A novel neural network traffic enforcement mechanism for ATM networks

Share Embed


Descrição do Produto

I08X

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 12, NO. 6. AUGUST 1994

A Novel Neural Network Traffic Enforcement Mechanism for ATM Networks Ahmed A . Tarraf, Student Member, IEEE, Ibrahim W. Habib, Member, IEEE, and Tarek N. Saadawi, Member, IEEE

Abstract-The Asynchronous Transfer Mode (ATM) principle has been recommended by the CCITT as the transport vehicle for the future Broadband ISDN networks. In ATM-based networks, a set of user declared parameters that describes the traffic characteristics, is required for the connection acceptance control (CAC) and traffic enforcement (policing) mechanisms. At the call set-up phase, the CAC algorithm uses those parameters to make a call acceptance decision. During the call progress, the policing mechanism uses the same parameters to control the user’s traffic within its declared values in order to protect the network’s resources and avoid possible congestion problems. In this paper, a novel policing mechanism using neural networks (NN’s) is presented. The mechanism is based upon an accurate estimation of the probability density function (pdf) of the traffic via its count process and implemented using NN’s. The pdf-based policing is made possible only by NN’s. This is due to the fact that pdf policing requires complex calculations, in real-time, at very high speeds which is not feasible via conventional mathematical approaches. The architecture of the policing mechanism is composed of two inter-connected NN’s. The first one is trained to learn the pdf of an “ideal nonviolating” traffic, whereas the second is trained to capture the “actual” characteristics of the “actual” offered traffic during the progress of the call. The output of both NN’s (which is an accurate estimate of the traffic bit-rate fluctuations in the next measurement period) is compared. Consequently, an error signal is generated whenever the pdf of the offered traffic violates its “ideal” one. The error signal is then used to shape the traffic back to its original values. The reported results, prove that our policing mechanism is very effective in detecting (and policing) all possible kinds of traffic violations (e.g., peak and mean violations).

T

I. INTRODUCTION HE HOPES OF CARRYING OUT a high speed net-

work in the near future has been pinned to the rising star of Asynchronous Transfer Mode (ATM). ATM is the technology recommended by CCITT [ 13 to provide voice, data, image and video services using a single integrated broadband network. The ATM solution is flexible enough to support a diverse mixture of multimedia traffic with different correlations and burstiness properties. Not only does the different types of multimedia traffic differ in their statistical characteristics but also in their service requirements specified by the quality of service (QOS). The QOS defines a set of guaranteed performance measurement parameters (e.g., cell loss rate, delay, delay variability) that each type of the multimedia traffic requires from the netThe authors are with the Electrical Engineering Department. The City University of New York, New York; NY, 10031. IEEE Log Number 9401462.

work. Due to the absence of a channel structure in ATMbased networks, the user’s traffic can easily overwhelm the network resources leading to serious congestion problems. Traffic enforcement (policing) is required to control the user’s traffic to an agreed “contract” determined at the call set-up phase. The policing mechanism controls the traffic via a set of user declared descriptors or parameters such as the peak bit-rate, mean bit-rate and the burst duration. These parameters are also used by the connection acceptance control (CAC) algorithm that decides to accept a new connection if the required QOS is guaranteed for the new connection while maintaining the QOS of the existing ones. Hence, at the call set-up phase, a traffic contract is “agreed-upon’’ between the network and the user. According to that contract, the CAC algorithm maintains a specific QOS for the user’s connection as long as the user does not violate his contractual parameters that characterize the traffic. If a violation of this contract (either caused by a malfunctioning terminal or due to a deliberate attempt by the user) is detected, the policing mechanism will enforce the original contractual parameters by an appropriate action. This action can be either cell-dropping [2] or throttling the source bit-rate [3]. The policing mechanism must fulfill the following requirements : 1. It must be capable of detecting and taking a swift control-action, whenever a violation occurs, on realtime basis. 2. It must be simple and cost effective to implement in hard ware. 3. It must be transparent for connections that respect the traffic contract and take no policing actions on their cells. On the other hand, it should act on all cells violating the contract in order to confine the traffic to its contractual agreement. 4. It must have a short dynamic reaction time (which is measured by the delay from the instance a violation occurs till it is detected) in order to avoid the flooding of the relatively small buffers in the network by violating cells. To meet these somewhat conflicting requirements, several policing mechanisms have been proposed, such as the leaky bucket [4]-[6], source rate throttling [7], window mechanisms [8], [9], and several others (see [lo], Ell] and the many references therein).

0733-8716/94$04.00 0 1994 IEEE

TARRAF et. al.: NEURAL NETWORK TRAFFIC ENFORCEMENT MECHANISM FOR ATM NETWORKS

In summary, any policing mechanism can be classified into a measuring algorithm, that monitors the traffic parameters to be policed and an action to be taken when these parameters are violated. Several options exist for the choice of the policed parameters. For example, one might consider a set of parameters such as peak and mean bit-rates. One might, also, choose a set such as the maximum number of cells (X) that can arrive in certain time interval T. There are two conflicting factors that influence the choice of these parameters: (1) The set must be simple enough, such that it can be easily measured and enforced. (2) The set must fully characterize the traffic bit-rate timevariations in order to enhance the statistical multiplexing gain required for bandwidth allocation. Most of the existing policing mechanisms attempt to police the peak and the mean bit-rates of the traffic. However the peak and mean policing functions check only one parameter of the probability density function (pdf) of the bit-rate of the source. The fundamental policing problem, with respect to mean cell-rate policing, is that the characteristics of the source must be estimated in a relatively short time period (sampling period), thus leading to an inaccurate estimation of the mean cell-rate and hence incorrect policing decision. An extension of the sampling period, on the other hand, also increases the reaction time of the mechanism, hence the policing action will not be in time to correct the violations. Another problem is caused by the fact that a simple set of traffic descriptors will not accurately characterize very rapid changes in the bit-rate time-variations of the traffic over short intervals, thus leading to an increase in the margin between the policed bit-rate and the actual one. Hence, reducing the effectiveness of the policing algorithm as well as that of the CAC algorithm. A promising policing mechanism is the one that attempts to police the pdf [ 121, [ 131. Unfortunately in most cases, this pdf can not be described by a well known mathematical model. Moreover it involves complex calculations of higher order moments that can be not achieved in real-time. For example, the Gabarit policing mechanism [ 121 approximates the pdf by a mathematically well defined distribution (e.g., a Gaussian distribution). However, due to the complexity of its implementation, it is unfeasible to police the Gaussian envelope continuously over all points. Hence, a stair-case shape function is used to approximate the Gaussian envelope leading to errors in the estimation of the policing parameters. In this paper, we adopt a novel approach to this problem using NN’s, which we call Neural Network Traffic Enforcement Mechanism (NNTEM). In the NNTEM, a traffic source is characterized via the pdf of its count process. The NNTEM method does not explicitly calculate any higher order moments in order to police the pdf of the traffic. It does that, however, via the NN transfer function which implicitly learns the pdf of the traffic count process through several millions of learning trials. Hence, the NNTEM does not rely upon the policing of simple parameters such as mean bit-rate, peak bit-rate, or burst dura-

1089

tion, but rather an elaborate and very accurate function (pdf) which includes all the statistical properties of the traffic. Moreover, the NNTEM actually predicts (through the NN leaning algorithm) any violations in the count process bit-rate variations. If the source traffic violates its “agreed-upon” characteristics declared via its pdf, a violating signal (error signal) is generated. This signal is used to regulate the traffic by dropping the violated cells. The advantage of the NN traffic enforcement approach over other mechanisms is the simplicity of implementing a policing function that can characterize (and predict) the bit-rate variations of the traffic (and hence any violations) and police it. The motivation behind our selection of NN’s is that they are very effective in learning and hence predicting, non-linear complex functions, thus making it an ideal tool to employ in ATM networks. Section I1 describes the proposed NN traffic enforcement mechanism, Section I11 reports simulation results, Section IV contains humerical results whereas the conclusions are given in Section V . 11. NEURALNETWORKTRAFFICENFORCEMENT MECHANISM (NNTEM) The most widely used NN learning algorithm is the back-propagation [ 141. Backpropagation is an example of a mapping network that learns an approximation to a function, y equal to f(x), from the sample x, y pairs. The fact that the function to be learned is non-linear presents no problem to backpropagation network (BPN). We used backpropagation successfully to characterize the multimedia traffic in ATM networks [15], [16]. Backpropagation NN’s have been used in a number of deterministic and stochastic problems [17], [18]. It has been found that these networks perform well in most cases with accurate results. Fig. 1 shows an example of a backpropagation NN with three layers: input layer, hidden layer and output layer. Each layer has a number of processing elements (PE’s) or neurons that are fully interconnected via adaptive weights (Wg)where the weight Wg connects the output of the ith neuron with the input to the j th neuron. The neurons in the input layer do not perform any processing on the input data, they simply store the input values. The hidden and output layers neurons carry out two calculations. First, , N and a bias (equal they multiply all inputs i = 1, 2, to a constant value of 1) by a weight and then they sum the result as: n

s. = ; c wjx; =I J

(1)

Second, the output of t h e j th neuron, Oj, is calculated as the sigmoid function (Activation function) of S j as:

oj = U (Sj)

(2)

where a(z) =

1

1

+ exp (-z)

(3)

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 12. NO. 6 , AUGUST 1994

Input Traffic

Predicted Traffic

1x1

IYI

hpltlqe

HddslW~

_ _ _ NN2 -~

A

* X i ]

-

wtmtme

Fig. 1 . BPN for traffic prediction

r-- Error Signal

~-

Fig 2 NNTEM structure

The backpropagation network learns by making changes in its weights in direction to minimize the sum of squared errors between its predictions and a training data set. The minimization uses the steepest descent algorithm [ 191. Assume that there are R input/output pairs x(~)Y(‘), available for training the NN. After presentation of pair r , the weights are changed as follows: where

In the Off-line training phase, NN 1 is trained to capture the actual pdf of the ideal non-violating traffic (a traffic that does not violate its pdf during the call progress). For example consider a multimedia source that generates three types of traffic (data, voice and video). NN1 is trained to learn the characteristics of each type and any possible combinations of those types (e.g., video and voice, data and voice, etc.). Hence, NN1 learns the characteristics of Wc! = Wp,- I ) A Wz! (4) several classes of traffic. In the call set-up phase, the user W!”- ’) is the value of a weight from neuron U in the hid- will declare the connection’s class of traffic and accorddenhnput layer to neuron U in the output/hidden layer at ingly, NN 1 will select the corresponding input-output :! is the value of the mapping function that contains specific information about step ( r - 1) (before adjustment). W weight at step ( r ) (after adjustment). When AW: is hid- the pdf of that class of traffic. den to output weight change, then: The NN learns the pdf of the traffic as it learns the relationship between current and past values of the number A W$’ 0 ( S k ) [ yp’ - 0 p’] 0; ( 5 ) of cell arrivals within a certain measurement period (count process). NN2 is trained to predict the future values of when A W g is input to hidden weight change, then: the actual offered traffic (which may exhibit violations in its pdf) during the call progress. In the on-line phase (the A W$) = c ~ ( S j ) C u ’ ( S ~ [) $) - Or’] W r -I ) X:” (6) call progress phase), the current and past values of the I Jk count process are fed to both NNl and “2. NN1 prewhere dicts the future values of the count process of the traffic @ ’ ( S k > = O(Sk)[l - ‘J(Sk)l (7) based upon its inputs and upon the actual pdf of the nonviolating traffic, whereas NN2 predicts the same for the After presentation of the first input/output pair one proactual offered traffic (i.e., it actually predicts the violaceeds with the second pair and so on, the weights are tions based upon past history). If the source traffic does changed with each presentation. not violate its ‘ ‘agreed-upon’’ characteristics, both NN 1 and NN2 produce almost the same outputs in the next time A . System Overview interval. Hence, the error signal between NN1 output and The architecture of NNTEM is shown in Fig. 2. In the NN2 output is almost zero and no policing action is taken. NNTEM, the traffic source is characterized via the pdf of While, if the NN’s detect any violations, an error signal its count process. The NNTEM method does not explic- is produced and a corresponding policing action will be itly calculate any higher order moments in order to police implemented. The error signal is function only of any vithe pdf of the traffic. It does that, however, via the NN olations that may occur in the shape of the traffic bit-rate transfer function which implicitly learns the pdf of the variations depicted by the count process. However, in ortraffic count process through many learning trials. Hence der to achieve an acceptable level of robustness in the the NNTEM, does not rely upon the policing of simple policing action, one might specify a certain threshold of parameters such as mean bit-rate, peak bit-rate, or burst the error signal. Any error signal that is less than the duration, but rather uses an elaborate and more accurate threshold value will be discard and no policing action will function which includes all the statistical properties of the be activated. traffic. Moreover, the NNTEM actually predicts (through 1. Trafic Source Model: In this paper, we used an ON/ the NN learning algorithm) any violations in the count OFF voice source model to demonstrate the efficaciousprocess bit-rate variations. If the source traffic violates its ness of the policing mechanism. Fig. 3 shows the char“agreed-upon’’ characteristics declared via its pdf, a vi- acteristics of the cell arrival process from a voice source. olating signal (error signal) is generated. This signal can This model has been used in earlier publications to model be used to police the traffic by different ways, the simplest packetized voice [20]-[22], still picture [8] and interacone is to use this signal to drop the violated cells. We tive data services. It allows the relevant parameters define two phases of operation: (1) Off-line training phase. namely peak bit-rate [ b ] ,mean burst (ON) duration [h], (2) On-line operation phase. mean OFF (silence) period [ k ] , mean bit-rate [ m ] ,to be

+

Lk:

1

TARRAF er. al.: NEURAL NETWORK TRAFFIC ENFORCEMENT MECHANISM FOR ATM NETWORKS

I

Talk-spurt ped&

I,- +--

Sllent perlod

I

I

Traffic Source

2

1091

Target network output

HF+m)

I

1

I

b

time

+ T +

I

1

Fig. 3. Two-phase ON/OFF process

varied independently of each other. The O N and OFF time periods are assumed to be exponentially distributed random variables [21]. During the O N period, a fixed number of cells is generated, each having a duration of T ms, whereas in the OFF period, no cells are generated. This model was used to generate the count process (X, T,), where ( X ) is the number of cells arriving in a time interval (T,). 2. NNTEM: The role of the NN1 and “2, in this application, is to capture the unknown complex relationship between the past and future values of the traffic [16]. In other word, the NN (“1 or ”2) is employed as an adaptive predictor that learns the stochastic properties of the traffic. Fig. 4 illustrates the basic idea in training a NN to act as a predictor. The cell arrivals process is represented by the data vector [ H ( i m)] which is the NN target output vector. The data vector [ H ( i m ) ] provides the NN with the count process information from which the predictions will be made. The NN predicts the count process variations by exploiting the inherent correlations that exist among the arrivals in the cell arrivals process. For training purposes, the input data vector [(H(i)] to the NN, is a delayed version of the data vector [ H ( i m)]. The NN, then tries to match the target output datalvector [ H ( i m)] with its predicted output data vector H [ ( i m)] by adjusting its weights. It, then, follows that when the input to the NN bypasses the delay unit, the output vector [A(i m)] is a prediction of the values the traffic will have in the future. The delay unit, shown in Fig. 4, delays its input [ H ( i m)] for m time steps. Assuming that the NN requires a negligible amount of time to compute its output from its input, the NN, a t e r training, provides estimates for the values of traffic [ H ( i m)] m steps in the future. This approach to adaptive prediction rests on the assumption of a parameterized class of models for functional relationship between the current and past values of the traffic and its later values, or equivalently between earlier values of the traffic and its current values. The NN model used is

+

+ output vector from the NN representing the expected traffic over the next measurement period (T,), see Fig. 5 . It then follows that,

T, = m

* T,

(9)

The NN input traffic pattern [H(i)] is expressed as

+

+

+

+

+

+

+ m)l

I

Fig. 4. Using neural network for adaptive predictions.

where h ( i ) is the value of the count process at the sampling instant i expressed in number of cells. The target traffic pattern for the next m measurement intervals [ H ( i m)] is expressed as

+

= NNf{[H(i)l, [WI)

(8)

NNf denotes neural network transfer function, [ W] represents the weight matrices of the hidden and output layers. The vector [H(i)] represents the count process over the past measurement period (T,) up to the present instant i. In this paper, the count process is used as a traffic descriptor in ATM networks due to its robustness against the delay variability of the cells’ interarrival times. Thus, the vector [H(i)] consists of m samples of the count process. These samples are obtained by samqling the arrival process at every sampling period (T,) [ H ( i m)] is the

+

[ ] + m) h(i + m - 1)

+

[fi(i

I

h(i

[ H ( i + m)]

=

h(i

(11)

+ 1)

As mentioned above, during the off-line training phase, NN1 learns the bit-rate variations of the traffic (in terms of the count process (X, T:)). It predicts future value(s) of (X, T,) (represented by [ H ( i m)]) based upon the current and past values of (X, T,) (represented by [ H ( i ) ] ) .By doing this prediction process, NN1 actually learns pdf of (X, T,). In other word, after completion of the training phase of NN1 and upon the introduction of the traffic pattern of (X, T,) to its inputs, it maps these inputs to its outputs via the already captured pdf of the non-violating traffic. Once NN1 learns the pdf of the ideal “non-violating” traffic, it does not learn any new traffic-patterns. During the on-line operation phase, it acts as reference bench-mark model for the ideal “non-violating’ ’ traffic. Whereas in the on-line operation phase (call progress) NN2 continues to learn new traffic-patterns which are violations of the original ideal traffic. It is clear from the previous discussion, that the traffic generated from a source can be categorized into one of two possible categories. Either a “non-violating” (ideal) traffic category, or a “violated” traffic category. If NN1 is exposed to a traffic that belongs to the ideal category, it produces outputs that are extremely close to the actual

+

1091

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 12, NO. 6, AUGUST 1994

No of cells

-* time

Fig. 5 . The sampled arrival process

future values. NN2 is trained “off-line” as well as “online”. The main purpose of NN2 is to predict the future value(s) of ( X , T,) for the two traffic categories mentioned above. During the off-line learning phase, NN2 is trained to learn both categories and learns how to predict the future values of both “violating” and “non-violating’’ traffic. It is important in the training phase, that NN2 encounters all possible violations that might occur (e.g., peak, mean, burst length violations. . . etc.). Although, the structure and the leaning rates of NN2 could be different from those of “1, both NN1 and NN2 still have the same size of input vectors and output vectors. In the on-line operation phase, both NN1 and NN2 are exposed to the same current and past values of (X, T,) through the input traffic vector [H(i)]. A violating signal (error signal) is generated by subtracting the output vector [&(i f m ) ] of NN1 from that of “2. This signal tells how far the current traffic, generated from the source, violates its “agreed-upon’’ contractual values. The violating signal will be zero if the traffic does not violate its “agreed-upon” pdf and will be non-zero if it does. This signal is then used to drop the violated cells, or more efficiently, it can be fed to another third NN that interprets this signal, together with some pre-defined optimization function. Based upon that function, the error signal could be ignored, amplified or multiplied by some factor. For example, one might wish to ignore the short term violations in the peak bit-rate as long as the mean bit-rate is not violated. Such added functionality can be simply incorporated by adjusting the optimization function such that less weight is given to the error signal generated by such types of violations. For example, consider two categories ofusers: namely category A and B. Category A users, are charged more per connection than category B users. Then, category A users could be allowed, for example, certain violations in their peak bit-rate based upon a pre-defined criterion, whereas category B users are not allowed any kind of violations. 111. SIMULATION RESULTS Extensive simulations were performed to obtain the NN’s data set, for both training and production phases and also to assess the performance of various NN’s architectures. The cell arrivals process, resulting from the

packetized voice source, was simulated on a Sun Sparc station 330 using the C language. Assuming a voice source peak bit-rate of 32 kbps and an ATM cell size of 53 bytes, we have a cell interarrival time T of 12 ms. The nominal values of the relevant parameters for that source are: bo = 32 kbps, ho = 350 ms, ko = 650 ms and mo = 11.2 kbps. A traffic that has these parameters and does not violate them during the call progress is an ideal “non-violated” traffic. NN1 is trained using the above parameters in the off-line phase of operation to learn the pdf of the count process. An important parameter which must be defined is the sampling period T,. The choice of this parameter is influenced by the type of the traffic and in this application has been found to be 50 ms (see [16] for more details). NN1 is trained to predict the count process for the next sampling period (T, = 50 ms) based upon 20 past values of the count process. Hence it contains 10 PE’s within the hidden layer and one PE in the output layer. The training data set of NNl(20, 10, 1) consisted of the past-history of the data sets of [ H ( i ) ]and [ H ( i + m)] vectors observed by running the simulation model for several hours. These past-history data are fed to NNl until the convergence of the mean squared error (MSE) is achieved. The rate of conversion of the network is illustrated in Fig. 6. It shows a plot of the MSE between the actual traffic and the predicted traffic as function in the number of times a data set has been presented to the network, it takes 3500 presentations until the conversion is achieved. Similarly, in this phase of operation, NN2 is trained to predict the count process for different types of traffic (violating and non-violating). The training data set of NN2 contains all possible traffic patterns that might appear due to any violations in its pdf. For simplicity, we considered the following cases of traffic violations: case( 1): The traffic violates both the peak bit-rate and the mean bit-rate. case(2): The traffic violates the peak bit-rate only. case(3): The traffic violates the mean bit-rate only. The “on-line” operation phase starts as soon as the call gets accepted by the CAC algorithm at the call set-up phase. During this phase, the learning rate parameter (a parameter which determines the speed of learning of the NN) of NN1 is set to zero, since it is not required from NN1 to learn any characteristics about the offered traffic. However, because the function of NN2 is to predict correctly the future values of the offered traffic, its learning rate is adjusted such that NN2 continues to learn new traffic-patterns. During this phase of operation, both NN 1 and NN2 are exposed to both the past and the current values of the traffic through the vector [ H ( i ) ] .An error signal is generated, by comparing the output of NN 1 and that of “2. This error signal is used to control the status of the cell dropping switch, see figure (1). If the source traffic is a non-violating one, both NN1 and NN2 yield almost the same value for the count process ( X , T,) over the next sampling period. In this case the error signal is zero and no cells will be dropped. While if the traffic violates its

TARRAF et. al.: NEURAL NETWORK TRAFFIC ENFORCEMENT MECHANISM FOR ATM NETWORKS

I093

4.00E-02 3.50E-02 3.00E-02 -2.50E-02 ~2.00E-02 -1.50E-02 -~ 1.00E-02 -5.00E-03 -O.OOE+OO

I

1

0

500

1000

1500

2000

2500

3500

3000

4000

No. of iterations

Fig. 6. Rate of convergence.

8001

contractual parameters, NN1 gives a prediction of the count process (X,T,) based upon the pdf of the ideal traffic stored in it, whereas NN2 predicts the violating traffic over the next sampling period. In this case, the error signal will not be zero, and the violating cells are dropped. In this paper, the NN's were simulated using HNC Inc. EXPLORENET 3001 package [23]. Several experiments were performed to tune up the parameters of NN1 and "2. The activation functions of these NN's were selected to be sigmoid where the steepness parameter (a factor that determines the non-linearity of the sigmoid function) was selected by trial and error method to achieve good results. IV. NUMERICAL RESULTS In this section, we demonstrate the effectiveness of the proposed NNTEM by demonstrating its capability to produce an error signal whenever a violation of the pdf occurs. Four experiments were performed to explore how the error signal is generated for the different violations cases mentioned in the previous section. Fig. 7 shows the pdf of the ideal traffic as well as for each of the different violations cases. Experiment I: In this experiment, the offered traffic during the call progress does not violate its contractual parameters. The pdf of this traffic is shown in Fig. 7(a). Fig. 8 shows the NN1 output compared with the actual one. It is clear that NN1 prediction is very close to the actual traffic values and NN1 capture the statistical properties of the ideal traffic as can be assessed by Fig. 9 which shows the autocorrelation function of the traffic produced by NNl compared with that of the actual traffic. As can be seen that NNl based on backpropagation algorithm does excellent job in characterization and learning the underlying the arrival process. Fig. 10 shows the error signal for that case, where the error signal is zero because the offered traffic to NN1 has the same pdf, as that captured by "1. No policing action is taken in this case. Experiment 2: In this experiment, the offered traffic violates its contractual parameters through an increase in both its peak and mean bit-rate. The resulting pdf of this traffic is shown in Fig. 7(b). In this case the traffic relevant parameters were b = 1.5bo, and m = 1.5mo. Fig. 11

400

z 200 0

2

4

L 6

"0

2

4

6

No. of cells

No. of cells

(d) (C) Fig. 7. Histogram of the count process. (a) No violations. (b) Peak and mean violations. (c) Peak violation. (d) Mean violation.

12

16

14

18

20

tirne(sec)

Fig. 8. Comparison of the output process from N N l with the simulation one.

shows the generated error signal versus time which indicates how far the offered traffic has deviated from its ideal case. For example, at time = 10 s the error signal e(t) is

-

I094

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 12. NO. 6 , AUGUST 1994 ~

'Ii

I\

o.8t , 1

sirn nu la ti on ....... NNI

0.3

),

I

8

",

0.25

0.6-

0.2 I Y (U

0.15

0.1

0.05

!

b!

i -0.2'

0

2

1

3

1 5

4

tirne(sec)

1

-0.4

I

~

~

-0.6-

-0.8-

I

0.35,

83.25

-5 0 2 0.15 01

0.05 n

"0

10

20

30 time (sec)

"0

10

20

40 tim

Fig 9 . Comparison of the autocorrelation function of the count process ("1, simulation).

-0.2

n

40

50

Fig. 1 1 . Error signal-peak and mean violations case

60

50

iec)

Fig. 12. Error signal-peak

violation case

about 0.27, hence about 27% of the offered cells at that time will be dropped. As can be seen from Fig. 11 that the error signal e(t) started with zero and increased rapidly in less than 0.5 s, hence the reaction time in this case is less than 0.5 s during which, about 42 cells were emitted. This reaction time is very small compared with the reaction time of window mechanisms [8], which is about 150 cells. Experiment 3: In this experiment, the traffic violates, only, the peak-bit rate such that b = 1.5bo. The pdf of this traffic is shown in Fig. 7(c). Fig. 12 shows the error signal for that case. From Figs. 11 and 12, we can see that the error signal generated from peak and mean violations and that generated from a peak violation only, have similar time varying characteristics and are very close in magnitude. This explains the reason that almost all existing policing mechanisms have tremendous difficulties enforcing both peak and mean bit-rates violations. Since it is difficult to distinguish traffic with peak bit-rate violations from one that incorporates both peak and mean bitrates violations. Experiment 4: In this experiment, the offered traffic violates, only, its mean bit-rate such that m = 1.5mo. The pdf of this traffic is shown in Fig. 7(d). Fig. 13 shows the error signal in that case. By reference to Figs. 11-13, it is clear that the error signal in that case is less in magnitude than those of other types of violations and has different time varying characteristics. This result confirms that peak and mean bit-rates violations are not detectable by the same measuring algorithm. Since a measuring algorithm that can detect the peak bit-rate violations must have very short reaction time, yet the same algorithm must have a longer reaction time to detect violations in the mean bit-rate. However, our algorithm can easily distinguish those two cases of violations as evident by comparing Figs. 12 and 13. It is clear from the above results that the NNTEM can easily detect both mean and the peak bit-rate violations

:

TARRAF et. a l . : NEURAL NETWORK TRAFFIC ENFORCEMENT MECHANISM FOR ATM NETWORKS

Although the training phase needs extensive number of trials, it does not affect the performance since it is done “off-line” and before the actual “on-line” or production phase of the NN.

0.1 -

0.08-

g0.06-

0.04-

0.02 -

50

L

O0-

1095

10

20

30 time (sec)

40

Fig. 13. Error signal-mean violation case

60

with a single NN architecture. The problem of policing both the mean and the peak bit-rates using a single set of parameters has been a challenging problem for all existing flow enforcement mechanisms. Our proposed NNTEM solves this problem via a pdf-based algorithm utilizing the learning capabilities of NN’s. For example, the Leaky Bucket mechanism [4],needs two different set of parameters in order to police the mean and the peak bit-rates. While the same NNTEM structure is capable of policing both the mean and the peak bit-rates. Another advantages of the NNTEM is it has very short reaction time compared with the jumping window mechanism, triggered jumping window mechanism, and others mentioned in [8].

V. CONCLUSION In this paper, a pdf-based neural network flow enforcement mechanism (NNTEM) is presented. In this mechanism, a set of two inter-connected backpropagation NN’s are used to characterize and predict any violations in the pdf of the traffic. The proposed NNTEM is suitable for the variable bit-rate multimedia traffic enforcement in ATM-based networks. This is achieved by utilizing one NN to capture the actual pdf of the ideal “non-violating’’ traffic, whereas the other NN is trained to adaptively characterize and predict any type of traffic violations by learning the relationship between the past and future traffic variations. The NNTEM polices the actual pdf of the traffic which includes all the statistical properties of the traffic (instead of simple parameters such as the peak and mean bit-rates). The error signal produced by the NNTEM can detect the individual contractual parameter violations as well as any combinations of the contractual parameter violations. The proposed system solves the fundamental problem challenging most of the existing flow enforcement mechanics which is the difficulty in policing both the mean and the peak bit-rates. The system can be easily configured and trained to incorporate any types of multimedia traffic by a simple modification of its input vector to include the learning of the pdf of the new traffic types.

REFERENCES

[ I ] CCITT Recommendations, I series (B-ISDN), July 1992. [2] L. Dittmann, S . Jacohsen, and K. Moth, “Flow enforcement algorithms for ATM networks,” IEEE J . Select. Areas Commun., vol. 9 , Apr. 1991. [3] I. Habib and T. Saadawi, “Access flow control algorithm for voicemultiplexers in ATM networks,” in Proc. IEEE MlLCOM’93. [4] M. Butto, E. Cavallero, and A. Tonietti, “Effectiveness of the leaky bucket policing mechanism in ATM networks,” IEEE J. Select. Areas Commun., vol. 9, Apr. 1991. [5] H. Chao, “Design of leaky bucket access control schemes in ATM networks,” in Proc. ICC 1991. [6] H. Chao, “Architecture design for regulating and scheduling user’s traffic in ATM networks,” in Proc. ICC 1991. [7] I. Habib and T. Saadawi, “Congestion control in video multiplexers in broadband networks,” J. Comp. Commun., June 1992. [8] E. Rathgeb, “Modeling and performance comparison of policing mechanisms for ATM networks,” IEEE J. Select. Areas Commun., vol. 9, Apr. 1991. [9] T . Aoyama, I. Tokizawa, and K. Sato, “ATM VP-based broadband network for multimedia services,” IEEE Commun. Mag., Apr. 1993. [IO] T. Saadawi, M. Ammar, and A. Elhakeem, “Fundamentals of Telecommunication Networks,” John Wiley & Sons, 1994. [ 111 I. Habib and T. Saadawi, “Controlling flow and avoidance congestion in broadband networks,” IEEE Commun. Mag., Oct. 1991. [I21 F. Denissen, E. Desmet, and G. H . Petit, “The policing function in an ATM network,” in Proc. 1990 Int. Zurich Sem. Digital Communications, Zurich, Switzerland, Mar. 1990, pp. 131-144. 1131 M. De Prycker, Asynchronous Transfer Mode. Ellis Honvood Ltd., 1991. [14] P. G. J . Lisboa, Neural Networks Current Applications. New York: Chapman & Hall, 1992. [I51 A. Tarraf, I. Habib, and T. Saadawi, “Neural networks for ATM multimedia traffic prediction,” in Proc. Int. Workshop Appl. of Neural Networks to Telecornmun., IWANNTP3 Conf., Princeton, NJ. [16] A. Tarraf, I. Habib, and T . Saadawi, “Characterization of packetized voice traffic in ATM using neural networks,” in Proc. of IEEE Globcom’93 Conf., Houston, TX, Nov. 1993. 1171 J . M. Zurada, Introduction to ArtiJcial Neural Systems. New York: West Publishing, 1992. 1181 R. Nielsen, Neurocomputing. Reading, MA: Addison-Wesley. 1989. [19] D. E. Rumelhart, J . L. McClelland, and PDP Research Group, Parallel Distributed Processing, vol. I. Cambridge, MA: MIT Press, 1986. (201 H. Heffes and D. Lucantoni, “A Markov modulated characterization packetized voice and data traffic and related statistical multiplexer performance,” IEEE J. Select. Areas Commun., vol. SAC-4, Sept. 1986. [21] K. Siriram and W . Whitt, “Characterizing superposition amval processes in packet multiplexers for voice and data,” IEEE J . Select. Areas Commun., vol. SAC-4, Sept. 1986. [22] J . Diagle and J . Langford, “Models for analysis of packet voice communications systems,” IEEE J . Select. Areas Commun., vol. SAC4, Sept. 1986. [23] HNC Inc., “HNC ExploreNET release 2.0 Operating Manual,” 1991.

Ahmed A. Tarraf (S’92) received the B.Sc. and M.Sc. degrees in electrical engineering from Ain Shams University, Cairo, Egypt, in 1986 and 1990, respectively. He is a Ph.D. student in the electrical engineering department of the City University of New York. From 1986 to 1991, he was a Research/Teaching Assistant in the Department of Computer and Electronics Engineering at the same university. During the same period, he was a Senior Design Engineer working in the area of data acquisition

109h

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 12, NO. 6, AUGUST 1994

and control systems. Currently, he is a Research Assistant at the City Univ e r d y of New York. His research interests are in the applications of artificidl neural networks in the area of access flow control and congestion avoidance in broadband networks. He has spent the summer of 1993 working with AT&T Bell Labs, NJ, where he is currently a part-time MTS.

ATM networks and the applications of artificial intelligence in high speed networks. Dr. Habib is a member of the IEEE Technical Committee on Computer Communications and is an Associate Technical Editor of the IEEE COMMUNICATIONS MAGAZINE.

Ibrahim W. Habib received the Ph.D. degree . 1991 from the City University of New York and the M.Sc. degree from Polytechnic University of New York in 1984, and the B.Sc. degree from Ain Shams University, Cairo, Egypt, in 1981, all in electrical engineering. From 1984 to 1988, he was a Senior Computer Communications Engineer in the Middle East, working on many computer networking projects such as IBM SNA. From 1988 to 1991, he was a Research/Teaching Assistant in the Department of Electrical Engineering of the City University of New York. He is currently an 4ssistant Professor of Electrical Engineering at the City College of New York, where he is con ducting research in the area of congestion control in

Tarek N. Saadawi received the B.Sc. and M.Sc. degrees from Cairo University, Cairo, Egypt, and the Ph.D. degree in electrical engineering from the University of Maryland. From 1977 to 1980, he was a Research Assistant in the Department of Electric Engineering at the University of Maryland, College Park. His research interests are in high-speed and wide-area networks. He has published exclusively in the area of telecommunications networks. He consults with industry and the U.S. government. MAGDr. Saadawi is a Technical Editor of the IEEE COMMUNICATIONS A Z I N E , former Chairman of the IEEE Computer Society of New York City, and is listed in Who’s Who in the East. He received the IEEE Region I Award in 1987.

Lihat lebih banyak...

Comentários

Copyright © 2017 DADOSPDF Inc.