Neural Networks In ATM Traffic Control
“ Infostream „05”
Introduction Asynchronous Transfer Mode (ATM) based networks are designed to be scalable, high-bandwidth, manageable, and have the flexibility of supporting various classes of multimedia traffic with varying bit rates and Quality of Service (QoS) requirements. Thus, they have the potential to create a unified communications infrastructure that can transport services with widely different demands on the network. An important difficulty of exploiting the potential of ATM optimally is the management and control complexity of the scheme itself. Since ATM simultaneously attempts to support voice, data and video applications which all have differing performance and QoS requirements, optimal utilization of the network resources requires complex, nonlinear, distributed control structures. In order to achieve its potential, ATM networks will need to accommodate several interacting control mechanisms, such as call admission control, flow and congestion control, input rate regulation, routing, bandwidth allocation, queue scheduling, and buffer management. Due to the complex nature of the above mentioned control issues, researchers are looking for solutions by application of Neural Networks and Fuzzy Logic to design intelligent control systems to various aspects of ATM network management, often supplementing the existing control techniques. Their motivation arises from the reported success of those techniques in various previously unsolvable or difficult control problems in many diverse fields. Elements of Traffic Control: Traffic Control is the key to high speed networking design. It is the act of managing network traffic, providing service guarantees to user connections and ensuring optimal utilization of network resources . to achieve this several elements of traffic control are used. The important once are Traffic Contracting/Connection Management . Traffic Shaping Traffic policing Priority Control Flow Control Congestion Control
ATM Connection Parameter:
Controlling traffic in ATM networks means monitoring and managing
traffic and service parameters
defined by the ATM forum and International Telecommunication Union-Telecommunication standardization (ITUT}. They are ATM QoS Parameters or Service Descriptors Traffic Parameters
ATM Service Categories: Providing desired QoS for different applications is very complex for example voices delay-sensitive but not loss sensitive, data is loss sensitive but not delay sensitive, while some other applications may be both delay sensitive and loss sensitive. ATM networks provide multiple classes of service to support the QoS requirements of device application as shown in table below. Service class CBR Characteristics This class is used for emulating circuit switching. The cell rate is constant with time. CBR applications are quite sensitive to cell delay variation. Ex: Telephone traffic, video conferencing and Television. nrt-VBR This class allows users to send traffic at rate that varies with time depending on the availability of user information. Statistical multiplexing is provided to make optimum use of network resources. Ex: Multimedia e-mail rt-VBR This class is similar to nrt-VBR but it is designed for applications that are sensitive to cell delay variation. Ex: Voice with Speech Activity Detection (SAD),Interactive Compressed Video ABR This class of service provides rate-based flow control and is aimed at data traffic such as file transfer and e-mail. Depending upon the states of congestion in the network, the source is required to control its rate. The users are declared a minimum cell rate, which is guaranteed to the connection by the network. UBR This service category is intended for applications having bursty traffic, but which are not sensitive to cell delay, cell delay variation and cell loss. If the network is unable to carry the offered load, the load belonging to this category is dropped. Ex: simple file transfer and e-mail. CBR=>Constant Bit Rate nrt-VBR=> non-real-time Variable Bit Rate rt-VBR=> real time Variable Bit Rate ABR=>Available Bit Rate UBR=>Unspecified Bit Rate QoS Requirements for various applications: Application Reliability Delay Jitter Bandwidth
E-mail File transfer Web access Remote Login Audio on demand Video on demand Telephony Videoconferencing QoS Guarantees:
High High High High Low Low Low Low
Low Low Medium Medium Low Low High High
Low Low Low Medium High High High High
Low Medium Medium Low Medium High Low High
There are essentially two ways to provide QoS guarantees. To provide lots of resources, enough to meet the expected peak demand with a substantial safety margin. This is nice and simple, but expensive in practice, and can't cope if the peak demand increases faster than predicted: deploying the extra resources takes time. To require people to make reservations, and only accept the reservations if the routers are able to serve them reliably. Functions And Procedure For Traffic Control In ATM Networks: ATM promises to support all these different requirements with a common network. To provide specified and guaranteed levels of Quality of Service(QoS) To use available network resources efficiently.
ATM Traffic Control Mechanisms: ATM Traffic Contracting/ATM Connection Management: A traffic contract is defined by two sets of parameter: Traffic Parameters: Peak Cell Rate (PCR) Sustained Cell Rate (SCR) Burst Tolerance (BT), etc
QoS Parameters: Delay Cell loss
ATM Traffic Shaping: ATM Traffic Shaping is preventive control method. The purpose of traffic shaping is two fold: To shape traffic entering an ATM network to make it conform to the traffic contract.
To reshape traffic within an ATM network to ensure that it destination without being discarded.
Private ATM Network
Public ATM network
LEAKY BUCKET ALGORITHM: Shaping Mechanism   Traffic shaping is typically implemented using the leaky bucket algorithm. There are two control parameters in the algorithm: the service rate of the pseudo-server and the pseudo-buffer size. The algorithm can receive a bursty traffic and control the output rate. If the excess traffic makes the pseudo buffer overflow, the algorithm chooses discarding the cells or tagging them with CLP=1 and transmitting them. Bursty data generated by a source is stored in buffer and sent out at a lower rate, there by transforming bursty traffic with high PCR to more stable traffic with lower PCR.The mechanism is referred to as Peak cell Rate Reduction. Various multiple connection shapers have been proposed. One method calculates the inter-departure time, other calculates cell departure time. A new shaper has developed that can follow multiple generic cell rate algorithm rules and considers all traffic parameters.
ATM Traffic policing: Traffic policing is the mechanism used at the entry point to an ATM network to ensure that an incoming cell stream conforms to the traffic contract.fig:3
Application (here violates contract)
CLP=1 ATM Network N2
GCRA and UPC/NPC-Mechanisms: The most commonly used algorithm is the Generic Cell Rate Algorithm (GCRA).GCRA is used to define the conformance of the cells with respect to the negotiated traffic contract. The algorithm checks every cell for its conformance. The GCRA is used to define conformance with respect to PCR and CDVT,as well as SCR and BT. GCRA is defined in terms of two parameters: Increment (I) and Limit (L) and is denoted as GCRA (I, L).GCRA checks the conformance to the traffic contract, using the twin concepts of excepted and actual cell arrival times. For an incoming cell, the expected arrival time is referred to as the Theoretical Arrival Time (TAT).The conformance of the cell is decided is based on its TAT.A cell is said to be conforming if it arrives at time or after. All cells arriving before time are marked as non-conforming. Fig:4
ATM priority Control: In ATM Network, the end-system may generate traffic flows of different priority. Mechanisms-Selective Cell Discard Selective Cell Discard is the only priority control functions specified in ATM, which is based on the CLP bits of the ATM header and stated The cells with CLP bit set to „0‟ are considered as high priority cells . The cells with CLP bit set to „1‟are considered as low priority cells.
If the network is forced to drop some cells, the network selectively drops the cells with lower priority. The cells can be either marked by the user or by the Usage Parameter Control (UPC) function at the network ingress point .
Benefits of Traffic Shaping When much traffic flows past a packet bottleneck (logical or physical) the benefits of traffic shaping are: Less jitter. Less dropped packets. Smaller Lag
In telecommunication, jitter is an abrupt and unwanted variation of one or more signal characteristics, such as the interval between successive pulses, the amplitude of successive cycles, or the frequency or phase of successive cycles. Traffic shaping is thus about regulating the average rate. This is dependent upon Spacing cells in time Queuing service schemes
Neural network controller and its structure  The three-layer neural network of the controller has 6 input neurons, 6 hidden neurons and 2 output neurons (denoted by 6-6-2), as shown in Fig.5,
where q(k), u(k), n(k) represent the queue length (or cell number) of the multiplexer buffer, the source coding rate and the corresponding user percentage at sample time k, respectively. Training for neural networks: In order to determinate the weight between the neurons, the neural network is trained so as to optimize the performance index function and get the accurate nonlinear mapping between real-value inputs and outputs. If the off-line trained neural network holds throughout the control process, it will not accurately describe the dynamic input-output modeling due to the traffic sources‟ time-varying, uncertainties and burst, and the control system performance will degrade greatly. If online training algorithm is used for neural networks, the control system will not satisfy the real-time requirement and high burst of the traffic sources will result in updating the weight sensitively, which leads to loss of the robustness of the control system. Therefore, the neural network training method based on moving-window is used, i.e., at every L sample times, the neural network is re-trained on the basis of the previous L sample data. The length L of the moving window reflects the sensitivity of the controller to the dynamics of the system since the weights of neural networks are updated only at the end of each L measure periods. This is a tradeoff in the selection of this parameter. On one hand, a small L implies frequent updates of the weights of neural networks and better performance results at the expense of the possible instabilities. On other hand, relatively longer value for L assures stable control actions at the expense of long training time and the danger of inaccurate modeling.
NEURAL NETWORK TECHINQUES: Neural Networks can be used to solve many of the serious problems encountered in the development of a coherent traffic control strategy in ATM networks. The main philosophy that favors networks over conventional programming approaches is their learning and adaptive capabilities, which can bee utilized to construct adaptive algorithms for allocation resources, thus providing highly effective tools for congestion control.
Most traditional implementation methods of the traffic control functions are based upon the results obtained from analytical performance evolution such as queuing models. Neural networks are applied to the call level control functions such as traffic measurements, traffic enforcement, and rate-based feedback congestion control. Moreover neural networks can be applied to the networks control functions such as optimal link capacity assignment and dynamic routing.
Neural Networks For Traffic Measurements: Neural networks can characterize and predict the bit rate variations of a complex stochastic process using a simple parameter such as the number of cells arriving within a certain time period. Neural network can predict the arrival process over a fixed measurements period of time based upon a sampled values taken from the previous measurement period.
Neural Networks For Connection Admission Control : Various methods have been developed for neural network connection admission control function, some of them are: Learning control method, hybrid Admission method and moving windows training approach. In the learning control method for admission, neural network creates a decision function by learning the behaviour of the operating multiplexer. The controller uses a three layer fully connected neural network with back propagation. The hybrid control method defines a subroutine of CAC called Link Admission Control (LAC), which is performed at each node along the network path connecting the source and destination node. The schema comprises a supervised multi-layer perception (MLP) combined with a tractable analytical approximation. A moving Window training approach estimates the traffic‟s entire congestive behaviour from its impact on the output queue via measurements of quantities such as mean delay loss and jitter.
Neural Network For Traffic Policing: A policing mechanism using neural network, called Neural Networks traffic Enforcement Mechanism (NNTEM), was introduced. It is based upon an accurate estimation of the probability density function (PDF) of the multimedia traffic .The architecture is composed of two inter-connected neural networks. The first
neural network is trained to learn the PDF of an ideal non-violating traffic, whereas the second neural network is trained to capture the actual characteristics of the actual offered traffic during the progress of the call. Suggestions: Traffic Shaping: We suggest from the earlier proposed controller, that a fuzzy controller be incorporated in a similar neural network (at N1 in fig0). As a fuzzy system (motivated by proposal in ) unlike other logic can yield intermediate output (not only true and false). This feature can be used to enable dynamic rate determination for received packets (increasing no. of output nodes unlike fig.2). The fuzzy rules can be determined based on the type of data (delay-sensitive, loss-sensitive etc.) received keeping in mind the QoS specifications.
Traffic Policing: A Neural Network may be designed (at N2fig. 3) with required no. of hidden layers and two output units-one representing lower priority and another representing higher priority. This can be used to set the CLP(Cell Loss Priority) bit. Later using Selective Cell Discard, these packets may be dropped if necessary. Learning methods taking into account the negotiated traffic contracts (for ex. Like the GCRA, the expected arrival time and the actual arrival time may be used to decide the functions) may be used.
Discussion and Concluding Remarks Research on applications of Neural networks and Fuzzy systems in ATM networks is being pursued by an active research community, and methods are being developed simultaneously. However, unlike consumer applications, there are no commercially deployed applications as yet. The reasons could be the lack of comprehensive performance comparisons between the best traditional techniques and the ones involving Neural networks and Fuzzy systems. This issue is closely related to The lack of comprehensive performance studies. The comparisons performed in the research studies usually have been undertaken in simplified networking scenarios, and testing on real hardware has not been undertaken yet except for some partial implementations. As a final note, by appropriate design of each individual strategy in a new multilevel fuzzy logic structure, and/or the integration of existing, or separately designed strategies, with their integration achieved via a fuzzy logic based supervisor, taking care of the overall “goodness” of the network and handling any interactions among the control functions, at the same or different levels.