2013 27th International Conference on Advanced Information Networking and Applications Workshops
Modeling and Evaluation of Machine Learning based Network Management System for NGN Abul Bashar College of Computer Engineering and Sciences Prince Mohammad Bin Fahd University Al-Khobar, Saudi Arabia 31952 Email:
[email protected] WDM IP/MPLS PBB-TE
Abstract—The recent emphasis on monitoring and managing telecommunication networks in more intelligent and autonomic manner has led to the emergence and popularity of Machine Learning based Network Management Systems. In order to study the behavior and assess the performance of such NMS, it is essential that a suitable modeling and evaluation framework exists. The work presented here addresses this need and proposes an autonomic NMS which employs the prediction capabilities of the Bayesian Networks (BN) models. To achieve this, it formulates and models the BN-based Decision Support System for providing real-time decisions with regard to the Call Admission Control (CAC) problem in the Next Generation Network (NGN) environment. Simulated experiments are performed to verify the suitability and practicality of the proposed models. The novelty and relevance of this research is demonstrated through offline modeling and online performance evaluation of BNAC (Bayesian Networks-based Admission Control) by considering the metrics of Packet Delay, Packet Loss, Queue Size and Blocking Probability. The paper concludes that BNAC approach performs better than the Peak Rate CAC in terms of online CAC functionality.
APPLICATIONS / SERVICES
GATEWAY ACCESS NETWORK
EDGE NETWORK
Fig. 1.
S ES EL WIR
ACCESS DEVICES
AL TIC OP IBER F
LA N
xD SL
SA TE
LL I
TE
IP CORE NETWORK
IEEE 802.11 GSM IEEE 802.16
IEEE 802.3 xDSL EPON
NGN and its Complex Architecture
connects the multi-user access network to the high speed core network to complete the NGN architecture. From the architectural, functional and management points of view, NGN has become a highly complex, dynamic and unpredictable system which the traditional Network Management Systems (NMSs) find difficult to control and manage [2]. This situation has led researchers to explore intelligent and autonomic approaches for improving the functionality of NMSs so that they can operate in a more efficient manner to achieve the desired management goals. Machine learning (ML) is one such approach which has the capability to address the issues of improving the network management functionalities by imparting automated and intelligent data analysis techniques to network management datasets. These techniques have the ability to learn the system behavior from past data and estimate future behavior based on the learned system model [3]. In this paper we present the application of a graphical modeling technique, Bayesian Networks (BN), to assist the NMS in solving performance management problems in a network. The original contribution of this paper is the application of BN to model the behavior of network elements (e.g., the routers) and to provide a mechanism to perform probabilistic inference, decision making and prediction to achieve desired QoS goals through the process of Call Admission Control. The work presented in this paper basically answers the following research questions, which, to our knowledge, have not been addressed before, and hence we claim the novelty of
I. I NTRODUCTION The domain of telecommunication networks is evolving at a rapid pace in terms of its architecture and services due to the advances in the underlying technology and also owing to the new demands placed by the telecom consumers. The current dominant technology in this domain is the Next Generation Network (NGN), which is capable of providing converged services with guaranteed QoS whilst offering enormous savings to the network and service providers by reducing their Capital Expenditure (CAPEX) and Operational Expenditure (OPEX) [1]. A typical architecture of such a converged network is depicted in Fig. 1. The majority of the existing Internet traffic is carried over a combination of wireline and wireless communication infrastructure. As seen from Fig. 1, this infrastructure includes customer premises, telephone exchanges, base stations and core trunk network links. In essence, the core network of NGN involves a consolidation of several transport networks, each built for a different service, into one core transport network based on IP (Internet Protocol) and MPLS (Multi Protocol Label Switching). The access network is a highly complex mix of wired and wireless technologies, including ADSL (Asymmetric Digital Subscriber Line), FTTP (Fiber-To-The Premises), WLAN (Wireless Local Area Network), WiMAX (Worldwide Interoperability for Microwave Access), etc. The edge network 978-0-7695-4952-1/13 $26.00 © 2013 IEEE DOI 10.1109/WAINA.2013.184
VoIP VoD IPTV
SERVICE USERS
1473
The task of a NMS is to monitor and control the functionalities and behavior of various constituent network elements so as to achieve a desired system state, given a set of pre-defined high-level business goals. The goal could be to maximize resource utilization (translated as maximizing revenue to service providers) or minimize service response times (equivalent to enhanced customer satisfaction). To achieve these objectives, there is a requirement for a standardized and efficient NMS. Several standard NMSs like the SNMP [6], CMIP and FCAPS (Fault, Configuration, Accounting, Performance and Security) [7] are used to manage the telecommunication networks. However, since NGN is a converged network providing data and voice services, it would have to be managed by the combination of existing NMSs. One aspect which is common to all NMSs is that they deal with collections of management data typically stored in Management Information Bases (MIBs) at network nodes. To analyze and interpret such massive data sets by a human operator would be almost an impossible task. Hence, we need ML techniques to automate such tasks by utilizing rich reasoning mechanisms to find patterns which are useful for network management. Various ML approaches have been proposed to address the issues related to network management functions. Decision trees have been used to achieve proactive network management by mining the data obtained from SNMP MIB objects [8]. Bandwidth broker design can be facilitated by applying Fuzzy Logic [9]. Predictive network fault detection in a telecommunication network has been addressed using Bayesian Belief Networks (BBN) [10]. Dynamic Bayesian Networks (DBN) have also been used for detecting network faults in realtime [11]. Bayesian reasoning based software agents have been used to achieve intelligent network fault management [12]. Reinforcement Learning has been used to provide efficient bandwidth allocation in Differentiated Services (Diff-
$33/,&$7,216 6(59,&(6
% $ ,3&25( 1(7:25.
' $&&(66 '(9,&(6
(
('*( 1(7:25.
; /
$/ 7,& 23 ,%(5 )
6/
6$
7( // ,7 (
$&&(66 1(7:25.
/$ 1
['
*$7(:$
ŽƐƐͺE
ϯϬϬ ϮϬϬ ϭϬϬ Ϭ
Ϭ
ϭϬϬϬ
ϮϬϬϬ
ϯϬϬϬ
ϰϬϬϬ
ϱϬϬϬ
^ŝŵƵůĂƚŝŽŶd/ŵĞ;ƐĞĐͿ
/RZ 0HG +LJK
Fig. 7. Fig. 5.
>ŽƐƐͺWĞĂŬͺZĂƚĞ
ϰϬϬ
Average Packet Loss Comparison for Online CAC
BN Model (BNAC)
The results of the first metric (i.e. Delay) are shown in Fig. 6. It is natural to see a drop in the average delay from 320 ms to 140 ms between the No CAC and BNAC scenarios and also the performance of the Peak Rate scenario lies in between them. The value of delay obtained through the BNAC approach is very desirable for a VoIP service, where the delay is required to be lower than 150 ms. The second metric is Packet Loss and its results are shown in Fig. 7. Again, it is clear that CAC functionality limits the traffic into the link and hence lesser packets are dropped (they decrease from about 330 dropped packets in the No CAC scenario to 30 dropped packets in the BNAC scenario). Once again the performance of Peak Rate approach lies between them. The third metric to be observed is the average Queue Size whose results are shown in Fig. 8. In this figure it can be seen that there is a decrease in the average queue size from 25000 bits (No CAC scenario) to 6000 bits (BNAC scenario). The reason is clear from the fact that less packets enter the queue as a result of the CAC mechanism and hence the queue occupancy drops to maintain the desired delay and packet loss. Another important metric which is shown in Fig. 9 is the call blocking probability. It is the percentage of calls/flows/connections/packets which are blocked due to unavailable resources in the network. The lower this value for any CAC scheme, the better. But it should also correspond to a lower Delay and Packet Loss metrics. It can be seen that
bilities of all the nodes (i.e. the probability of each state (Low, Med, High) of the BN node). It can be seen that the BN structure presents the unknown dependencies between the measured variables, which cannot be normally derived by a human operator by data inspection. In situations where there are many variables, this structural relationship becomes a useful tool for the network manager to study the cause and effect variables which helps him to understand the network behavior. CPTs for each of the nodes were calculated using the EM algorithm and not shown here due to space limitations (refer [14] for details). They provide the strength of the relationships between the states of all the nodes of the BN. By studying these probabilities the network manager can deduce how the various statistics affect each other. Online Performance Evaluation: The BNAC model obtained above will now be used in making real-time decisions in an online scenario. For comparative study, the online performance evaluation will be done by running the simulation in three scenarios, namely, without any CAC mechanism (termed No CAC scenario), with the BN based CAC model (termed BNAC scenario) and Peak Rate CAC method (termed Peak Rate scenario, which is a traditional CAC scheme). The performance of CAC functionality will be evaluated using the metrics of Delay, Packet Loss, Queue occupancy and call Blocking Probability (BP).
1477
YƵĞƵĞͺEŽͺ
YƵĞƵĞͺWĞĂŬͺZĂƚĞ
was demonstrated in comparison to the traditional Peak Rate CAC method. As a possible enhancement to this work, the BNAC solution is planned to be extended to model the dependencies among multiple routers in a centralized setup. Further, it is planned to develop BNAC as a comprehensive online learning and decision support system.
YƵĞƵĞͺE
ǀĞƌĂŐŐĞYƵĞƵĞ^ŝnjĞ;ďŝƚƐͿ
ϱϬϬϬϬ ϰϬϬϬϬ ϯϬϬϬϬ ϮϬϬϬϬ
ACKNOWLEDGMENT
ϭϬϬϬϬ Ϭ
Ϭ
ϭϬϬϬ
ϮϬϬϬ
ϯϬϬϬ
ϰϬϬϬ
ϱϬϬϬ
^ŝŵƵůĂƚŝŽŶdŝŵĞ;ƐĞĐͿ
Fig. 8.
R EFERENCES
Average Queue Size Comparison for Online CAC WͺEŽͺ
WͺWĞĂŬͺZĂƚĞ
WͺE
ůŽĐĐŬŝŵŐWƌŽďĂďŝůŝƚLJ
ϭ Ϭ͘ϴ Ϭ͘ϲ Ϭ͘ϰ Ϭ͘Ϯ Ϭ Ϭ
ϭϬϬϬ
ϮϬϬϬ
ϯϬϬϬ
ϰϬϬϬ
ϱϬϬϬ
^ŝŵƵůĂƚŝŽŶdŝŵĞ;ƐĞĐͿ
Fig. 9.
The authors would like to acknowledge the support of Prince Mohammad Bin Fahd University, KSA for providing the facilities to perform this research work.
Blocking Probability comparison for Online CAC
this probability is zero for No CAC scenario, as all the calls are accepted into the network. Therefore, this scenario has the worst performance. However, the BNAC approach provides a lower blocking probability as compared to the Peak Rate and hence is better in this metric as well as the metrics of Delay, Packet Loss and Queue. This is because, in the Peak Rate method the CAC decision is purely made on the peak rate value of the incoming call and hence if there is not enough bandwidth it will be blocked. However, the BNAC approach on the other hand, looks at all possible incoming calls with different rates and service class and can just accommodate a lower bandwidth call where the peak rate approach might fail. This clearly demonstrates the intelligence and superiority of the BNAC approach over the Peak Rate approach. VII. C ONCLUSION This paper addressed the need for providing a framework for implementing a Machine Learning based Network Management Systems. To this end, it proposed and implemented an autonomic NMS which employs the prediction capabilities of the Bayesian Networks (BN) models. It successfully demonstrated that BN are capable of representing highly dynamic systems, by efficiently modeling unknown and complex relationships between network elements. BNAC solution was developed in an offline mode and then its performance evaluation was performed for metrics of Packet Delay, Packet Loss, Queue Size and Blocking Probability. BNAC’s superior performance
[1] ITU-T, “General overview of NGN,” ITU-T Recommendation,Y.2001, Dec. 2004. [2] A. Pras, J. Schoenwaelder, M. Burgess, O. Festor, G. M. Perez, R. Stadler, and B. Stiller, “Key research challenges in network management,” IEEE Communications Magazine, vol. 45, no. 10, pp. 104–110, Oct. 2007. [3] E. Alpaydin, Introduction to Machine Learning. MIT Press, 2004. [4] OPNET Technologies Inc., “Opnet Modeler 16.0,” http://www.opnet.com, (Last accessed : Oct 2012). [5] Hugin Expert A/S, “Hugin Researcher 7.3,” http://www.hugin.com, (Last accessed : Oct 2012). [6] D. Harrington, R. Presuhn, and B. Wijnen, “An architecture for describing SNMP management frameworks,” in RFC 3411, IETF, Dec. 2002. [7] “TMN Management Functions,” ITU-T Recommendation M.3400, Feb. 2000. [8] P. G. Kulkarni, S. I. McClean, G. P. Parr, and M. M. Black, “Deploying MIB data mining for proactive network management,” in 3rd International IEEE Conference on Intelligent Systems (IS 2006), Sep. 2006, pp. 506–511. [9] S. Sohail and A. Khanum, “Simplifying network management with fuzzy logic,” in IEEE International Conference on Communications (ICC 2008), May 2008, pp. 195–201. [10] C. S. Hood and C. Ji, “Proactive network fault detection,” IEEE Transactions on Reliability, pp. 333–341, Sep. 1997. [11] J. Ding, B. Kramer, S. Xu, H. Chen, and Y. Bai, “Predictive fault management in the dynamic environment of IP network,” in IEEE International Workshop on IP Operations and Management, Oct. 2004, pp. 233–239. [12] E. U. Ekaette and B. H. Far, “A framework for distributed fault management using intelligent software agents,” in IEEE Canadian Conference on Electrical and Computer Engineering 2003, vol. 2, May 2003, pp. 797–800. [13] T. C. K. Hui and T. Chen-Khong, “Adaptive provisioning of differentiated services networks based on reinforcement learning,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 33, no. 4, pp. 492– 501, Nov. 2003. [14] A. Bashar, G. P. Parr, S. I. McClean, B. W. Scotney, and D. Nauck, “BARD: A novel application of Bayesian reasoning for proactive network management,” in 10th Postgraduate Conference on Telecoms, Networking and Broadcasting (PGNET 2009), Jun. 2009, pp. 161–166. [15] A. Bashar, G. P. Parr, S. I. McClean, B. W. Scotney, M. Subramanian, S. K. Chaudhari, and T. A. Gonsalves, “Employing Bayesian belief networks for energy efficient network management,” in 16th IEEE National Conference on Communications (NCC 2010), Jan. 2010, pp. 1–5. [16] J. Qi, F. Wu, L. Li, and H. Shu, “Artificial intelligence applications in the telecommunications industry,” Expert Systems, vol. 24, no. 4, pp. 271–291, Sep. 2007. [17] H. Steck, “Constrained-based structural learning in Bayesian networks using finite data sets,” Ph.D. dissertation, Department of Informatics, Technical University Munich, May 2001. [18] S. L. Lauritzen, “The EM algorithm for graphical association models with missing data,” Computational Statistics & Data Analysis, vol. 19, no. 2, pp. 191–201, Feb. 1995.
1478
