Quality of service management in switches by means of intelligent agents

Size: px
Start display at page:

Download "Quality of service management in switches by means of intelligent agents"

Transcription

1 Quality of service management in switches by means of intelligent agents Antonio Barba Martí Dept. of Applied Mathematics and Telematics. Escuela Técnica Superior de Ingenieros de Telecomunicación Universidad Politécnica de Catalunya (UPC). c/ Jordi Girona 1-3, Module C-3, Campus Nord, Barcelona, SPAIN Abstract Broadband area networks require of new resource management mechanisms for switching nodes in order to provide quality of service in real time. A specific switch has been designed in order to support the operation of a cooperative multi-agents infrastructure. This design is developed as a solution with the purpose of providing an optimum quality of service, according to different types of traffics, to the suscribers Introduction In this article a network management architecture with the aim to provide the best quality of service to the user is designed. A switching architecture supports an intelligent multi-agents infrastructure in order to reduce the congestion situations in networks [1]. The intelligent agents can be defined as autonomous software entities that allow to interact according to an intelligence by itself. These mobile agents represent active or portable objects moving from a switch to others to access remote resources or even to find other agents and to cooperate with them. In this case, the distribution of management tasks is known as delegation management. This paradigm uses a descentralized management that allows to offer an important increment in processing resources in the network nodes. By means of this architecture the network bandwidth grows together with the number of managed nodes. 2. Multi-agent architecture for the quality of service management The switches have been designed with the objective to process different types of traffic. The proposed mechanism of traffic management is based on the use of predictive agents migrating on the distributed structure of the network nodes. This system provides an admision control of connections that allows the statistics multiplex of multiple calls according to a specific quality of service for each call. Management Center Agent Domain Level System Level Agent Agent Switch 1 Switch 2 Fig. 1. Scheme of a multi-agent architecture in a network 1 This work has the support of the spanish CICYT: TEL

2 The management process developed by the multi-agent structure have four different steps. First step, each node monitoring the relevant management information periodically by means of the agent. In the second step, this information is sent to the management center for an intelligent processing. As a consequence of this process, a serie of decisions about the network elements are adopted. Using the agents delegation, the agents are sent following the connection paths through the switches until the end terminals. Finally, the defined decisions adopted in the managed switches are applied to the network. The advantage of the scheme based on mobile intelligent agents basically takes the benefit from the descentralization of processing tasks from the management center to the managed nodes, using these agents for this function. This aspect is more important according to the growth of the network to manage. Furthermore, an effective management delegation requires a cooperation among agents with the objective to know the complete status of the network in real time. This thing allows to act in congestion or fault situations with a high speed. 2.1 Network architecture The framework of the traffic and congestion controls is defined taking into account the following network architecture. TRANSPO NETWORK ACCESS ACCESS TRANSPO NETWORK ACCESS TRANSPO NETWORK TRANSPO NETWORK ACCESS Fig. 2. Network architecture. The system offers two different networks. One of them is a access network with direct connection with users. The communications would be multiplexed with a total bit rate of 155 Mbps. The another one, a transport network connects nodes by means of links, to 622 Mbps between nodes of the access and fixed network or 2'5 Gbps between nodes within the transport network. The nodes can process different kinds of traffic explicitly and with different priorities. In this case a structure of M/D/1 queues with different priorities depending on the traffic classes is defined. The node can process, N, (agents) traffic and signaling (e.g. RM cells) as a special type of traffic.

3 2.2 Model of nodes Both kind of nodes, access and transport nodes, use M/D/1 queues in the three different stages of information processing. This queue model is the most appropiate in ATM networks. In this case the size of the cells is fixed and deterministic the type of service. In the first stage the information cells are routed to different queues and processors depending on the type or priority of traffic. In this stage the headers of the cells are also processed according to the congestion algorithm. In the second stage occurs the switching of cells and the processing of the agent algorithm due to signaling and monitoring information cells. In the third stage the information cells are assembled and routed to the established link. The type of traffic in the input connections is a Markov Modulated Poisson Process (MMPP) [4]. The traffic in the and N connections is a result of a stadistic multiplexation of a set of Poisson sources. In case of N, the signal is the addition of M identical minisources with two states (on, off) and it has a mean rate of: E( λ) = MpA, where p is the probability that a minisource is in on state and A is the compression factor of the video signal in bits/pixel. This thing allow to dispose of a variable rate in order to simulate a N or traffic burst. In case of a M/D/1 queue, the mean number of cells in queue, inclusively the service one is as follows: En ( ) = ρ 1 ρ ( ) ( ) 1 ρ 2 and the mean delay in the system: ET ( ) = 1 µ 1 ρ ( ) ( ) 1 ρ 2 where ρ = µ λ is the traffic intensity, lambda is the mean rate of cells arriving to the queue and time of service in a cell. Both expressions are related by the Little formula: E(n) =λe(t) 1 µ is the mean In the defined network models a priority system has been used. It has been assigned a top priority to the traffic due to the transmission is in real time, interactive and time critical. The other types of traffic (N and ) have less priorities because the communications are not in real time. In the stages where the three types of traffic use the same processor, the cells with less priority are not served until the cells with the highest priority have been processed. Finally, it has been considered the (agent) traffic as a special signaling traffic. With these hypothesis and taking into account the M/D/1 queue formulas the following expressions have been formulated: λ C = a 0 λ 0 + a 1 λ a n λ n λ V = b 0 λ 0 + b 1 λ b n λ n λ a = c 0 λ 0 + c 1 λ c n λ n λ S = s 0 λ 0 + s 1 λ s n λ n The λ 0,λ 1...,λ n are the input traffics in the access node. The λ c, λ v,λ a,λ s are the different types of traffic classes. λ Ti : Total bit rate at the entrance of the third stage in the node with "i" output. This delay is the addition of the mean delays in the three stages. Repeating the same in the other cases of traffic is obtained:

4 The mean number of cells in every stage is obtained multiplying the mean delay in every stage and the mean bit rate at the entrance of this stage (Little formula). Doing the same in the fixed network, the following expressions for conmutation delay are obtained: The mean number of cells in the system is obtained such as the previous case. a0 u1 u11 x y N, b0 N u2 N u22 z c0 u33 N, s0 u3 4 N, Fig. 3. Scheme of a access node. The design of the nodes requires also the definition of the size of the buffers and the speed of the processors. The design has been done taking into account a determined scenario of traffic. This scenario works with a 50% of Rt traffic, 40% of N traffic and 5% of user (Agents) traffic. In this case, the signaling charge is similar to the traffic [2, 3].

5 a0 b0 c0 a1 b1 u1 u2 u3 u1 u2 N u11 u22 i k j c1 u3 u33 4 Fig. 4. Scheme of a transport network node. 3. Cooperation among intelligent agents The agents act in the switch to establish an adequate quality of service testing differently each type of traffic. Four classes of traffic have been defined; that is, N, signaling (SIG) and classified according to the priority level. The agents will adopt different management policies according to quality of service requirements defined for each traffic class and priority level. Furthermore, a management policies in the scenario of management nodes in the network can be defined. That is, a P1 policy where the agent of the node only can use a b percentage of the common available space not used by the rest of services in addition to the maximum size assigned to the service. And similarly, a P2 policy is defined with the purpose of a management agent may reduce the size of the buffer in a factor a. The types of cells used on the agents communication among switching nodes are of the type OAM F4 (Operations, Administration and Maintenance) of performance management with types of functions like monitoring or reports. These cells have an restricted actuation in each management domain, they have the same path but different virtual channel than the user cells. Its emission period is similar to the ring way of the cell in each domain. de gestión de rendimiento con tipos de funciones de monitorización y de reports. Estas celdas tienen un ámbito de actuación restringido a cada dominio de gestión, tienen un mismo camino pero distinto canal virtual que las celdas de usuario. Su período de emisión si bien tiene en cuenta la situación de congestión de los nodos, es de un orden de magnitud aproximado al Delay de circulación de la celda por cada dominio. 3.1 Managed node agents The switching nodes of the transport network can be modelled by means of specific buffers according to the supported kinds of traffic. In the same sense, a memory buffer of common cells with the end of reducing the information losses to the minimum in case of buffer overflow or network congestion.

6 In the figure 5 can be seen a buffer with variable threshols in function of the charge of cells in the input of the node. a*ed b*edbc LMB NA Servicio ED EDBC Espacio reservado por otros servicios: +N+ Fig. 5. Scheme of the node buffer of the transport network with the threshold levels according to a congestion situation. The results of the operations that perform the agent of the managed node are transferred to the agent of the manager node by means of intelligent agents and with a periodicity depending on the priority specified by the management center. The structure of the agents can be deployed easily with specifications defined in stardards like IETF, OMG and the ATM Forum. The parameters sent as information with the agents are the following (in case of the traffic ): - Delay: Delay of medium service acording the mean length of the buffer (LMB) and the TASA. - Tmaxsum: increment in b*(nº slots of available packets /TASA). - Tmaxres: decrement in a*[(na-lmb)/tasa]. - Trend: trend = d(ed)/dt. being: - Mean length of the buffer (LMB). - Available space in the common buffer (EDBC) with (a, b < 1). - Tasa of the port of output (TASA). - High threshold of the buffer (NA). The use of the parameter Tmaxsum is defined in the management policy P1 while the parameter Tmaxres would be specified in the management policy P2. Furthermore, the performed measurements are defined periodically according to different service delays. 3.2 Transport network management by agents Together with the resource optimization policies of the nodes (P1, P2) are integrated other kind of policies that tries to improve the quality of service in each connection of the network and taking into account the global status. In this sense the manager agent relates the switching nodes with the requirements of quality of service specified by the suscriber. The agent of the management node tries to apply the management policies corresponding to the rest of network nodes. The policies more oriented to services are the following: P3 policy by means of which the manager agent distributes the credit of service time among the switches. A P4 policy by means of which the node more congested will receive more credits than the node less congested. If it is necessary, the NA level is decremented in order to maintain the same transit delay of the packet (P) specified in the setup phase of the connection. Finally, a P5 policy in every domain where the maximum transit delay experimented by a packet crossing the domain would be restricted to the max P specified in the setup connection. The management center performs an intelligent control by means of the defined agents. The system does a distributed jerarquic infrastructure of agents. In each node the agents perform the control of the thresholds, and configuration parameters of the buffers in order to send the information messages to the user. Now, different intelligent agent algorithms are explained. A cost function is defined taking into account the congestion levels in all the nodes of the network. For this purpose the agent collect the congestion levels of the bufferscorresponding to the different types of traffic (N,, ). The following terms are defined:

7 f(cnmi): cost function of the switch i (CNMi). d(ed)/dt, d(edn)/dt, d(edsig)/dt, d(ed)/dt: trends corresponding to the buffers of different types of traffic, N, Signaling and respectively. Tmaxsum, TmaxsumN, TmaxsumSIG, Tmaxsum: Increments in b*(nº slots of available packets/tasa) for the different types of traffic and buffer. Tmaxres, TmaxresN, TmaxresSIG, Tmaxres: Decrements in a*[(na-lmb)/tasa] for each different type of traffic and buffer. Delay, DelayN, DelaySIG, Delay: Delays for each different type of traffic and buffer. Then, it is defined a generic cost function: f(ci) = K1*f() + K2*f(N) + K3*f(SIG) + K4*f(). with 0 < K1, K2, K3, K4 < 1 being each of the trends defined such as f() = d(ed)/dt The managed node acts providing the following parameters, in this case for the traffic : Delay_CNMi = LMB/TASA Tmaxsum_CNMi = b*(nº of free packets/tasa) If LMB < NA then Tmaxres = a*[(na-lmb)/tasa] if not Tmaxres = 0. f() = d(ed)/dt The managed node updates the NA threshold to the reception of each agent being: NA = NA+ t*tasa In relation to the agent of the management node, continuously would do the following operations for each buffer integrated in the nodes. Furthermore the global delay of the CTD connection is limited particularly for the traffic: If MaxP > Delay_CNM1 + Delay_CNM2 then /*Policy P5*/ Case f(_cnm1) < 0 y f(_cnm2) >0 do Policy P4 f(_cnm1) > 0 y f(_cnm2) < 0 do Policy P4 f(_cnm1) > 0 y f(_cnm2) >0 do Policy P3 The manager agents would do complex processes taking into account the information presented by the agents of the nodes according to the function f(cnmi). Nowadays we are testing the impact of this kind of agent algorithms in the functionment of an advanced managed network since a global point of view. 4 Traffic management by agents in the nodes The proposed traffic management mechanisms are based in the use of predictive agents distributed in the network nodes. These nodes support agent algorithms with diverse threshold conditions according to the traffic levels of the network. The purpose of this traffic control is to predict and in consequence to reduce the congestion.

8 The agents are integrated in an intelligent network that tries to predict traffic and to take measures in order to anticipate a congestion situation in the network. The agents allocated in the nodes and other elements of the network are autonomous and communicate themselves to adapt the traffic of the network. These systems have also Connection Admission Control (CAC) allowing the statistic multiplexation of multiple calls according to a specified quality of service in each call. 5. Quality of service agent The intelligent agent uses a set of threshold levels activated according to the congestion level of the node. In each node is analized the ocupation of the buffers. In the case that the number of cells overflows the defined levels, the node eliminates the cells with low priority (CLP = 1). In front to the congestion, these cells (e.g. high quality video cells) are eliminated in order to protect the performance of the traffic with highest priority. In case of the traffic at the entrance of the access node consist of bursts that carry out the saturation of the buffers, this mechanism is not enough to avoid the congestion and losses of prioritary cells. Then a feedback to the information sources is applied and by means of the activation of the bit EFCI / EBCI in the data cells, the source node changes the priority level (CLP = 1) to the information cells sent to the network. If the congestion is mantained, the access node eliminates these cells and the quality of the service of the subscriber is preserved. The speed in the recuperation of the network after a congestion situation depends on the feedback. The feedback depends on the processing power of the nodes (delay algorithm), transmission speed in the links, measurements and propagation delay. A high speed in the feedback allows a reduction in the number of lossed cells. In addition, the quality of service is improved. 6. Conclusions In the article is presented a switching architecture that allows to provide quality of service management specified for a determined type of traffic (e.g. ). This quality management can aply also to other types of traffic. The studied mechanism that prevents the packet losses is based on an agents structure that act providing a selective discard of packets taking into account the packets flows sent to the network. In other side, all the management system is based on an intelligent multi-agent structure supported by appropiate switches. The use of intelligent agents optimize the behaviour of the system doing the parameters reconfigurables in real time. The structure of the agents can be done compatible with the defined standards IETF, ATM Forum and OMG. References [1] Recommendations about Traffic Management. v. 4.0, ATM Forum. [2] Ahmed Mehaoua, Yousses Iraqi, Adel Ghlamallah. An Intelligent Multi-Agent Architecture for Dynamic regulation of ATM Congestion Control Parameters. pp Management of Multimedia Networks and Services. IFIP. Chapman Hall [3] Jörg P. Müller. The Design of Intelligent Agents. A Layered Approach. Springer Verlag [4] Stuart Russell, Peter Norvig. Inteligencia Artificial. Un enfoque moderno. Ed. Prentice Hall [5] David L. Tennenhouse, A Survey of Active Network Research. IEEE Communications Magazine. January [6] M. Veeraraghavan, T. F. La Porta and R. Ramjee. A Distributed Control Strategy for Wireless ATM networks. Wireless Networks. p J. C. Baltzer AG, Science Publishers. [7] A. Barba. Gestión de red ATM en una arquitectura multi-agente a partir de parámetros de control de congestión. 1º congreso catalán de inteligencia artificial. p Octubre Tarragona.

9