with an exp erimental implementation in the Internet environment.

1 Introduction The future distributed systems will be heterogeneous. They will consist of computer systems end systems with widely varying capabilities, interconnected by netw orks which oer dierent levels of service. Moreover, the levels of service that can be supported by the various components of the system will vary during the life time of a connection, especially in the case of user mobility. Two methods of dealing with this heterogeneity and variability have been proposed. One approach we call the guaranteed approach is based on the notion ofnegotiating and re-negotiating level of service or quality of service QoS that will be provided. System architectures that attempt to negotiate this type of QoS contracts assume that some form of service guarantees can be provided by both the end systems and the netw orks.The other approach we call the best-eort approach assumes that service guarantees cannot be obtained from the system. These architectures are based

on algorithms that enable the applications to adapt themselves to the operating environment. In systems that proposed which are based on guarantees, the application designer needs to build in the lev els of operation the application will support. In addition, the designer needs to specify system resource required to run the application at each of the abo ve specied levels, for ev ery platform it is to run on.This information is then be stored in an application management information base. The trigger for QoS renegotiation is generated either as a result of lack of system resources or user request. The re-negotiation simply adjusts the application's QoS level to alter the resource usage to a sustainable level. Because of the service guarantee requirements, and the notion of QoS re-negotiation, the QoS architectures do not scale to environments where there are wide uctuations of service levels 13 . In addition, it requires apriori knowledge of the requirements of each level of operation on all available operating environments. This signicantly complicates the application designers task, and leads to complex QoS management arc hitectures that are di cult to be released in practice. The systems based on the best eort approach on the other hand attempts to lower the resource usage to a sustainable level by changing operational parameters. F or example, if there is a bottleneck it may emplo y a more aggressive compression algorithm. With tis type of scheme the user makes the choice and the adaptation algorithms are built into the application. Hence they do not require any specialised system services for resource reserv ation or apriori kno wledgeof resource requirements. How ev eras the adaptation is carried out using localised information, conten tion for resources among adaptive applications may lead to unstable system behaviour 25 . The "guarantee" based and "best-eort" based QoS management pro vides tw o alternateviews of resource management which is required for providing end-to-end QoS. Internet, today represents the broadest heterogeneous net w ork,and the services pro vided via the Internet is thus the best illustration of best-eort based designs. On the opposite end of the spectrum are

Proceedings of The Thirtieth Annual Hawwaii International Conference on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

networks based on ATM technology. These networks, have been designed to provide performance guarantees. Thus the services using the Constant Bit Rate CBR facilities of ATM netw orks would be examples of guaran tee based designs. Howev er as mentioned, both schemes suer from serious drawbacks. We believe that it is possible to overcome the drawbacks of the schemes based on guarantees and adaptive approaches by combining adaptivity with user control by providing the user with: exible controls to allocate resources to applications dynamically, and information that is needed to make the appropriate resource allocations. This alleviates the necessity to have explicit knowledge about application resource requirements as the user decides the resource allocations. Also, as the user who has a global view makes the allocations, applications will not suer from stability problems. This paper presents a QoS management framework that is based on the above principles with some experimental results that will illustrate its viabilit y. The paper is organised as follows. Section tw o pro vides an overview of the related work on QoS management arc hitectures and adaptive algorithms. It shows that both the QoS Architectures that ha vebeen proposed in the literature, and schemes based on adaptive algorithms can be represented using a common generic model. Section three highlights the drawbacks of these approaches by examining the implementation of tw o QoS management schemes implemented on tw opublic domain applications. This section illustrates why both approaches on their own cannot satisfy end-toend QoS management requirements. Section 4 presents our framework which is based both on adaptation and reservation. Finally, section 5 discusses our approach in light of some early experimental results.

2 Related Work A number of research groups are addressing issues associated with end-to-end QoS management. This work however has either concentrated on the development of QoS management frameworks or developing adaptive algorithms for specic applications. In this section we will cite tw otypical examples from each category to highlight the dierence with our methodology.

2.1 Guarantee based QoS Management The Lancaster QoS Architecture 4 provides a framew ork which is based on the concepts of the ISO reference model. The framework consists of 6 layers, and three planes. The low erfour layers are equivalen tto the lo w er four layers of the OSI Reference Model with enhanced protocols to support multimedia communications and the required policing. The fth layer is similar to the session layer of the OSI reference model, but in the QoS-A it deals with synchronisation of dierent media streams. The sixth layer is an integration layer which provides a uniform interface to the applications thus facilitating application portability. How ev er, due to its complexity, no full realisations of the architecture is available. The QoS Broker is another QoS Management Arc hitecture proposed by the researchers at the University of Pennsylvania 23 . It is less complex than the Lancaster model, and considers the system to be a collection of three components. Two of the components, application and transport, are considered to be within the end systems and managed by the end system QoS Manager, referred to as the QoS Broker. The other component, the netw ork,is considered to be managed separately by a netw ork QoS Manager and is assumed to provide a service to the QoS Broker. This model is somewhat similar to the model presented in this paper. The difference is that it is based on a negotiation concept. As a result, the system infrastructure necessary to support ha ve again been proven to be too complex 24 .

2.2 Best Eort based QoS Management As mentioned earlier, there are a large number of adaptive application designs as well. However in this section w e willonly pro vide an overview of tw o such applications. The other applications use very similar techniques. The Row ena Project 5 has developed a distributed MPEG pla yerthat adapts to the operating environment. In this implementation, users can specify their desirable display rate. When end system resources becomes scarce, the clien t will attempt to keep up the presentation rate by dropping frames as required. The number of dropped frames is used to estimate the netw ork or the system load. In the case of network load, the server is requested to reduce the frame rate. VOSAIC Video Mosaic is another example of a adaptive application 6 . It is a WWW bro wser that incorporates real time video and audio into standard


1060-3425/97 $10.00 (c) 1997 IEEE

hypertext pages. In VOSAIC, again the server receives information from the client and uses an adaptation algorithm to optimise the transfer bandwidth. These as well as the other reported adaptive algorithms all infer the state of the system from feedback messages. As a result the applications take decisions without global knowledge. These suer from the problems such as those highlighted by 8, ie when one application detects system congestion and reduces its load, another sees the availabilit yas an indication of plentiful resources and increases its usage. More importantly, they do not attempt to integrate system level resource management with application adaptivity as we propose.

as those reported in 28, 22, that take into account requirements of multimedia applications. Net w ork lev el quality of service will be provided using emerging Integrated Service Netw ork's 18, 26 facilities to make bandwidth reserv ation with a dened level of quality. For example in B-ISDN systems, it will be possible to dene four classes of reservation in line with the four trac classes that have been dened. In the emerging Internet model IntServ, it is proposed to provide trac control in the netw ork nodes,based on sc heduling, admission control, and policing. Application

3 Guaranteed QoS Management The basic function of QoS management can be seen as attempting to provide the services that can be oered with the available resources to a group of applications to: at least satisfy each application's minimum needs, and to maximise the utilisation of the available resources. In other words, it is the matching of available resources to possibly con icting demands. A system can be thought of as consisting of a number of subsystems. A distributed system will consist of two subsystems, namely the operating and the communication subsystems. These subsystems will in turn depend on the next low erlevel. F or example, in the case of the communication subsystem it will be the intermediate netw orks, and in the case of the operating system, it will be the managers of the resources such as CPU, Memory and IO devices. This is schematically shown in Figure 1. Thus the above matching process needs to occur at various levels. A t the highest level, it will be applications requesting certain levels of service from the system supporting it. At a low er level, the system can be thought of as consisting of a number of subsystems. This can be viewed in a similar manner to how the communication systems are represented in the OSI Reference Model 10. At the dierent levels la yers of the system, the Quality of Service QoS that can be oered to the level above will depend on the services oered by andor the availability of resources at the level below. The end system quality of service will have to be provided by using new resource management schemes such

Communication Subsystem

Operating System

CPU

Memory

I/O Devices

Networks

Figure 1: Generic QoS Management Model Thus, at each level, it is necessary to co-ordinate the services that can be provided to the lev el abo vein a collectiv e fashion.These co-ordination activities are in eect the QoS Management of that layer and is shown by the shaded boxes in Figure 1. The co-ordinating activities may vary in complexity. One extreme is to utilise the traditional operating system techniques to manage the end system resources and standardised protocols to pro videcommunication. This in eect does not provide an y system lev el QoS Management functionality apart from the resource management oered by the operating system. These ha vebeen sho wn to be unsuitable for distributed multimedia applications 17 . The alternative is to provide system level support for QoS or develop algorithms that will enable the application to adapt to the operating environment.


1060-3425/97 $10.00 (c) 1997 IEEE

3.1 System Level QoS Management

System level QoS management involv es one or more of the following features, at various levels of the system: keeping track of the level of service that can oered with the resources it has at its disposal, ie. accounting, make decisions as to how best to use these resources, ie. logic to determine what need to be done according to specied QoS policy, once a service level is accepted, reserving the required resources, ie. the mechanisms that are necessary to implement the QoS policy, and monitoring to ensure that the oered service and usage are at the agreed levels, ie. policing This is schematically represented in the Figure 2 below.

eringincreasing the amount of information or the frequency with which information is exc hanged. These tw o referred to as throughput adaptation and dela y adaptation respectively, attempt to optimise the use of available resources. Thus, as opposed to guaranteed architectures based on reservations, they operate as well as they can within the available resources - ie best-eort. There is evidence that applications can successfully use delay and throughput adaptation techniques.

3.2.1 Delay Adaptation

Multimedia applications can easily adapt to variable dela ys, if indeed these delays remain reasonably stable. The well known "proof of concept" is the "vat" 19 audio-tool, which maintains a constantly upgraded estimate of the average and standard deviation of the transmission delay. Once these parameters are known, one can compute a correctly sized "pla yout buer". The signal distortions which the varying delays could cause are thus compensated.

3.2.2 Throughput Adaptation

Figure 2: System Level Quality of Service Management Each of the above tasks is complex and there are possibly many dierent ways of providing the required support for these tasks. Moreover, methods of providing this system infrastructure are being investigated. As a result there are no complete system based on system lev el QoS management.

3.2 Best Eort QoS Management

Best eort QoS management schemes are based on adaptive application designs. An application is said to be adaptive if it can reduceincrease the amount of resources it requires whilst running. F or example in the case of the end system, the CPU load can be varied according to the video encodingdecoding scheme chosen eg. CPU load can be determined by the occupancy of transmission buers or deadline misses . In the case of the communication system, bandwidth can be adjusted depending on the state of network congestion that can be inferred from the loss rate or round trip delays . An adaptation will result either in a low-

Multimedia applications can easily cope with some amount of net w ork congestionor transmission errors, by adapting their throughput to the current state of the operating environment. There are in that case dierent mechanisms available. V ery often, the signal is naturally redundant. Missing images can be replaced by the next value, missing sound segments can be interpolated from the preceding and following segments. These techniques are in all cases media dependent, and in fact vary with the compression algorithm in use. Some form of explicit redundancy may be required, but that is routine work in an engineering laboratory. Multimedia applications can easily adapt to variable transmission capacity. There are in fact tw o dierent techniques, variable compression ratios and hierarchical encoding. Hierarchical encoding is a well known capability of sev eral w avelet compression algorithms. An image is sent into a set of packets which provide increasing definition. When the netw ork is unloaded, all pac kets arriv e. If the netw orkis congested it drops rst the high denition parts, which are carried in low er priority pac kets. This system is quite attractive, specially for large broadcast where closed loop control would be impractical.


1060-3425/97 $10.00 (c) 1997 IEEE

Variable compression techniques alter some characteristic of the coder to adjust the information rate. The most commonly used schemes either vary refresh rate or quantisation resolution. INRIA video conferencing system IVS has successfully uses variable compression techniques 27. The senders constantly poll the net work to estimate the quality of reception, compute the available bandwidth and set the compression ratio accordingly see gure 3. This mechanism is used to con trol both video and audio signals. feed−back

raw video

video coder

network

undesirable features. Firstly, as sho wn by 8, application behaviour tendsto div erge rather than con verge. That is one application will be using up more resources than the other. Secondly, the user does not have any con trol over which application is to receive the resources. The only control the user has is to terminate the application, which is less important. The divergence is caused by the application lev el adaptation algorithms taking decisions based on information inferred from feedback messages. F or example, in the case of a simple feedback based adaptation scheme, similar to the one used in the rst versions of IVS, the feedback message may be delayed due to net w ork congestion or processing load at the receiver. How ev er, the transmitter has now way of determining the cause of delay. This is sho wn in Figure 4 where the delays caused by end system load aggravates the situation.

Figure 3: Net w ork feed-bac k control

4 Implementation Issues

300 "ivs_adaptation"

250

200 max_rate(Kbps)

The usage of dead-reckoning within distributed simulations, for example, has the same potentialities as varying the frame rate of a video signal the precision of the object descriptions could vary in much the same w ay as the number of bits per pixel. Applications could also de ne priorities in the information they are transmitted. This is the case in the MAGIC testbed, where the TerraVision application organises its data in three dierent les: high priority, low priority, transmit if room 20.

150

100

50

0 0

Much of the discussion about system level QoS management assumes that applications have xed requirements, and that they are not going to run if the system cannot pro vide the required resources. We ha vealready evidence to the contrary for many applications, especially in the case of audio and video transmissions which are supposed to make the bulk of "real time transmissions. F or example, the CPU requirements of a MPEG video stream will depend on the con tent of the video stream. Therefore, even if the infrastructure necessary for reserving all the resources could be provided, it will not be possible to develop eective system level QoS management schemes. Adaptive application designs perform well when used in isolation. Howev er, when more than one adaptive application is executing concurrently, they exhibit tw o

50

100

150 Time(sec)

200

250

300

Figure 4: IVS Throughput with Receiver Load This can be overcome by reserving system resources as described in section 3. The eect of reserving CPU on the above implementation is shown in gure 5. How ever, to exploit this, rst, mechanisms for reserving all the necessary resources in the system has to be developed. Secondly, the resource requirements of the applications should be kno wna priori. Methods of reserving netw orkresources are fairly w elladvanced. Moreover, as mentioned earlier, determining the resource requirements of each application is virtually impossible.


1060-3425/97 $10.00 (c) 1997 IEEE

These adaptive applications in turn will run on top a tw olev el QoS management hierarchy. The rst lev el will be dealing with the user lev el QoS andthe other will deal with the application level QoS. Adaptation also allo ws for the ltering of lower priority pac kets in the net works nodes in order to adapt the bandwidth, the processing capabilities, or the required level of service of a group member. Adaptation does not require hard reservation mechanism like VC establishment, but works well with dynamic trac control based on sc heduling and admission control.

300 "ivs_adaptation"

250

max_rate(Kbps)

200

150

100

50

0 0

50

100

150 Time(sec)

200

250

300

Figure 5: IVS Throughput with Receiver Load with CPU Reservation

5 An Adaptive QoS Management Framework We have sho wn in the previous sections that a QoS model based solely on either system level mechanisms guaranteed or on application level adaptation best eort is not ecient. This analysis sho ws that the requirements for a viable QoS framework are: Scalability: adding a new application or a new participant to a group application should not change the QoS of ongoing applications beyond speci ed limits. Individuality: the user should be able to select the Quality of Service of eac h application individually. A single quality for a group of users is not acceptable. Simplicity: QoS pro vided at the application lev el should be simple to use and easy to implement. It should avoid complex layered QoS architectures with associated mapping functions Flexibility: The QoS should be able to uctuate with the availability of system resources within limits. A QoS management scheme should have a contract with the application. Such a contract should not be broken because of transient conditions. In the following sections we present an adaptive QoS framework which will satisfy the above requirements.

5.1 Proposed F ramework

The proposed framework separates the application level adaptation and system lev el adaptation. It assumes that the applications will employ adaptive algorithms.

5.2 The Application Level QoS The application level QoS deals with information that can be bound to the application itself, ie. the requirements of the application to execute correctly . Thus w e separate this information from the information associated with using the application. The information application level QoS management facility will use in the proposed framework are: Minimum CPU and memory needs, Peripheral equipment requirements, Reliability no retransmission or retransmission, and error control, Type of Trac allows to select among window ow control or rate ow control, and Data delivery order: in sequence or out of sequence The application level QoS is de ned during the application design phase. The rst two items will be used by the user level QoS management facility to determine whether the requested application can run. The remainder will be used to organise the communication subsystem. All these information are then provided to the User QoS management facility through a QoS MIB Management Information Base. The QoS MIB will be maintained manually , or using management protocol and tool. In the case of the communication subsystem, the application QoS management facility will be able to select a communication subsystem tailored for the given application and operating environment. Approach for this application QoS facility could be similar to those described in 12 or 14. Both approaches use formal descriptions of the application eg. Petri nets and ESTEREL speci cations to automatically bind end-toend control mechanisms to the application.


1060-3425/97 $10.00 (c) 1997 IEEE

5.3 The User level QoS User level QoS is associated to the operating environment user personality, user needs, available netw ork connection, user budget not to the application itself. Thus user level QOS will be dened for each session. In our model we propose to use the following parameters: Price - the maximum amount of money the user w ould be prepared to spend. Preference - whether system should attempt to provide delaythroughput adaptation.

management capabilities are available, it may utilise a scheme similar to Unix "nice" command to adjust the priorities of the executing processes. If more advanced resource management scheme suc h as these described in 28 are a vailable, these will be used. If the allocations of resources are not correct, and the user is not satised, the process can be repeated. If no changes can be made, the engine will inform the user the reason as to why no change is possible. User

We do not believe that there is a need for more parameters. In fact, the most critical criterion to select a QoS level is generally expected to be the price. This is analogous to the airline model . When one buys an air line ticket, the QoS parameters are the pric e and the class. In classic OSI-like QoS management arc hitecture, the user would have to specify something like the depth of the seat, the type of meals calories, take-o dela y, size of the plane, etc. Each of these parameters w ouldbe specied with four dierent values, namely minimum, average, maximum, and target value. We believe that the problem of QoS description is the same in a communication netw ork and in the airline model. The User QoS is selected at each instantiation of an application, ie. at application run time. This will be facilitated by a generic user QoS management facility. The User QoS management facility has t wo tasks: Create and maintain a transmission context resource reservation, access to integrated service, negotiations with remote entities. Inform the User of the "state" of the transmission. This is schematically shown in gure 6. Once the user indicates the adequacy of the quality she is receiving, the User QoS engine will interact with the end system resource manager and the application lev el QoS manager to determine the allocation of resources to the application. This will be done within the constraints of the specied cost and privileges of the other applications that are executing concurrently. T odo this and guarantee that the other applications are not aected adversely, it is necessary for the User QoS management facility to have a global view of the system. The end system resource allocations will be made according to the availabilit y of resource management facilities of the host system. F or example, if no resource

Application MIB User QoS engine Application

Local resource (CPU, memory)

Integrated Service network

Figure 6: User QoS engine A consequence of this approach is that there is no changes required to the application. The User QoS engine is the intermediatory in the class of service negotiation. A user wishing to start an application just has to call the User QoS management facility and ask it to start the application with a dened lev el of service. This management facility is in charge of chec king the application parameters, making the reservations locally and on the net work, and starting the application. Then, once the conditions are accepted by the user, the role of the QoS management facility is just to "inform". Maintaining the QoS lev el will be done by application's adaptation algorithm and the netw ork con trol. Under any circumstances, this model is simpler than an y of the proposed schemes rstly , because what is guaranteed is not the specic qualit y,but rather the best quality that can be pro vided under the experienced conditions. Secondly, the resource allocations are done iteratively by the user, rather than by the system. Finally, in this model, the QoS managers only takes a decision as to whether to execute an application or not. It nev er tak esan yother decision, but simply informs the user of the possible alternatives. Thus its role is to maximise eciency and inform the user. It is up to the application user make the decisions.


1060-3425/97 $10.00 (c) 1997 IEEE

6 Discussion This paper analyses the requirements of QoS management. It justies its use in future networking environments especially for supporting time-constrained multimedia applications. It then categorises the current approaches to QoS management into tw o categories and illustrates the shortcoming of these tw oschemes. Finally, it presents a no vel hybrid QoS management scheme which overcomes the limitations of the schemes that have been proposed thus far. The major advan tage of the proposed scheme is its simplicity, both in terms of the end system support and the communication subsystem. In this section we discuss the consequence of the proposed scheme on an enhanced protocol architecture, as an indication of this, ie it will lead to a simple network environment where reservation will only play a limited role. Classic reservation mechanisms require virtual circuit management. This solution is not exible as it cannot dynamically vary the allocated bandwidth in accordance with the available resource on the network. Trafc control pro vides the same advantages as reservation but with more exibility and a more ecient management of netw ork resources.It simply requires scheduling of packets in dierent level of priority queues, and admission control mechanism to ensure that the queues are not overloaded. Trac control is v ery similar to application adaptation in that it allo ws packet ltering, ie. dropping packets of lo w er priorit y to solve congestion situations. If w eaccept the idea that all members of the same service class ie. sharing the same communication channel should be treated equally, w e nd very little use for reserv ation. In fact, w esee tw ousages of the reserv ation procedure: a reven ue opportunity for netw ork operators and a prioritization tool for netw ork recipients. Netw ork operators may wish to implement a reserv ation service that will giv e more resources to whoever pays more. Whether or not this makes sense economically depends of the cost of the procedure, including the signalling, accounting and billing. If w e accept the idea that a user's share of the resource is xed, the main usage of reservation could be to assign user-dened priorities. Receiver triggered reserv ation tools such as RSVP could be used to dene which data ows are more important, eg. the audio track of a video transmission, and to make sure that if the netw ork has to drop something it does not drop the important streams. Applying a combination of fair sharing and class based queuing

has the immense advan tage of keeping the network very simple. All fairness enforcement are local. Routers can pic k which ever algorithm they prefer to enforce them, and costly signalling procedure and fragile ow set-up mechanisms are not needed. We ha vemade some experiments with an adaptive application the videoconferencing tool VIC and with basic trac control implemented in a router 21: Our schedular uses tw oclasses of data ows. The low priority or normal trac which is not controlled, is a best-eort class where non real-time trac is sent. The high priority class is dedicated to real-time packet VIC packets in our case with 900 kbps bandwidth reserved A round-robin algorithm is used to guarantee the reserved bandwidth. An admission con troland a classier modules ha ve been designed to manage incoming data ows on router interfaces. In the experimental conditions, the User QoS is predened in order to simplify the implementation and to focus on service guarantee. Two sources send data to the router. A non real-time source which goal is to congest the network. And a real-time source made of VIC packets. The performance results are obtained from the senderreceiver reports delivered by VIC. We keep VIC sending at a constant rates 50, 200, 400, 600, 800, 1000, 1200, 1400 and 1600 bits per second. We compare the received rate with the sent rate gures 7 and 8. The results observed conrm the expected beha viour. In case of low net work load, the con trolled trac is a little slow erthan the same trac sen t through ordinary FIFO without pac ket classication gure 7. The dierence is due to extra processing consequent to trac control mechanisms. The advantage of the trac control TC extension is visible and sensitive when the netw ork is congested gure 8.The bandwidth allocated to the real-time application without trac con troldecreases to an unacceptable lev el. Instead of decreasing, it remains stable because VIC automatically adapts to the available bandwidth. With trac con trol, the bandwidth is maintained at the guaranteed level of 900 kbs. Similar measures made on loss rate show that in that case too, VIC adapts to the request bandwidth to send no more than 900 kbps the number of loss is stable too at 900 kbps.


1060-3425/97 $10.00 (c) 1997 IEEE

1600 "without_TC" "with_TC"

Bandwidth Received

1400 1200 1000 800 600 400 200 0 0

200 400 600 800 1000120014001600 Bandwidth Sent

Figure 7: Performance with adaptive application and trac control netw ork not congested 900 "without_TC" "with_TC"

Bandwidth Received

800 700 600 500 400 300 200 100 0 0

200 400 600 800 1000120014001600 Bandwidth Sent

Figure 8: Performance with adaptive application and trac control netw ork congested This proves the feasibility and the eciency of a simple QoS management scheme with a very little n umber of parameters in our experiment, parameters are preference throughput adaptation and price 900 kbps. Similarly, w ehave dev elopedan experimental User lev el QoS management facility by using a modied version of the stride schedular 7 . It w as tested using again a modied version of the Oregon MPEG player 5 . Early results show that this type of QoS management can be used with video and audio applications very eectively. Moreover, that they result in better user satisfaction. Therefore , it could be concluded that the hybrid QoS management framework that has been presented will result in a simplied and more ecient arc hitecture

of the communication system, and provide user with the user with more control. This in turn will lead to more ecient utilisation of network resources no hard reserv ation, but trac control better user satisfaction. However, there are a number of issues that still need to be addressed before the scheme can be fully assessed. Firstly, how the dierent feedback mechanisms will interact when there are multiple resources need to be investigated further. Secondly, user reaction to having con trol of the QoS levels need to studied. We are currently in the process of addressing these by attempting to develop a simulation model that models the operation of the proposed framework, and integrating the w ork reported in 21 and 7 .

References 1 M. B. Abbot, and L. L. Peterson. Increasing Netw ork Throughput by Integrating Protocol Layers. IEEE ACM Trans. on Networking. October 1993. 2 J-C. Bolot, T. Turletti. A rate control for packet video in the Internet. IEEE INFOCOM '94. Toronto. 3 J-C. Bolot, T. Turletti, I. Wakeman. Scalable feedbac k control for multicast video distribution in the Internet. A CMSIGCOMM '94, Vol. 24, No 4. October 1994. 4 A. Campbell, G. Coulson, and D. Hutchison. A Quality of Service Architecture. ACM Computer Communication Review. April 1994. 5 S. Cen, C. Pu, R. Staehli, C. Cow an and J. Walpole. A Distributed MPEG Video Audio Player. 5th International Workshop on Network and op erating System Support for Digital Audio and Video. Durham, April 1995. 6 Z. Chen, S. Tan, R. Campbell and Y. Li Real Time Video and Audio in the World Wide Web 4th International World Wide Web Conferenc e Massachusetts, December 1995. 7 H-S. Cho, and A. Seneviratne. Dynamic QoS Control without the Knowledge of Resource Requirements. submitted for publication. 8 C. Compton and D. T ennenhouse Collaborative Load Shedding for Multimedia Based Applications International Conference on Multimedia Computing Systems Boston, May 1994


1060-3425/97 $10.00 (c) 1997 IEEE

9 A. Danthine, Y. Baguette, G. Leduc, and L. Leonard. The OSI 95 Connection-mode Transport Service Enhanced QoS. 4th IFIP conference on High Performance Networking. Liege. December 1992. 10 A. Day, H. Zimmermann. The OSI Reference Model Pr oceeding of the IEEEDecember 1983 11 S. Deering. RFC 1112: Host extensions for IP multicasting. August 1st, 1989. 12 M. Diaz, K. Drira, A. Lozes, C. Chassot. De nition and Representation of the Quality of Service for Multimedia Systems. IFIP International Conference on High speed Networks. Palma Spain. September 1995. 13 C. Diot, C. Huitema, and T. Turletti. Multimedia Application should be Adaptive. HPCS Workshop. Mystic CN. August 1995. 14 C. Diot, I. Chrisment, and A. Richards. Application Level Framing and Automated Implementation. IFIP International Conference on High speed Networks. Palma Spain. September 1995. 15 Recommendation H.261: Video codec for audiovisual services at px64 kbitss. International Telecommunication Union ITU-T. 1993. 16 C. Huitema, T. T urletti. H.261 software codec and w orkstation Videoconferencing. Pr oceedings of INET '92. Kobe, Japan. August 1992. 17 D. Hutchison, G. Coulson, A. Campbell, G. Blair Quality of Service Management in Distributed Systems Networks and Distributed Systems Management, Ed M. Solomon Addison Wesley, 1994 18 S. Shenker, D. Clark, and L. Zhang. A Service Model for an Integrated Services Internet. IntServ charter: http: www.ietf.cnri.reston.va.us html.charters intserv-charter.html. October 1993.

19 V. Jacobson. vat - X11 based audio conferencing tool. Unix manual pages. F ebruary 1993. 20 Y. G. Leclerc, S. Q. Lau Jr. T erra Vision: A T errain Visualisation System. SRI International, T echnic al note N 540.April 22, 1994. 21 M. May and C. Diot. An Experimental Implementation of Trac Control in IP Networks. Submitted to IFIP Pr oto colfor High Sp eedNetworks workshop. Sophia Antipolis. October 28-30, 1996. 22 C. Merecer, H. T okuda. Processor Capacity Reserv es: Operating System Support for Multimedia Applications. International Conferenc e on Multimedia Computing Systems. Boston, May 1994. 23 K. Nahrstedt and J. Smith The QOS Broker. IEEE Multimedia Magazine Spring, 1995. 24 K. Nahrstedt and J. Smith A Service Kernel for Multimedia End Stations. Proceedingof the IWA CA September, 1994. 25 A. Seneviratne and H. S. Cho. Quality of Service Management Mapping in Multimedia Systems. IEEE Multimedia Networking Conferenc e. Aizu, Japan. September 1995. 26 W. Stallings ISDN and Broadband ISDN with F rame Relay and ATM Prentice Hall 1995 27 T. Turletti. The INRIA Videoconferencing System IVS. ConneXions - The Interoperability Report. V ol. 8, No 10, pp. 20-24.October 1994. 28 C. Waldspurger and W. Weihl Stride Sc heduling: Deterministic Proportional Share Resource Management. MIT T echnical Memorandum MIT LCS TM528 June 1995


1060-3425/97 $10.00 (c) 1997 IEEE