Admission Control Policies in Integrated Services Networks

Transcription

1 Admission Control Policies in Integrated Services Networks Roberto Canonico, Simon Pietro Romano and Giorgio Ventre Dipartimento di Informatica e Sistemistica, Università di Napoli Federico II, Napoli, Italy {canonico, sprom, ventre}@grid.unina.it Abstract. Modern integrated services networks must offer a wide range of communication services. For these networks, the ability to provide individual data flows with Quality of Service guarantees is of paramount importance. To obtain this goal, allocation of resources in the network infrastructure can be performed through explicit reservation, via the undertaking of a contract by the two interested parties: the network and the client. The strength of such a contract, and, consequently, of the QoS guarantees, depends on the capability of the network to control the effective usage of the resources by its clients. A possible solution relies upon admission control, which limits the number of simultaneous connections in a network. This paper discusses the main issues related to the admission control problem and illustrates some of the solutions presented in the literature. Keywords: Quality of Service, admission control, scheduling. 1 Introduction The ongoing convergence of computing and telecommunications technologies has made viable and economically convenient the utilization of modern commmunication infrastructures as integrated services networks, which support both traditional distibuted applications and new multimedia services, offering communications services characterized by a wide range of Quality of Service (QoS) requirements. Due to the variety of the adopted solutions, however, there exist differences in the way network architectures support such requirements. In particular, in the case of multimedia applications, two different approaches can be exploited. The first, referred to as adaptive, tends to adjust the application behaviour to the actual performance of the communication service provided by the network. The second approach consists in exerting some form of pro-active control over the achieved performance. This usually requires that applications declare the expected Quality of Service when a new request of service is submitted. In this case the network has to perform a preliminary check, best known as admission control, to verify that the requested performance is compatible with the current state of utilization of resources. Such an approach implies that usage of resources is governed by a resource manager, which allocates resources to different calls. Many solutions have been proposed in the literature for the admission control problem, each tailored to a specific kind of network. Yet, the problem is still open and

2 an ultimate solution is far to be accepted. This paper examines the problems related to admission control in modern networks from a general point of view. Section 2 describes the difference between best-effort and guaranteed-qos communication services. Section 3 introduces the admission control problem. Section 4 illustrates the parameters which typically define the quality of a communication service in an integrated services network. Sections 5 and 6 illustrates two fundamental aspects involved when performance guarantees are to be offered: traffic characterizations and service disciplines, respectively. Section 7 describes two approaches which relax the concept of guarantee, in order to maximize the usage of resources by real-time traffic. Section 8 discusses the admission control policies implemented in the switches of an ATM network. Finally, section 9 illustrates a solution to the real-time communication problem proposed by Tenet group at the University of California at Berkeley. 2 Guaranteed Quality-of-Service communication services Nowadays, the vast majority of computer networks provides communication services conforming to a service model known as best-effort, which can be described as the network does not guarantee minimal performance, so always strive to achieve better performance. The present Internet is based on the best effort philosophy, which makes possible to share the communication resources by exploiting the advantages of statistical multiplexing. Since the best-effort service does not take account of application needs, it may happen that some applications are provided considerably greater performance than they need, possibly at the expense of other applications. For this reason, in the case of limited resources, a crucial goal for best-effort systems is to achieve "fairness", that is to equitably divide the shared resources among all users. Networks which offer only best-effort services, typically do not have any form of admission control. Alternative to best-effort, a different approach consists in offering guaranteed quality of service, which can be described as reliably provide a level of performance that the user finds acceptable, no matter how other users are behaving. This approach is resource-conserving, assuming resources are expensive. As different applications have different requirements, it is usually assumed that the communicaton infrastructure provides a different QoS to each request of service (hereafter referred to as call). This makes necessary the introduction of resource allocation mechanisms to adjust network resources (such as bandwidth, buffer space, etc.) to the provisioned QoS. Of course, in order to allow the computation of the required amount of resources, it is also necessary that some characteristics of the traffic generated by each call are declared and that a mechanism of traffic policing enforces the traffic generated by each call to conform to the declared characteristics. An additional consequence of the guaranteed QoS approach is the necessity of some sort of pricing mechanisms that distinguish different QoSs; otherwise, the applications will always choose the highest available QoS. Resource reservation, pricing and billing mechanisms add a significant level of complexity to the network architecture, as they require to maintain state information per each established call in the network elements.

3 In the past years, several solutions have been proposed for the introduction of guaranteed communication services in modern computer networks. They all reflect one of the following two approaches: mapping new service models over existing networking technologies; designing new networking technologies with native guaranteed QoS communication services. The first approach is that followed by the Internet Engineering Task Force. It aims at re-engineering the existing Internet, by designing new communication services (e.g. Guaranteed Service, Controlled Load Service) which may be provided over heterogeneous internetworks, by means of new communication (e.g. IPv6) and reservation (e.g. RSVP) protocols. In this case, applications will be provided an architecture-independent API, and in each network element the requested communication service will be mapped over the native network-specific service, by a local QoS manager. The alternative approach consists in natively deploying the advanced features of new networking technologies (such as ATM) to provide guaranteed performance communication services. In this case, no mapping is required, but of course this approach does not work in heterogeneous environments. 3 Admission control Let us refer to a packet switched network with explicit per-call resource allocation. Such a network is composed of network elements, linked by communication links, and arranged in an arbitrary topology. Each network element switches traffic (i.e. packets) from one of the input links to one of the output links. Incoming packets are served according to a service discipline, which is responsible of the packet scheduling. Fig. 1 shows how packets are switched from a source A to a destination B, through a number of intermediate network elements, (N1, N2, N3). Fig. 1. Interconnection of network elements in a packet-switched network

4 The Internet, which is based on the IP network protocol, uses connectionless transport, where information is dynamically routed on a per-packet basis (datagram), depending on congestion and availability. This approach has proved to offer many advantages, such as robustness to failure and ability to dynamically route around points of congestion. Further, the absence of a per-connection state in network elements leads to a tremendous simplification in their software. On the other hand, it makes QoS guarantees difficult to realize. Telecommunication networks, instead, are typically based on connection-oriented transport, where information is constrained to traverse the same route from source to destination. This approach enables resource allocation along that route to actively control QoS. This distinction, however, is narrowing. The Internet Engineering Task Force is defining new Service Models (Guaranteed Service, Controlled Load Service) for the Internet, which require the introduction of per-session state information in the network elements (routers). At the same time, there is an effort to better support connectionless services on connectionoriented architectures, like ATM (e.g. IP switching). To offer QoS guarantees to individual calls, integrated services networks must be able to perform some kind of admission control. On a per-call basis, an admission control algorithm checks if the incoming call s Quality-of-Service requirements can be met and if admitting the call would reduce the service quality of established calls. This decision depends on the choice of scheduling disciplines and on the set of services provided by the network. If the previous check fails, the call is either delayed, until resources are available, or rejected. In packet switching networks, the admission control can be carried out in three different ways. The first approach, which we call centralized admission control, consists in realizing a single admission manager, which is responsible of admitting or rejecting the service requests for the whole network, on the basis of the resources utilization in all the nodes and links. In this way the use of QoS-aware routing strategies is enabled, i.e. it is possible the choice of a path within the network which, at the same time, meets the client requirements and maximizes the efficiency of the network s resources usage. We refer to the second approach as distributed coordinated admission control. It consists of performing the admission choice in a distributed manner, by performing a partial admission control algorithm at each node of the network. This approach removes the potential bottleneck of the centralized admission manager, but makes more difficult the selection of paths within the network according to QoS-aware strategies. In this case we use the term coordinated to indicate that the requested QoS refers to the end-to-end communication, so the admission algorithm progresses by cumulating the QoS parameters from the access node to the destination end-point, and the final admission decision is made at the destination on the basis of the data collected along the entire path. Finally, a third approach, which we call distributed un-coordinated admission control, consists in a priori splitting the required end-to-end performance parameters among the various nodes which form the data path in the network and in executing independent admission control algorithms in each of the nodes. The existing solutions to the admission control problem can be classified according to the above taxonomy.

5 4 QoS requirements description In this section we will briefly introduce the parameters which are commonly assumed to define the Quality of Service for a communication service. Typically, the following three QoS performance attributes are considered to be guaranteed: latency, that is the delay that the transmitted data will experience through the network. Within the network, latency is made up of two components: the propagation delay over the links and the service time in network elements (routers, switches). The latter comprises the variable queueing delays in network elements buffers. The end-to-end delay also includes the time required to accumulate a packet at the source (packetization delay). Related to latency, is the jitter, that is the variation of the end-to-end delay from its average value. throughput, that is the amount of data per time unit (averaged on a time interval T), successfully transferred from the source to the destination, through the network. reliability. In packet-switching networks, the major cause of unreliability is loss of packets due to buffers overflow in network elements. The maximum packet loss rate is usually used as a measure of the reliability of a communication service. Each of the QoS parameters for a given service request can be guaranteed deterministic bounds (no violations are admitted) or statistical bounds (the bound holds with a given probability). Guaranteeing deterministic real-time bounds is, of course, more expensive in terms of resources. Different applications have significantly different requirements for each of these three performance attributes, as shown in Table 1 [9]. Table 1. Communication requirements for several kinds of traffic Traffic Type Delay Requirement Throughput Requirement Interactive Simulation Low Moderate-High Yes Network monitoring Moderate Low Yes Virtual Terminal Low Low Yes Bulk Transfer High High Yes Message Moderate Moderate Yes Voice Low, constant Moderate No Video Low, constant High No Distributed Computing Low Variable Yes Reliable Delivery Network Control Moderate Low Yes

6 For interactive applications latency is often a critical element of subjective quality; thus, transport latencies are often required to be both short (tens or hundreds of milliseconds) and guaranteed. Guaranteed latency is particularly important for interactive applications conveying continuos-media streams. These applications typically perform a synchronous reconstruction with strict temporal requirements, and thus any data arriving outside of a time interval is not used, just as if it had been lost. This has led to attempts to guarantee bounded delays and bounded delay variations (jitter) in packet networks. Other applications have less critical latency requirements. Sporadic media often require reliable delivery, which can only be achieved over unreliable transport services through multiple transmissions, with the side effect that latency cannot be guaranteed. Reliability in transport is adversely affected by congestion, which may cause loss by buffer overflow, and bit errors caused by noise or interference in transmission (which may cause loss if they occur in the packet headers or corruption if they occur in the packet payload). The techniques available for improving reliability, including forward error-correction coding, diversity, and acknowledgment and retransmission protocols, have the fundamental side effect of increasing latency. Continuous media can tolerate reasonable levels of loss and corruption with adequate subjective quality. Computer networking, on the other hand, has typically dealt with sporadic media and thus has focused on transport techniques such as packet switching and statistical multiplexing, combined with transport protocols (like TCP) that guarantee reliable delivery at the expense of indeterminate delay. 5 Traffic models and traffic characterization To effectively manage traffic, a network provider must know not only the requirements of individual applications, but also their "typical" behaviour. A traffic model summarizes the expected behaviour of an application or an aggregate of applications. In terms of rate characteristics, continuous media can be represented by a continuous stream of bits with variable or constant bitrate, whereas sporadic media may have periods of very high bit rate alternating with idle periods. A common way to model data sources is by means of an on-off model, which may be used for voice, as well as bursty sources, including images and video [12]. Such a model assumes that the packet arrival rate is determined by the state of a continuous-time Markov chain. When the Markovian process is in the ON state, packets are generated at fixed intervals of T = (1/R) seconds; when the process is in the OFF state, no packets are generated. An on-off source is described by three parameters: the rate of packets transmission during ON intervals, R (in bps); the average length of a ON interval (burst), 1/β (in seconds); the average length of an OFF interval, 1/α (in seconds). The probability distributions of the ON and OFF intervals are typically assumed to be exponential. The on-off model is commonly used in simulation studies (e.g. [12]) to represent sintetic sources with given characteristics. However, when an application has to declare its traffic characteristics, which is required for both

7 admission control and traffic policing, a simpler description (traffic characterization) is preferred. This is usually done by means of a set of parameters, called traffic descriptors. An example of traffic characterization is the token bucket filter [8], [10]. A token bucket filter is defined by two parameters, a token rate r and a bucket depth b. The bucket is a buffer filled up with tokens, generated at a constant rate r. No more than b tokens can be accumulated in the bucket. Every time a packet of size p is to be transmitted, p tokens are removed from the bucket. In brief, a source conforms to a token bucket filter (r,b) if there are always enough tokens in the bucket whenever a packet is generated. More precisely [2], let us denote with t i and p i the generation time and size, respectively, of the i-th packet. A traffic source conforms to a (r,b) token bucket filter if the sequence {n i }, defined by: n 0 = b n i = MIN[b, n i -1 + (t i t i -1 ) r p i ] satisfies the condition: n i 0 for all i. The quantities n i, if nonnegative, represent the number of tokens residing in the bucket after the i-th packet leaves. Essentially, the r parameter specifies the continually sustainable data rate, while the b parameter specifies the extent to which the data rate can exceed the sustainable level for short periods of time. Fig. 2. Token bucket filter In [6], an alternative way to describe traffic characteristics is presented. It consists in declaring the following four parameters for each traffic source: the minimum packet interarrival time x min the average packet interarrival time x ave over an averaging time interval the duration I of the averaging time interval the maximum packet size s max The average packet interarrival time x ave is calculated over the busiest time interval of duration I. Specifying x min and x ave together with s max corresponds to declaring a peak bandwidth and a long-term average bandwidth, respectively. The x ave /x min ratio is a measure of the source burstiness, which is also affected by the choice of the I parameter. In fact, given x min and x ave, I determines how long the source can continuously send packets at the peak rate 1/x min, so that the larger I, the burstier the traffic.

8 6 Service disciplines in network elements In a network element, incoming packets are selected and then re-transmitted to one of the output links. The term service discipline refers to the policy adopted for scheduling packets for re-transmission. The choice of a service discipline is made according to different criteria. The provision of guaranteed QoS services requires the achievement of two different goals: isolation and sharing. Isolation is the ability to protect flows from misbehaving clients and from traffic distortion due to network load fluctuations. Sharing is the ability of mixing traffic of different sources, so to obtain the advantages of statistical multiplexing. The First Come First Served (FCFS) service discipline is effective for sharing, but it does not provide any isolation [2]. On the contrary, isolation is provided by rate-based service disciplines, which are able to provide each flow with a minimum service rate, independent of the traffic characteristics of other flows [14]. In the literature, several service disciplines to support the QoS in the context of high-speed networks have been proposed. A brief description of the main of them follows. Weighted Fair Queueing (WFQ). The WFQ discipline emulates the ideal (but unimplementable) Generalized Processor Sharing (GPS) scheduling discipline [8]. The GPS serves packets as if they were in separate logical queues, visiting each nonempty queue in turn and serving an infinitesimally small amount of data from each queue, so that in any finite time interval, it can visit every logical queue at least once. In this way, the GPS provides a fair allocation of the available bandwidth among different data streams. The WFQ approximates the GPS discipline, by computing the finishing time for each packet, that is the time it would take to completely serve a packet with a GPS server, and then by transmitting packets in order of these finishing times. Finishing times are more appropriately called finish numbers, to emphasize that they are only service tags that indicate the relative order in which packets are to be served, but they have nothing to do with the actual times at which packets are served. Virtual Clock (VC). Very much like in WFQ, a VC scheduler stamps packets with a tag, called virtual transmission time, and packets are served in order of their tags. These tags are computed so to emulate a time-division multiplexing. Delay-Earliest Due Date (D-EDD). With the D-EDD service discipline, the server negotiates a service contract with each source, providing to respect a delay bound as long as the source obeys a peak and average rate. To do so, the server maintains three queues for deterministic delay service, statistical delay service and best effort service. In each queue, packets are ordered by increasing deadlines. A packet s deadline is set to the time at which it should be sent to meet the delay bound, had it been received according to the contract. Packets in the first queue are given, in the case of a conflict, higher priority than packets in the statistical queue. However, when the transmission of the head packet in the statistical queue does not violate the deadline of the head packet in the deterministic queue, the head packet of the statistical queue is selected for transmission [6]. Stop-and-Go (SG). SG is based on a framing strategy, which divides the time axis into frames of constant length T. According to the SG discipline, a packet arrived in a switch on a incoming link l during a frame f should be transmitted onto the selected

9 output link during the next frame. Bandwidth is allocated to each connection as a fraction of the frame time. Hierarchical Round Robin (HRR). HRR is a service discipline that also uses a multilevel framing strategy. The HRR server has several service levels, where each level provides round-robin service to a fixed number of slots. A channel is allocated some number of slots at a selected level, and the server cycles through the slots at each level. The time a server takes to service all the slots at a level is called the frame time at that level. The key to HRR lies in its ability to give each level a constant share of the bandwidth. Jitter-EDD (J-EDD). The J-EDD discipline extends D-EDD to provide end-to-end delay jitter bounds (that is, a bound on the minimum as well on the maximum delay). This is accomplished by marking each packet with the difference between its deadline and actual finishing time. A regulator at the entrance of the next hop switch holds the packet for this period, so to provide the desired lower bound for the end-to-end delay. The fore-mentioned service disciplines differ in many aspects. For instance, WFQ, VC and D-EDD are work-conserving, that is a server implementing one of those disciplines is never idle when there is a packet to send. On the contrary, Stop-and-Go, HRR and J-EDD are non-work-conserving. In a non-work-conserving discipline, each packet is assigned, either explicitly or implicitly, an eligibility time. A packet cannot be transmitted before its eligibility time, even when the server is idle. Non-workconserving service disciplines can be used to bound the end-to-end delay jitter inside the network. This reduces the buffer requirements in downstream switches, due to the absence of traffic pattern distortions inside the network. In the presence of misbehaving clients or traffic distortions due to network load fluctuations, a certain data flow may ask a network element to be served at a higher rate than the bandwidth allocated to it. With respect to this problem, Zhang and Keshav in [14] classify rate-based service disciplines into two categories: rate allocating service disciplines, which serve packets at the higher rate as long as it will not affect other flows performance guarantees; rate-controlled service disciplines, which never serve packets at a higher rate under any circumstances. WFQ, VC and D-EDD fall in the first category, while Stop-and-Go, HRR and J- EDD belong to the second category. Notice that the rate-controlled service disciplines are also non-work-conserving, as only non-work-conserving disciplines can place an upper bound on the service rate of a channel. 7 Measurement based admission control The approach consisting in providing a priori (deterministic or statistical) bounds to perfomance parameters is usually referred to as a hard guarantees approach, as it is typical of hard real-time environments [5]. Admission control algorithms for communication services requiring hard guarantees are based on worst case performance bounds. In the literature, worst case delay bounds have been proposed

10 for different service disciplines in network elements. For a network with FCFS servers and on-off sources, simulation studies [12] have demonstrated a dramatic difference between the point at which the actual (simulated) delay distribution becomes small (i.e. less than 10-5 ) and the analytical delay bound. To maximize the utilization of resources by real-time traffic, several alternative soft guarantees approaches have been proposed. These approaches relax in different ways the concept of guarantee, typically introducing post-facto bounds for QoS parameters. In [2], a particular class of multimedia applications, called play-back applications, is introduced. These applications can adapt, to some extent, to variable network conditions and so they do not need hard QoS guarantees. The applications of this kind typically transmit a continuous media stream over a packet-switched network, reconstructing the original media stream at the receiver side. By buffering the incoming data and then replaying the continuous media stream, the receiver can compensate the delay variation (jitter) introduced by packet queueing in the network. The receiver replays the received stream by delaying its play-back point of a time interval. This delay lets the receiver accumulate incoming packets in a FIFO buffer. At startup the play-back point is set to a value calculated on the basis of the knowledge of the network load. In so-called tolerant adaptive applications, the playback point is adaptively adjusted to the minimum delay that still produces a sufficiently low loss rate. When incoming packets arrive after their play-back point, then the play-back point is readjusted upward, so that other packets may be buffered. The major inconvenience of this approach is the service disruption caused by playback point adjustments. For applications of this kind, which can tolerate the loss of a certain fraction of packets, but still require an upper bound on end-to-end delay, a predicted service is defined in [2]. This kind of service is based on traffic load measurements to estimate the actual delay bound. When network conditions are relatively static, this approach let the applications minimize their playback point delay, i.e. this approach minimizes the post facto delay bound. To implement the predicted service, a classical FCFS service discipline is argued to be the best. In fact, with a FCFS service discipline in network elements, every time a single source produces a burst of packets, all the other sources packets are temporarily delayed. This effect, which is named a multiplexing of bursts, is clearly unacceptable for guaranteed services, but, leads to smaller post facto delay bounds than other policies, under the same link utilization. Alternatively put, the introduction of a predicted service aims at admitting a greater number of calls, as long as the applications accept periodic, unpredictable disruptions in the service, as a consequence of the admission of new calls or of changes in the behaviour of other traffic sources. Unfortunately, several types of applications requiring QoS guarantees do not tolerate service disruptions, so they will unlikely benefit of such approach. More recently, the Integrated Services Working Group (IntServ) of the IETF has defined the so-called Controlled Load Service, which has the same characteristics of best-effort delivery in absence of network overloading. To guarantee such a behaviour, a new form of admission control has been proposed, called measurementbased admission control (MBAC), whose goal is simply to ensure that adequate bandwidth and packet processing resources are available to handle the requested level of traffic. MBAC is derived from the idea that sources cannot describe their traffic in

11 advance, so call admittance is based on a nominal description of the source. The actual source behaviour is then measured to automatically construct an appropriate traffic descriptor. Such an approach relies on the assumption that the future behaviour of a source may be predicted on the basis of the knowledge of the traffic generated by that source in the past. This kind of service can fit the requirements of tolerant adaptive applications, but, of course, cannot provide firm delay and packet loss rate bounds. 8 Admission control in ATM networks So far, we have discussed of generic heterogeneous integrated services networks, by presenting several solutions designed to be applied in internetworking environments and not bound to a specific kind of architecture. In this section we present the admission control problem in ATM networks, whose peculiar characteristics well suit the QoS requirements of an integrated services network. These admission control tests are required either when users intend to natively deploy the ATM connectivity services or when they want to map generic communication services (like those proposed by the IETF) over the basic ATM connectivity services. ATM networks are cell-switched, that is data is tranported through the network in the form of fixed-size packets, called cells. ATM connectivity is based on the notion of "virtual circuit", that is all the cells belonging to a virtual connection are forced to travel along a fixed route. This architecture enables faster routing and also reduces the addressing overhead per cell (since only local addressing is required). ATM networks basically provide two kinds of connections: CBR, Continuous Bit Rate, which is merely described by its constant transmission rate; VBR, Variable Bit Rate, which is described by a set of parameters, according to a leaky bucket model. When a new call arrives, requiring a virtual connection with specified QoS (bandwidth, delay, loss probability, etc.), whether to admit or not the call is decided by the admission control. Once the call has been admitted, the traffic generated by this call must be monitored to ensure it does not violate its declared characteristics. This control, exerted both at the user-network interface (UNI) and within the network, is referred to as a policing function, and in ATM networks is more specifically called Usage Parameter Control (UPC). The implementation of admission control policies within the switching nodes of an ATM network interacts with the signalling mechanisms, which are currently object of the standardization process carried out by the ITU and the ATM Forum. Admission control for CBR calls is simple because a CBR call i can be completely described by its rate, ρ (i). If a link has capacity C and has an admitted load L, then a new call i can be admitted if and only if L+ρ (i) < C. If a CBR call also has a delay requirement, and the link scheduler can provide delay bounds, the call may fail the admission control test if the best delay bound available from the scheduler is worse than the call s delay requirement. A call that fails the admission control test is either

12 rerouted, delayed till link becomes available or denied. A switch controller that denies calls instead of delaying them is said to implement a loss system. VBR calls send data in bursts, so that their peak rate (during a burst) may be much greater than their average rate. As a link s capacity increases and it carries more and more connections, the probability that all the sources simoultaneously send a burst into the link becomes small. Thus, if the number of sources is large, a burst from one source is likely to coincide with an idle period from another. In this conditions, the admission control algorithm can admit a call as if it were sending a CBR stream with a rate close to its long-term average rate. This considerably simplifies the admission control algorithm, but can result in delay bound violations due to statistical fluctuations. Thus, by characterizing the behaviour of an ensemble of bursty sources, we can make statistical delay guarantees to each source. This approach works well only when the number of sources is large. The easiest admission control algorithm for VBR calls is to treat them as CBR calls with a rate set to their peak rate (peak-rate allocation). Thus, the switch controller reserves enough resources to deal with a call even if it has no idle time between bursts. Clearly, this leads to an unefficient use of resources, both in the switches and in the network, precluding the use of statistical multiplexing. Other possible approaches are less conservative than peak-rate allocation; they use the results of the scheduling theory to allocate resources so that a connection meets its performance requirements even in the worst case. This avoids making statistical assumptions about other sources. For example, if the scheduler implements WFQ, we can allocate sufficient bandwidth at each switch so that the worst-case delay along the path is bounded, and sufficient buffers so that no packets are lost. This would simoultaneously meet the bandwidth, delay and loss bounds. 9 Admission control: the Tenet proposal The Tenet group at the Univerity of California at Berkeley has proposed a connectionoriented reservation-based approach to real-time communication in packet switching networks [4]. According to the Tenet approach, a real-time channel is established, and resources are reserved for it, after a round-trip establishment phase, which involves a distributed admission control and setup process. When the network accepts a service request, both the network and the client are bound by a form of contract. The client specifies the characteristics of its traffic, according to the (x min, x max, I, s max ) traffic characterization described in 7, and gives a pledge to respect them for the duration of the contract. As long as the the client keeps its part of the contract, the network promises to provide a service conforming to the client s requirements, guaranteeing a priori, in the absence of network failures: an end-to-end delay bound D; a delay violation probability bound Z; a buffer overflow probability bound W; a delay jitter bound J.

13 If Z is 1, a channel is said to be deterministic, otherwise it is referred to as statistical. A guarantee on the bandwidth is obtained from the traffic specifications, because the network promises to absorb the offered load as described by the client, within the agreed delay bound. The Tenet group has developed a Protocol Suite which implements the forementioned approach. In this suite, the channel setup process has been implemented in the Real-Time Channel Administration Protocol (RCAP). The setup process starts when an application invokes an establish_request primitive, which contains performance requirements, traffic characteristics and addressing information. As a consequence, a setup request message is issued by the source endpoint. This message travels along the connection path, through a number of intermediate nodes (switch, router, gateway), causing several admission tests and computations to be performed at each node. If the new channel passes all the tests in a node, the message is forwarded, with some state information about the current node, to the next node on the route. In case of failure of any test in a node, a reject message is sent back to the source end-point, by releasing the reserved resources in each intermediate node. The destination end-point (which is presumably a workstation) is the last point along the path where the acceptance or reject decision for a channel request can be made. This decision is transmitted back to the source by means of a reply message, along the channel s route. When a node is revisited during this message return trip, the resources previously reserved are either committed or released. The tests performed at each node are concerned with the availability of sufficient bandwidth in the links, as well as processing power and buffer space in the node. These tests are based upon worst case analysis of the scheduling mechanisms and of the traffic specifications, so that after the setup process the network can a priori guarantee the required performance. In [3] and [6] the admission tests are specified, assuming a Multiclass D-EDD scheduling policy is adopted in the network nodes. The characteristic of the Tenet approach of defining the scheduling policy to be adopted in the intermediate nodes is also its major drawback. The current trend in the IETF is just the opposite; a set of communication services have been defined and the implementation of these services is left to product manufacturers. According to this view, only a standard signalling mechanism is necessary, while even the admission control policies are considered implementation specific, and so they are left out of the standardization process. 10 Conclusions Admission control remains probably one of the most controversial research topics in the area of networking. The appearance of new, demanding applications requiring communication services suited for continuous media such as digital video and audio has indeed highlighted the limits of many of the existing network architectures. However, it is still unclear how to proceed in order to fill the gap between the applications'requirements and the network capabilities. One of the possible approaches to this problem is to design network architectures capable of reserving

14 resources for the existing communication sessions. In this case, if we provide the network with adequate information about the expected load induced by any new service request, the network itself will be capable of computing the amount of resources needed to manage the additional traffic. To offer communication services with guaranteed performance, the network will accept new requests only if these will not jeopardise the existing connections. In this paper we have tried to show some of the problems related to the adoption of such approach in the design of new network architectures and services. In fact, the capability of controlling and reserving resources, does not come at no cost, in terms of complexity, reliability, and generality of the solution itself. We have therefore presented three solutions to this problem that differ in the level of the control that can be applied on the architecture and, consequently, on the strength of the quality of service guarantees that can be offered to the applications. In spite of their differences, from these proposals and from the others that were presented in the last years, it appears that the solutions adopted for the provision of integrated communication services in future network architectures will affect not only the services and the applications that will be available, but also the way users will have access to them. References 1. R. Braden, D. Clark and S. Shenker. Integrated Services in the Internet Architecture: an Overview. Internet RFC 1633, D. Clark, S. Shenker and L. Zhang. Supporting Real-Time Applications in an Integrated Services Packet Network: Architecture and Mechanism. Proc. of ACM SIGCOMM 92, pp 14-26, D. Ferrari. Real-Time Communication in Packet-Switching Wide Area Networks. Technical Report TR , International Computer Science Institute, Berkeley, California, D. Ferrari, A. Banerjea and H. Zhang. Network Support For Multimedia. A Discussion of the Tenet Approach. Technical Report TR , International Computer Science Institute, Berkeley, California, D. Ferrari, J. Ramaekers and G. Ventre. Client-Network Interactions in Quality of Service Communication Environments. Proceedings of HPN92, 4th IFIP Conference on High Performance Networking, Liege, Belgium, pp. E1/1-E1/14, D. Ferrari and D. Verma. A Scheme for Real-Time Channel Establishment in Wide-Area Networks. IEEE JSAC, vol.8 no. 3, pp , April S. Floyd. Comments on Measurement-based Admission Control for Controlled-Load Services. (Draft paper available on the World-Wide-Web, at the URL: ftp://ftp.ee.lbl.gov/papers/admit.ps.z), Lawrence Berkeley National Laboratory, Berkeley, California, S. Keshav. An Engineering Approach to Computer Networking. Addison Wesley, B. Leiner. Critical Issues in High Bandwidth Networking. Internet RFC 1077, C. Partridge. Gigabit Networking. Addison Wesley, G. Ventre. Real-Time Interaction Issues in Distributed Multimedia Computing. GRID Internal Report, 1995

15 12. D. Yates, J, Kurose, D. Towsley and M.G. Hluchyj. On per-session end-to-end delay distributions and the call-admission problem for real-time applications with QOS requirements. Proceedings of ACM SIGCOMM 93, H. Zhang and D. Ferrari, Improving Utilization for Deterministic Service in Multimedia Communication, Proceedings of 1994 International Conference on Multimedia Computing and Systems, pp , H. Zhang and S.Keshav. Comparison of Rate-Based Service Disciplines. Proceedings of ACM SIGCOMM 91, pp , H. Zhang and E. W. Knightly. RCSP and stop-and-go: a comparison of two non-workconserving disciplines for supporting multimedia communication. ACM/Springer-Verlag Multimedia Systems Journal n.4, pp , 1996