Admission Control (CAC) of ATM networks. The performance of MRAN CAC is
compared with that of the conventional CAC controller. OPNET was used as.
NEURAL CALL ADMISSION CONTROL OF ATM NETWORKS USING OPNET Mohit Aiyar, Shefali Nagpal, N. Sundararajan and P. Saratchandran S1-B4b-06, School of Electrical and Electronic Engineering Nanyang Technological University, Singapore, 639798. Email:
[email protected];
[email protected]
ABSTRACT
its traffic parameters provided during connection setup. However, this mode of operation possesses significant drawbacks.
This paper investigates the use of the newly developed minimal radial basis function neural network called, Minimal Resource Allocation Network (MRAN) for Call Admission Control (CAC) of ATM networks. The performance of MRAN CAC is compared with that of the conventional CAC controller. OPNET was used as the simulation tool to evaluate the performance of both the controllers. Under identical heavy traffic scenarios, the MRAN controller provided better results than the conventional controller. While the conventional controller arrived at a value of 0.9 for the Call Blocking Ratio, the MRAN based scheme showed a notable performance enhancement with a figure in the proximity of zero. Similarly, the Link Utilization was improved from about 0.12% for the conventional scheme to almost 65% for the MRAN based CAC.
In the practical design of CAC facilitating systems, there are a number of events to be considered such as deriving a proper model for the traffic source as described in [1]. It is difficult to obtain the complete statistics on input traffic to a network. As a result, it is not easy to accurately determine the equivalent capacity or effective thresholds for multimedia high-speed networks in various bursty traffic flow conditions. This is where conventional schemes perform poorly because they provide optimal solutions only under a steady state. Networks are forced to make decisions based on incomplete information so that the decision process is full of uncertainty. Because of the unpredictable statistical fluctuations of the system, these control schemes will always be subject to decision errors, which degrades performance.
1. INTRODUCTION
Because of their ability to efficiently adapt and predict, when incorporated in ATM networks, neural networks stand out as a very viable and suitable alternative. The neural network learns how to adapt to unexpected traffic as a non-linear function from noisy data. Some of the research work done in the application of neural networks for CAC has been described in [2] and [3]. Comparative studies done in [4]–[6] have shown that the proposed neural network approaches significantly improve system performance compared with conventional approaches.
Call Admission Control is the mechanism whereby the decision to permit a call to enter the network is made. This decision to admit calls requires information about the connection’s traffic characteristics and is based on various Quality of Service (QoS) parameters. The CAC along each node and along the end to end path of the connection will use these values to determine the traffic model and to decide whether to accept or reject the request for setup. If the CAC at every node along the path decides that there are enough resources to provide QoS guarantees, the request is accepted; otherwise it is rejected. The ATM node accepts the new call only if the QoS of the already connected calls is expected to satisfy their required values even after the new call is connected
In this paper, recently developed MRAN neural network methodology is used to enhance the performance of the conventional CAC controller. MRAN is a minimal Radial Basis Function (RBF) Neural Network with a sequential learning algorithm recently developed by Y.Lu et al [7], which combines the growth criterion of RAN with a pruning strategy to realize a minimum RAN.
In this paper, Call Admission Control (CAC) approaches which do not employ the use of the neural networks, are referred to as conventional schemes. In most cases, conventional CAC schemes perform a simple comparison of the requested bandwidth and the available bandwidth on the associated traffic links where the information on each connection’s traffic is contained in
The paper is organized as follows: Section 2 provides an overview of the conventional CAC Algorithm and also provides an insight into the location of the algorithm in the OPNET modeling environment. The MRAN
1
7simulation. Some of these values were modified later to simulate varying situations. However, some of these values remained constant to conform to the design specifications.
algorithm is briefly dealt with in Section 3 and the simulation results and the location for the incorporation of the algorithm in the MRAN based CAC scheme are presented in Section 4. Conclusions from the study are summarized in Section 5.
2. CONVENTIONAL CAC ALGORITHM This section highlights the results obtained on simulating the conventional CAC algorithm and proceeds to depict the location of the algorithm in the OPNET environment. Both, the conventional and the neural network based schemes, were simulated under identical heavy load highly-multiplexed traffic situations, as depicted in Table 1. Traffic nodes were placed on either side of an ATM backbone network. Traffic of various service categories were multiplexed on these sources in order to create a high traffic load scenario. Figure 1: Simulated Network Topology Figure 1 depicts the prototype network model simulated. The duration of the simulation varied from 10 to 45 minutes depending on the traffic scenario simulated.
Link Bandwidth 20Mbps
Traffic Source
Cbr_stn1
Traffic Contract / Mbps PCR 15.0
SCR 5.0
MCR 0.0
Ubr_stn1
20.0
10.0
0.0
Vbr_stn1
20.0
10.0
0.0
Application Configuration
OPNET follows a very structured hierarchical modeling methodology, and the location of the existing conventional CAC algorithm was traced through this layered structure, starting from the network layer to the node layer onwards to the process model and the finite state machines. Finally, the involvement of child processes and external files was investigated. The results of the investigation yielded that the conventional Call Admission algorithm is located in the ams_atm_support_v3.ex.c external file. The child processes, namely ams_atm_call_src_v3, ams_atm_call_net_v3, ams_atm_dr_v3 and ams_atm_call_dst_v3 make reference specifically to the ams_atm_vp_supports_traffic() function in this external file for the execution of the CAC algorithm. These child processes perform call handling tasks which specifically relate to the establishment and termination of virtual channel / virtual path assignments. The ams_atm_mgmt_v3 process spawns a child process, ams_atm_dr_v3. This child process performs all the necessary routing for calls within the ATM Layer. Two other child processes of ams_atm_mgmt_v3, namely ams_atm_call_src_v3 and ams_atm_call_net_v3, invoke ams_atm_dr_v3 to obtain routing information for the calls they are handling. ams_atm_mgmt_v3 itself invokes ams_atm_dr_v3 to create routing information t o the other nodes in the network. Other functions used for CAC are ams_atm_vp_vc_find(),
SAMPLE Custom Application: x 60 sessions/hour x 1 session = 5min x 60 requests/hour x Session Rate = Constant x Session Duration = Constant x Packet Size = 5000 Type: Database Appln, Email and FTP. Custom Application: x 1000 sessions/ hour x Same as above Custom Application: x 1000 sessions/hour x Same as above
Table 1: Input Traffic Characteristics The above table provides an example of the type of traffic that was multiplexed on the various traffic sources. These traffic parameters were assigned to each of the nodes/links before the commencement of the
2
ams_atm_vp_vc_alloc() and ams_atm_vp_vc_free(). Besides containing the relevant instructions for the execution of CAC, this external file also contains miscellaneous functions that are needed for call handling.
associated functions present in the ams_atm_support_v3.ex.c file such as ams_atm_vp_vc_find(), ams_atm_vp_vc_alloc() and ams_atm_vp_vc_free(). This information helps the algorithm to streamline its flow into various sections depending on the category of incoming traffic. For each traffic category, relevant comparisons are made to determine whether the call should be accepted or rejected. As stated earlier, this decision is largely dependent on the input values of PCR, SCR, MCR and MBS, which are by and large, constant.
In essence, each and every node at the network level implements the Call Admission Control algorithm via the ATM Management module at the node level. The process model associated with this node component dynamically generates the child processes listed above which in turn make reference to specific functions in the the ams_atm_support_v3.ex.c external file.
Under a densely-congested heavily-multiplexed traffic scenario, the simulation results obtained are displayed in Figure 2.
The CAC algorithm guarantees that the sum of the Peak Cell Rates is x Less than or equal to the maximum allocated bandwidth for the Constant Bit Rate category x The sum of Sustainable Cell Rates for the Variable Bit Rate category x The sum of Mean Cell Rates for the Available Bit Rate category
2a
Call Blocking Ratio
1 .2 1 0 .8 0 .6 0 .4 0 .2 1746
1638
1530
1314 13 86
1422
1206 12 60
990
1098
882
11 34
774
666
558
450
342
234
18
126
0
T i m e (s)
No call admission is applied to the Unspecified Bit Rate category. All calls belonging to the UBR category are admitted irrespective of their traffic contracts. Therefore, the algorithm essentially checks the service category that the incoming traffic belongs to and applies the CAC rule accordingly.
0 .0 0 1 4
2b
Link Utilization
0 .0 0 1 2 0 .0 0 1 0 .0 0 0 8 0 .0 0 0 6 0 .0 0 0 4 0 .0 0 0 2
17 64
16 38
15 12
10 08
88 2
75 6
63 0
50 4
37 8
25 2
0
12 6
0
1728
1620
1512
1404
1296
1188
1080
972
864
756
648
540
432
324
216
0
2c
0 .0 5 0 .0 4 5 0 .0 4 0 .0 3 5 0 .0 3 0 .0 2 5 0 .0 2 0 .0 1 5 0 .0 1 0 .0 0 5 0 108
Before the commencement of a simulation, the user specifies the input traffic parameters as part of the traffic contract. A QoS level is defined as a combination of the QoS category (CBR, RT-VBR, NRT-VBR, ABR, UBR), QoS parameters (pp-CDV, max-CTD, CLR) and traffic parameters (PCR, SCR, MCR, MBS). The ensuing conventional CAC algorithm then utilizes to a large extent the static traffic parameter values. During the occurrence of a call, the ATM Management module is invoked, which in turn makes a call to the relevant CAC associated functions in the ams_atm_support_v3.ex.c external file. The primary controlling CAC function in this file is the ams_atm_vp_supports_traffic() and at the outset it ascertains, whether the flow of traffic is in the forward or reverse direction and accordingly, formats the values of PCR, SCR, MCR and MBS. Once this is done, the algorithm determines the QoS class reflected in the incoming calls traffic contract. This calculation also uses the Maximum Allocated Bandwidth, Minimum Guaranteed Available Bandwidth and Maximum Available Bandwidth. These figures are calculated during the course of the simulation using CAC
Traffic Received (pkts / sec)
T i m e (s)
T i m e (s)
Figure 2: Conventional CAC results
Figure 2a shows a significantly high value of approximately 0.95 for the Call Blocking Ratio. This is very ineffective since almost 95 percent of the calls that are being presented to the network are being rejected. The reason for this behavior is that the QoS requested by the stations is higher than the supported network QoS. Similarly, Figure 2b depicts a very low value for Point to Point Link Utilization owing to heavy traffic and a more stringent CAC. Figure 2c presents the Custom Application Traffic Received at the network destination in packets per second. Due to heavy congestion in the network and an ineffective CAC scheme in operation, this value is very low.
3
efficient manner to realize a compact RBF network with better approximation accuracy. Furthermore, the proposed MRAN adaptive control scheme provides for adjustments of the centers, widths and also the weights which result in better approximation for the input–output nonlinear functions.
This approach has some inherent shortcomings, which are very crucial for modern day B-ISDN traffic. Essentially, traffic control methods such as those described above, require specific knowledge of the statistical behavior of the input traffic declared via its traffic descriptors. Traffic characterization using simple parameters, however, often ignores most of its important correlation and burstiness properties, and thus leads to incorrect parameters, which could seriously degrade network performance.
4. MRAN BASED CAC ALGORITHM The incorporation of MRAN is proposed to assist in the dynamic allocation of bandwidth in accordance to the dynamic nature of the varying input traffic.
After studying the severe congestion scenario in the depicted scenario and noting that it is quite hard to overcome the congestion for dynamically changing multiplexed traffic, the MRAN controller is integrated into the loop. The exact desired location for incorporating the MRAN algorithm was to be determined. The MRAN controller is allowed to operate and its efficiency in removing the congestion is assessed. Prior to that, a brief overview of the MRAN algorithm is provided below.
In practice, therefore, transmission bandwidth requirement should be assessed using online traffic measurement. If this can be done relatively accurately and efficiently, then dynamic bandwidth allocation can be done by adaptively changing the transmission bandwidth using an online measurement of traffic demand. This dynamic assignation of bandwidth can then be used to facilitate Call Admission Control. The key question, however, is how to effectively predict abrupt changes in the incoming traffic using the online traffic measurement. Traffic on ATM networks is highly bursty, non stationary and correlated. It’s inter-arrival time is complex and erratic. To effectively predict the traffic and dynamically facilitate the calculation of parameters to be used later for Call Admission Control, the MRAN neural network methodology has been adopted in this research initiative. The use of MRAN has been proposed to make the calculation of the available bandwidth and other figures of merit used in the CAC algorithm that are more dynamic and reflective of the actual status of traffic on the network.
3. MRAN ALGORITHM The MRAN is a recently developed minimal Radial Basis Function (RBF) Neural Network, which combines the growth criterion of RAN with a pruning strategy to realize a minimum RAN. In the classical approach to radial basis functions (RBF) network implementation, the basis functions are usually chosen as Gaussian and the number of hidden units (i.e., the centers and widths of the Gaussian functions) is fixed a priori based on the properties of the input data. The weights connecting the hidden and output units are estimated by a linear least squares method. The MRAN learning algorithm succeeds the Resource Allocation Network (RAN) and the Resource Allocation Network via Extended Kalman filter (RANEKF) research initiatives. The MRAN overcomes the problems stated above by incorporating the growth criteria of RAN networks and concurrently running a neuron pruning strategy based on the relative contribution of each hidden unit to the overall network output. The resulting topology leads to a minimal topology for RAN. The pruning strategy used prunes hidden nodes that do not contribute significantly to the output of the network, or that are too close to each other. This is done by observing the output of each of the hidden nodes for a period of time, and then removing the node that has not been contributing a significant output for that period.
The use of the MRAN algorithm and the exact location for incorporation was investigated and the outcome of this exercise is briefly outlined in the following section. Figure 3 depicts the use of MRAN in the dynamic calculation of allocated bandwidth. The modification of the process flow done in order to incorporate the MRAN neural network is in sync with the OPNET distributed process architecture. As highlighted, during the generation of a call, the cell rate and predicted cell rate are calculated and their values are transmitted over the process model – external file interface. These values are then accorded global access status in the external file so that they can be used as and when the CAC functions are executed. Simultaneously, the recalculation of available bandwidth is hard coded into the external file. Before devising this algorithm, it was determined that the incorporation of MRAN and the calculation of the
In the proposed scheme, the network builds up the hidden neurons from the input data. It does this in an
4
Within the OPNET simulation environment itself, for purposes of coherent comprehension of the simulation events, the authors defined statistics that were used to display the actual load versus the predicted load.
original and predicted cell rate must be done before the execution of the CAC functions. Therefore, the process model call generation state was thought to be the most suitable place for incorporation. Redirecting the flow of the process model for an interim period incorporated the neural network and the flow, as provided by the OPNET-provided conventional CAC algorithm, was rejoined once the neural network functionality has been completed.
Incorporation of MRAN took considerable time and effort. The factors that would determine the precise location of the MRAN neural network would be primarily the functionality of the neural network and the interface required for passing the required parameters to the external file. The fundamental functionality of the MRAN was determined to be the prediction of the arrival cell rate into the ATM network model. OPNET does not provide a custom made statistic for purposes of tracking cell rate. This has to be calculated by the developer. Therefore, the neural network was deemed to be located in the traffic generation unit of the traffic source. The redirection of flow within the external file is depicted in Figure 4.
Start
Call Generation
Calculation of Cell Rate
Time Series Generation for MRAN to
predict cell rate
Transfer Variable to external file
Execution of other call generation
At start of simulation, log OPNET sim time t1 At every arrival, log arrival time t2 Calculate interim time = t2 – t1 Log packet arrivals (P) in (t2-t1) s Cell Rate = P/(t2 – t1)
As can be seen, this approach of incorporating the MRAN algorithm was modular. Therefore, in a similar vein, it facilitates the incorporation of any other neural network algorithm for comparison with conventional CAC schemes. This can be the basis of future research efforts.
Set NN feedback MRAN_Y1 MRAN_Y1 = initial cell rate Function Call MRAN() Set MRAN_IP3 = MRAN_IP2 Set MRAN_IP2 = MRAN_IP1 Set MRAN_IP1 = MRAN_Y1 Log MRAN_OP
ams_atm_support_v3.ex.c global access status to transferred variable Recalculation of Available Bandwidth using predicted cell rate Use of recalculated bandwidth values to facilitate CAC
END
Figure 3: Incorporation of MRAN Using the neural network to dynamically alter the values of PCR, SCR and MCR in the external file, ams_atm_support_v3.ex.c did the recalculation of the available bandwidth.
5
Start Invoke Child processes
Call Blocking Ratio
1
5a
0.8 0.6 0.4 0.2 1784
1665
1546
1427
1309
1190
952
1071
833
715
596
477
358
239
1.8
121
0
Invoke CAC fns in ext file – ams_atm_support_v3.ex.c
T i m e (s)
0 .7
Calculate bandwidth parameters
0 .5 0 .4 0 .3 0 .2 0 .1 58.1
54.7
51.3
47.9
44.5
41
37.6
34.2
30.8
27.4
23.9
20.5
17.1
13.7
10.3
6.84
0
0 3.42
Reassign PCR, MCR and SCR based on MRAN predicted values
Link Utilization
0 .6
5b
0 .0 6 0 .0 5 0 .0 4 0 .0 3 0 .0 2 0 .0 1 1775
1670
1566
1462
1357
1253
1148
1044
940
835
731
626
522
418
313
CMP = ((PCR