Packet Scheduling Based on Learning in the Next Generation Internet Architectures

Transcription

1 Packet Scheduling Based on Learning in the Next Generation Internet Architectures Alencar Melo Jr. USF - São Francisco University, CCET, Brazil alencar@nedprof.usf.com.br Juan Manuel Adán Coello PUCCAMP Pontifical Catholic University of Campinas, Informatics Institute, Brazil juan@ii.puc-campinas.br ABSTRACT With multimedia applications, the Internet traffic increased a lot in volume, but also suffered major alterations in its nature, requesting other network services besides the current best-effort service. To satisfy this new situation, network architectures QoS-oriented such as the Integrated Services and Differentiated Services architectures have been discussed in several IETF groups. In this work we propose a methodology to implement packet schedulers, an element of fundamental importance to the new services, based on fuzzy control. The main innovation of this proposal is the employment of rule induction algorithms as C4.5 to infer the fuzzy controller s main rules. The approach has several advantages over the current ones, specially flexibility, efficiency and the possibility of being applied in several environments. 1. Introduction Internet architecture and main protocols have had their current features set about twenty years ago. Early applications, such as remote terminal, file transfer and e- mail deal essentially with discrete data, not requesting rigid temporal restrictions. The protocols just assure the fidelity, basically using retransmission mechanisms. The main characteristics of the current TCP/IP protocols include: the supply of only one best-effort service class; admission control mechanisms do not exist, the applications do not need to obtain previous permission to begin the transmission of their packets; they do not offer delay warranties in packet delivery. The enormous success of the IP protocol is due mainly to its simplicity, which allows it to run virtually on any communication subnetwork, from technologies notconnection-oriented like Ethernet to technologies connection-oriented like ATM. The connectionless philosophy characterizes the IP protocol; the main functional features of the protocol are concentrated in the final systems, leaving the intermediate systems with few attributions. However, the IP protocol supplies only one class of service for packet delivery, called best-effort. An application that uses best-effort service is subject to unbounded delays. The last years have witnessed a trend of isolated and specialized networks, such as the telephony, television and radio networks gathering around Internet and causing an impressive traffic increase. Internet traffic increased a lot in volume and also suffered great alterations in its nature. With multimedia applications computers started to process audio and video, which are continuous media. Implementations of multimedia applications when developed on the current Internet do not work in an appropriate way, due mainly to the great delays involved. The main differences between discrete data and continuous media can be seen in the Table 1. In section 2 and 3 we discuss some aspects of quality of service and the nature of multimedia applications. Integrated and Differentiated Services Architectures are described in section 4. General aspects of packet scheduling are discussed in section 5 and in section 6 the proposal to design packet schedulers based on learning is presented. Finally, we present our conclusions. 2. Quality of Service The requirements of the communication services for multimedia applications exceed the current data-oriented communication services. The term Quality of Service (QoS) is used to characterize the ability of any network element to provide its users with a service that presents a superior performance than the current best-effort service, taking into consideration the typical QoS parameters such as bandwidth, delay in packets delivery, jitter (variation of the delay) and error rate. In almost all multimedia applications low delays are fundamental.

2 Table 1.. Discrete data X continuous media Discrete Data Continuous Media Bandwidth low high Sensibility to Delays little sensitive (asynchronous applications) very sensitive (isochronous applications) Error Rate no loss some loss (detection and recovery) Traffic Pattern bursty (without superior limit of bandwidth use) continuous flow (with superior limit of bandwidth use) Type of Communication point-to-point multipoint Synchronism none synchronized transmission (audio and video flows) A central concept in QoS is the capacity of the network to differentiate between distinct types of traffic or service, so that the users can have more than a single class of service (CoS), making possible the choice of the most appropriate services for each application [1]. CoS implies that different types of traffic can be classified and managed differently through the network. The term QoS has a broad connotation, including CoS, admission control, policing and traffic reshaping, packet scheduling, congestion control, routing etc. The word guarantee in QoS is controversial; many advocate that it is practically impossible to guarantee a consistent and previsible performance with no loss in packet-switched networks [1]. To serve the communication requirements of multimedia applications with the current best-effort infrastructure we have basically two options: to maintain the utilization level extremely low, which implicates in overprovision the whole network; this alternative is not economically viable now, in spite of some authors to consider that in the future the bandwidth can be free [1]; to turn the applications adaptive to the current load of the network, in other words, to alter the applications instead of modifying the current infrastructure. Adaptive techniques have been used with success in several applications for the Internet but, for applications that need stringent guarantees, such as telemedicine, they are not enough. Therefore it is necessary new network architectures that can differentiate applications and serve them according to their needs with different classes of services in an efficient way. The decision of altering the current best-effort architecture, providing new services, still generates some controversies. However, in [3] it is shown that network architectures capable of doing resources reservations bring significant benefits compared to besteffort architectures in several circumstances, no matter how cheap the bandwidth becomes in the future. The nature of multimedia applications for which the new network architectures are being developed is discussed in the following section. 3. Nature of Multimedia Applications The vast majority of distributed multimedia applications are of the type playback real-time [4]. In these applications a source takes some signal, packetizes it and transmits it over the network. The network introduces some variation in the delay of the delivered packets, causing jitter. The receiver unwraps the data and tries to reproduce the signal in a faithful mode, presenting the signal at some fixed offset delay from the original departure time. The performance of a playback application can be measured along two dimensions: latency and fidelity (faithful, distorted or incomplete presentation). Applications that involve interaction among the two ends of the connection such as telephony are more sensitive to latency; other applications such as movie transmission are less sensitive to latency but they require a larger bandwidth. The jitter has great importance in the performance of playback applications because the buffer size required for a connection is directly proportional to the jitter. According to the sensibility to fidelity loss, real-time playback applications can be classified into intolerant or tolerant. Intolerant applications do not admit fidelity loss and use a fixed offset delay larger than the maximum delay, avoiding late packets. The performance is independent of packet-arrival times as far as they arrive within the limits of delay. Tolerant applications admit some late packets, not needing to use an offset delay larger than the maximum delay. They can try to reduce their latency by a dynamic adjustment of the offset delay, observing the actual delays in a recent past. These applications require a packet delivery service that tries to improve the distribution of delays as a whole and do not need delay guarantees for individual packets. According

3 to [4], the great majority of future playback applications will be of the tolerant type. Applications that involve remote control of processes, stocks market etc. also require a reliable superior limit of delay but they are not multimedia applications. Non-real-time or elastic applications such as ftp use immediately the data rather than buffering and they always choose to wait for a late packet than continue without it. The performance depends more on the average delay than on the tail of delay distribution. Next, two main proposals under discussion at IETF are described so that the Internet can offer different levels of QoS to the users. 4. Quality of Services Architectures QoS should be supplied by flow, where a flow is an abstraction of a sequence of related packets that result of an activity and request the same QoS. A flow can have several sources and receivers. In general, for a flow to receive end-to-end QoS, path alterations or flaws of routers should not happen, the pattern of traffic should respect the descriptors previously accorded and the packets should not suffer fragmentation. The new proposed architectures present two fundamental alterations in relation to the current besteffort architecture: the applications have abilities to reserve network resources, mainly bandwidth and the network controls these resource reservations, mainly through admission control [5]. The admission control can be accomplished in an explicit mode, making use of a signaling protocol, or in an implicit mode, through the conditioning of traffic in the domains borders. The Quality of Services Architectures can be classified in two types: quantitative QoS: they are the architectures that make explicit reservation of resources for each flow produced by the applications, also called microflows. The state of the flows inside the network is maintained during all their life cycle, making use of signaling protocols; an example of this type of architecture is the Integrated Services (IntServ) architecture [6]; qualitative QoS: they are the architectures that simply give priority to the packets of applications that have more rigorous QoS requirements and do not make reservations for individual micro-flows. In the Differentiated Services (DiffServ) architecture [7], the packets belonging to the individual micro-flows are classified in the border of the network in aggregated flows, which can receive different levels of services. The packets of a micro-flow possess common source address, destination address and port number. In the DiffServ architecture, several micro-flows can be grouped into one class of service, sharing the resources associated to the class, but losing the isolation among them. The IntServ architecture maintains the micro-flows isolated, with resources reservations specific for each one, making possible to offer services which present a reliable superior limit of delay (guaranteed service), but its scalability over large networks is limited. The DiffServ architecture, due to aggregation of flows, cannot offer guaranteed service. In [4], other problems related to the isolation and sharing of resources reservations for different flows are discussed Integrated Services Architecture The Integrated Services architecture [6] intends to extend the current functionality of IP architecture, supplying other services besides the best-effort, in order to accomplish the QoS requirements of real-time applications. The new proposed services are the guaranteed service [8] suitable for intolerant real-time applications and the controlled-load service [9] suitable for the requirements of adaptive real-time applications. The guaranteed service supplies the following guarantees to the packets in conformity with a token-bucket filter [4]: superior limit to the end-to-end delay, bandwidth and no packet loss due to buffers exhaustion. The applications inform the routers about the traffic characteristics of their flows and the resources reservations required through the signaling protocol RSVP [10]. Each router along the path allocates, for each admitted flow, a certain bandwidth and buffer space. The controlled-load service provides a service to data flows that is almost equivalent to the best-effort service under an unloaded network, using admission control to ensure that the QoS will not deteriorate with the increase of the network load. Besides the new services and the signaling protocol, the other basic components of the IntServ architecture are the packet classifier that maps each received packet into one class (all the packets of a same class receive the same QoS); the packet scheduler that schedules the packets belonging to the classes, so that the QoS requirements can be satisfied; and the admission control that implements a decision algorithm that the routers use to determine if a new flow can have its requested QoS guaranteed without violating the guarantees of the previously accepted flows. The signaling protocol RSVP allows senders, receivers and routers of a unicast or multicast session to communicate with each other so that the necessary state in the routers to support the guaranteed and controlledload services can be established. In the IntServ architecture, the traffic of a flow is characterized mainly by a token-bucket filter (r, b), where r is the rate of tokens per second to supply the bucket and b is the bucket capacity. When a packet is transmitted some tokens should be retired from the bucket in

4 accordance with the size of the packet. The token-bucket filter allows controlled bursts, limited by the amount of available tokens in the bucket. The policing of flows is made in the borders of network, comparing the traffic of a flow with its descriptor to verify conformity; packets that are not in conformity are forwarded as best-effort packets. It can be necessary to do traffic reshaping in internal routers in order to soften distortions introduced along the path, readjusting the flows to their token-bucket filters. The applicability of the IntServ architecture and of RSVP in long distance networks as Internet has its constraints. Due to the necessity of maintaining the state of each flow inside the network, the solution becomes scarcely scalable. Gradual development of this architecture is difficult because the routers need to suffer major modifications and the applications need to generate RSVP signaling. Therefore, in the short term it is more attractive to use scalable architectures that offer QoS to the users in a softer way, without imposing major modifications in the current architecture Differentiated Services Architecture In the Differentiated Services architecture [7], packets of individual micro-flows are classified in the border of the network to one of a small number of aggregated flows, by marking (or remarking) the packets headers according to pre-determined policy criteria, established in a service level agreement (SLA) between networks that share a border. The packets can also be marked prior to entry into a DiffServ domain, by the sender host. Each aggregated flows receives in the inner routers one class of the available services in the domain, which possess different priorities. Traffic in disagreement with its profile cannot obtain the desired QoS and/or incur in extra cost. The traffic profiles are derived from the SLA. This architecture achieves scalability by implementing complex classification and conditioning traffic functions only at network boundary nodes and then apply a peraggregated flows service in the middle of the network. In the inner routers, the packets are simply forwarded according to the class of service indicated in their header. DiffServ can be employed over large backbones and the existent applications are benefitted because it is not necessary an end-to-end signaling protocol as RSVP. In the header of each packet, the DS field denotes the class of service (PHB) that the packet should receive at each hop as it is forwarded through the network [11]. The DS field is defined through the Type-of-Service (TOS) octet in IPv4 or through the Traffic Class octet in IPv6. The PHB (Per-Hop Behavior) defines the service offered by a router to a behavior aggregate (BA), that is an aggregated flows whose packets have the same markings in the DS field. PHBs constitute a granular way to allocate bandwidth and buffer resources in each router and they are implemented by packet scheduling and buffer management algorithms. Inside the DiffServ domain the packets can receive one of the PHBs that are being now considered in IETF: Expedited Forwarding (EF) [12], that minimizes delay and jitter, providing the highest QoS level and that discards any traffic that exceeds the defined traffic profile, or Assured Forwarding (AF) [13], that defines four relative classes of services and each class of service supports three levels of drop precedence. Out of profile AF traffic is not delivered with as high probability as the traffic in profile, which means it may be demoted but not necessarily discarded. The IntServ and DiffServ architectures can be seen as complementary because they could be used together in different contexts in order to supply the users with an end-to-end QoS. In [14] a framework is proposed in which RSVP signaling is used for dynamic provisioning and admission control in DiffServ regions, located in the core of the network, where the scalability is vital. In the borders of the network, near to the applications, the IntServ can be used, making the provisioning for each micro-flow. The packet scheduler is a vital component in any QoS architecture, being fundamental to assure the QoS requirements of the flows. In the following sections, some aspects of packet scheduling are discussed and it is presented a novel proposal to design schedulers, using techniques based on machine learning and fuzzy control. 5. Packet Scheduling Packet scheduling only becomes an important problem when statistical fluctuations in traffic sources cause considerable queues in routers. Otherwise the FIFO algorithm is a good solution to the problem. A scheduling algorithm has two orthogonal tasks: to decide the order in which the requisitions are served and to manage the queue of pending requisitions. A scheduler can influence the delay (and consequently the jitter), bandwidth and loss rate. When designing a packet scheduler, the following degrees of freedom (fundamental choices) should be considered [15]: classes of services: the number of classes of services and the relationship among them should be determined; work-conserving versus non-work-conserving: a work-conserving class of service never allows the link to remain idle when there are packets awaiting to be transmitted, while a non-work-conserving class of service can let the link to be idle even when it has packets to serve with the purpose of traffic reshaping;

5 degree of flows aggregation: the scheduler may have a queue per flow, a queue for all flows (same QoS for all) or a queue for each different class of service; service order: it is the order in which the scheduler serves the packets from a class of service. The scheduler can use a non-reordering (FIFO algorithm) or a reordering strategy; reordering may cause a considerable overhead to compute the packet labels. Ideally, a scheduler should be scalable and have a low complexity in order to support a large number of flows in high-speed links, like OC-48 (2.4 Gbps); it should also provide protection so that a minimum bandwidth is guaranteed for each flow, independently of the behavior of other flows and provide fairness, supplying a fair allocation of resources for each flow, according to their needs [15]. In the next section, the proposed approach to design packet schedulers in QoS architectures is presented. 6. Packet Scheduling Based on Learning Current approaches for packet scheduling usually require a great amount of on-line calculations, thus making unfeasible their use in high-speed links; they also attempt to satisfy a single QoS parameter, such as maximum delay or low jitter. The main objective of this work is to delineate a proposal for packet scheduling that makes use of methods based on machine learning and fuzzy control. Machine learning is concerned with the construction of computer programs that improves automatically its performance on the same task through experience, using techniques such as artificial neural networks, decision trees etc. [16]. Fuzzy control is based on a relatively simple idea of fuzzy set, that is a generalization of an ordinary set by allowing a degree of membership for each element [17][18]. A membership degree is a real number on the interval [0, 1]. The employment of fuzzy control in packet scheduler s construction is appropriate, mainly because of the following reasons: fuzzy control allows decision making in situations where the available information is incomplete or imprecise; when using connectionless protocols as the IP protocol this feature may be very useful; several problem variables do not have borders clearly delineated, such as the resource utilization (bandwidth/buffer) in the router, that can be characterized as very high, high, medium and low; the scheduling decisions may be well described by heuristic rules such as: If (class of service s buffer utilization is high and fairness bandwidth is low) then (class of service transmission priority is high) the mathematical model for the system behavior is very complicated; the rules can regard several QoS parameters; the fuzzy control rules are usually simpler and easier, often requiring a few rules, providing efficiency and robustness. The main deficiency of pure fuzzy systems is their incapacity for learning, because they have no memory. This can be solved associating fuzzy systems with techniques that perform pattern recognition such as rule induction [16]. A packet classifier in QoS architectures accomplishes basically a mapping task, receiving packets belonging to several flows and gathering them in queues according to the performance requirements, resources utilization and degree of flows aggregation. Different implementations of the same packet delivery service may use internally different structures of queues. The process of queues identification is fundamentally a clustering task and it may be done by unsupervised learning methods as the Kohonen s maps [19]. Once known the queues, a rule induction algorithm as C4.5 [20] can be used to induce a packet classifier structured in the form of rules. After the packets have been classified in different FIFO queues, the packet scheduling problem is reduced to the selection of one of the non-empty queues and to the transmission of the packet located at its head. This procedure really simplifies the problem because there is no packet reordering inside each class. The scheduler can be implemented as a fuzzy controller with inputs like flows traffic descriptors, resources utilization level, fairness degree and other environment features. The output defines the queue from which the next packet will be transmitted. The Figure 1 outlines the proposal for packet scheduling. The main innovation of this proposal is the employment of rule induction algorithms as C4.5 to infer the fuzzy controller main rules according to the steps below: 1. an appropriate network model is chosen, defining the routers, hosts and links characteristics; 2. IP packets are injected in a random packet scheduler. After a packet is scheduled (or not, in the drop case), it can be labeled with one or more of the following attributes: queue delay, packet dropped or not etc. The information about scheduled and unscheduled packets together with the other current network attributes is stored; 3. using the information collected in step 2, it is used C4.5 to induce decision rules; 4. from the obtained rules, the most effective ones are selected to be used in the fuzzy controller; 5. the initial rules may be refined and new rules may be added to the controller using human expert-based methods.

6 Figure 1.. Fuzzy scheduler based on learning The fuzzy scheduler main rules are obtained starting from the decisions taken by a random scheduler during an appropriate time interval, using a rule induction algorithm. Behind each decision taken in an aleatory way by the random scheduler, there are rules more or less effective for the problem. The rule induction algorithm can discover the most useful ones. The obtained rules are easy to understand and can be adapted by human knowledge. This cannot be easily accomplished when using black-box learning methods such as neural networks. Once obtained a packet scheduler, this can be evaluated by scheduling a new packet set with the fuzzy scheduler and the random scheduler and then comparing the performance of both according to the considered QoS parameters. 7. Conclusions In this paper, we have presented a novel proposal to design packet schedulers in QoS Internet architectures, using techniques based on machine learning and fuzzy control. The main innovation of this proposal is the employment of rule induction algorithms to infer the fuzzy controller main rules. Once obtained a fuzzy packet scheduler, this can be used in different environments with different traffic patterns. This only requires a calibration of the linguistic terms contained in the fuzzy rules or, in other words, it only needs a tuning the membership functions. Data Mining tools may be used to explore the available data, finding the critical points in the fuzzy controller input and output spaces. The membership degree at the critical points is 1.0 and suitable membership functions can be obtained by dropping the belief value to zero in an appropriate way. The proposed methodology will be employed in future works in order to evaluate the applicability of the techniques based on machine learning and fuzzy control in packet scheduling, a very complex problem for which it is difficult to obtain a sound mathematical model. 8. References [1] Ferguson, P., and G. Huston, Quality of Service, John Wiley and Sons, New York, [2] Negroponte, N., Being Digital, Knopf Publishing, [3] L. Breslau, and S. Shenker, Best-Effort versus Reservations: A Simple Comparative Analysis, ACM Computer Communication Review, vol. 28, nº 4, October 1998, pp [4] D. D. Clark, S. Shenker, and L. Zhang, Supporting Real- Time Applications in an Integrated Services Packet Network: Architecture and Mechanism, Proceedings of SIGCOMM 92, pp [5] S. Shenker, Fundamental Design Issues for the Future Internet, IEEE Journal on Selected Areas in Communications, vol. 13, nº 7, September 1995, pp [6] R. Braden, D. Clark, and S. Shenker, Integrated Services in the Internet Architecture: an Overview, RFC 1633, June [7] S. Blake, et al., An Architecture for Differentiated Services, RFC 2475, December [8] S. Shenker, C. Partridge, and R. Guerin, Specification of Guaranteed Quality of Service, RFC 2212, September [9] J. Wroclawski, Specification of the Controlled-Load Network Element Service, RFC 2211, September [10] L. Zhang, et al., RSVP: A New Resource Reservation Protocol, IEEE Network, September 1993, pp [11] K. Nichols, Definition of the Differentiated Services Field (DS Field) in the IPv4 abd IPv6 Headers, RFC 2474, December [12] V. Jacobson, K. Nichols, and K. Poduri, An Expedited Forwarding PHB, RFC 2598, June [13] J. Heinanen, et al., Assured Forwarding PHB Group, RFC 2597, June [14] Y.Bernet, et al., A Framework for Integrated Services Operation over Diffserv Networks, Internet Draft, draft-ietfissll-diffserv-rsvp-03.txt, September [15] Keshav, S., An Engineering Approach to Computer networking: ATM Networks, the Internet, and the Telephone Network, Addison-Wesley, Reading, [16] Mitchell, T.M., Machine Learning, New York, McGraw- Hill, [17] T. Munakata, and Y. Jani, Fuzzy Systems: An Overview, Communications of the ACM, vol. 37, nº 3, March 1994, pp [18] Gomide, F., and Pedrycz, W., An Introduction to Fuzzy Sets - Analysis and Design, MIT Press, Cambridge, [19] T. Kohonen, Self-Organized Formation of Topologically Correct Feature Maps, Biological Cybernetics, 43, 1982, pp [20] Quinlan, J.R., C4.5: Programs for Machine Learning, Morgan Kaufmann, 1993.