Fuzzy Active Management for Assured Forwarding Traffic in Differentiated Services Network E.S. Ng, K.K. Phang, T.C. Ling, L.Y. Por Department of Computer Systems & Technology Faculty of Computer Science and Information Technology, University of Malaya ezylite@perdana.um.edu.my; kkphang, tchaw, porlip@um.edu.my Abstract-Congestion control mechanism such as Random Early Detect (RED) allows more efficient usage of network resources. RED uses static s (high and low) in combating congestion. However, as network becomes more dynamic with the introduction of multimedia and real time application, this static approach becomes inefficient. This paper proposed a dynamic RED mechanism, namely, Fuzzy Active Management (FuzAQM). FuzAQM monitors the condition of a network in real time and changes its RED parameters based on the congestion level using fuzzy control. FuzAQM is implemented and tested using Ns2 network simulator. The simulation results showed that the proposed mechanism improves the total network throughput. FuzAQM also ensures a fairer treatment of lower priority packets and at the same time maintains the throughput of higher priority packets in the network. I. Introduction The increasing important of Internet and the introduction of multimedia contents and web applications triggers the requirement of higher bandwidth and quality of service. There is a need to redesign the Internet architecture to ensure better utilization of the Internet resources. More efficient congestion control algorithms and Quality of service (QoS) mechanisms are needed [1]. This has initiated the proposition of Internet services such as Integrated Services Architecture (ISA), Random Early Detection (RED) and Differentiated Services (DiffServ) [2]. The purpose of DiffServ is to deliver an aggregated quality of service in IP networks. Further advancement has been witnessed for DiffServ framework with the integration of Active Management such as RED to preferentially drop packets before serious congestion begins [3]. The combination of DiffServ with RED is an good way of realizing service differentiation. RED uses two s to control drop probability of a packet. These s, once set, will remain unchanged (static). However, as network traffic is dynamic in nature, the static nature of RED makes it unable to cope with the dynamic nature of the network. In view of this, in this paper a dynamic queue management mechanism for DiffServ network is proposed. This enables the queue management to be more dynamic and responsive towards changes in the network. The rest of the paper is structured as follows. Section II presents the related works. Section III discusses the proposed FuzAQM mechanism in details. Section IV tables the simulation environment and results. Section V analyzes the simulation result and Section VI is the conclusion. II. Related works The proposed FuzAQM model uses fuzzy logic to determine the condition of the network and manipulate the queue management accordingly. Several researches in this area using fuzzy logic have been proposed. [4] focused on developing a drop based congestion control mechanism, which control or drop packets upon its arrival and drop those packets with lower service precedence first before dropping packets with higher service precedence. Reference [5] presented an active queue management scheme called Fuzzy Explicit Marking in DiffServ framework. This model uses a fuzzy controller to examine the dynamic network changes and drops packets or resets packets Explicit Congestion Notification (ECN) bit accordingly. Reference [6] defined a congestion index to indicate the degree of network congestion. An intelligent packet dropping mechanism based on fuzzy logic is then
being used to optimize router performance. Reference [7] examined a fuzzy based Connection/Call Admission Control to provide effective congestion control in ATM network. It is clear from these researches that fuzzy based approach gives better control than conventional systems and in maintaining network QoS. III. Proposed FuzAQM Mechanism The proposed Fuzzy Active Management (FuzAQM) model uses fuzzy logic to convert the traditional static queue management model into a dynamic model. FuzAQM uses the current queue occupancy and the rate of change of queue as input to its rules to determine whether to pass or drop a packet. Fig. 1 shows the current queue length membership function graph and Fig. 2 shows the rate of change of queue membership function graph. These membership functions are derived directly from empirical tests. Fig. 1: Current Length Membership Function. Fig. 2: Rate of Change of Membership Function. It should be noted that at the beginning, a set of initial values of the RED minimum and maximum s and drop probability is assigned to each type of packet. As packets begin their transmission in the network, these values are reset and fine tuned dynamically according to the degree of seriousness congestion in the network. In this way, a dynamic active queue management mechanism is created. Parameters used in the inference proves are: the current queue length and average queue length. The latter will be used to find out the rate of change of queue. A typical fuzzy inference rule takes the following form: If Input X and Input Y, then Output is Z 1. 2. 3. 4. 5. 6. 7. 8. Eight rules have been defined and listed as follow: Change of is Decreasing Fast, then Output is Z 1 Change of is Decreasing, then Output is Z 2 Change of is Increasing, then Output is Z 3 Change of is Increasing Fast, then Output is Z 4 Change of is Decreasing Fast, then Output is Z 5 Change of is Decreasing, then Output is Z 6 Change of is Increasing, then Output is Z 7 Change of is Increasing Fast, then Output is Z 8 where X and Y represents the current queue length and the rate of change of queue respectively. Each input will be inference to produce the weighted output. The final output will be defuzzified using the summation of all rule output (w i z i ) and averaged by the summation of all the output strength (z i ).and the value produced will reflect the current congestion situation at the bottleneck link and further assessed. //sugeno method From the fuzzy inference process based on these two inputs, a final output will determine the condition of the bottleneck link. The first step in assessment step taken is to cast the value into 9 categories of seriousness of the bottleneck link. The value used is from 1 to 9 where 1 represents the least serious bottleneck link condition and 9 represents the most serious. Further action could be then taken to reconfigure. the RED s to adapt to the condition of the bottleneck link accordingly. Based on these 9
situations or degree of seriousness, the corresponding action will be taken. In this case, two virtual queues, vq1 and vq2 are maintained for the higher priority (DiffServ Code point 20) and lower priority packets (DiffServ Code point 21) respectively. Each queue maintains its minimum and maximum as well as its drop probability. Vq1 in general will have higher value for minimum and maximum and lower drop probability compared to vq2. Each time, when a packet enters the physical queue, the fuzzy rules will be fired and return a value. Based on the return value and the type of packet (higher or lower, the s of the virtual queues are reset. In other words, dynamic s are used in the proposed mechanism instead of the static one as in RED. In RED these s will remain unchanged once they are set. If the packet is a higher priority packet, the s of the high priority virtual queue will be readjusted to a higher value (as shown in Fig. 3) before the packet enters that virtual queue. This will result in more high priority to be queued and the queue utilization will be higher. Note that in this case, the s of the lower priority virtual queue are unaffected. Similarly, if a low priority packet enters the physical queue and if the network in not congested, the of the low priority virtual queue will be reset to a higher value. Fig. 4: Resetting of Thresholds when the network is congested. IV. Simulation The simulation is carried out using the ns2 network simulator [8]. The simulation uses a dumb-bell topology with 14 routers and varying numbers of source nodes depending on the traffic load to be generated. For light traffic, the number of source node is limited to 24 nodes. For heavy traffic, the number of sources is doubled. Fig. 5 shows the network topology for the simulation of light traffic. Fig. 5: Network Topology Used for Simulation Initial minimum Initial maximum Table 1 below summarizes the number of nodes and its respective roles for both simulations. New minimum New maximum Fig. 3: Resetting of Thresholds when the network is not congested Conversely, if the network is congested, the set of s will be reset to a lower value as in Fig. 4. Table 1 Type of Nodes and Traffic Load Type Quantity Light Traffic Heavy Traffic Routers 14 14 VoIP 8 16 FTP Clients 4 8 FTP Servers 4 8 HTTP Clients 4 8 HTTP Servers 4 8 Total 38 62 Two types of traffic are used: DiffServ Code point 20 and 21 representing high priority packets and low priority packets respectively. Simulations are carried out to study and compare the performance of FuzAQM under light and heavy traffic. V. Results Analysis Initial minimum Initial maximum
Throughput = Total packets successfully traveled through the bottleneck link Total packets received at the bottleneck link Fig. 6 shows the throughput of packets going from r6 to r7. On the other hand, Fig. 7 the throughput of packets going in the reverse direction i.e. from r7 to r6. 88.00 83.00 r7r6 Throughput age for Heavy Traffic Simulation results showed that when traffic is light, FuzAQM outperforms the non fuzzy counter part. 100.00 99.00 97.00 96.00 95.00 94.00 92.00 91.00 100.00 99.00 97.00 96.00 95.00 94.00 92.00 91.00 r6r7 Throughput age for Light Traffic r6r7 Fig. 6 r7r6 Throughput age for Light Traffic r7r6 r7r6 Fuzzy Fig. 7 r6r7 Fuzzy Similarly, in the case of heavy traffic, simulation results depicted in Fig. 8 and 9, FuzAQM also outperforms the non fuzzy counter part. 78.00 Fig. 9 Table 2 Average packets successfully traveled through the bottleneck link for light traffic CodePoint R6r7 R6r7 fuzzy R7r6 R7r6 Fuzzy 20 (high 301911 315623 305062 316354 21 (low 122022 162745 122386 159094 Table 3 Average packets successfully traveled through the bottleneck link for heavy traffic CodePoint R6r7 R6r7 fuzzy R7r6 R7r6 Fuzzy 20 (high 491941 431260 492692 435888 21 (low 68835 147575 65339 148692 The throughput of packet of higher and lower priority is depicted in table 2 and 3. In table 2, when the traffic is light, the fuzzy approach demonstrated higher throughput for both high and low priority r7r6 r7r6 Fuzzy r6r7 Throughput age for Heavy Traffic 88.00 83.00 78.00 r6r7 r6r7 Fuzzy Fig. 8 packets. In table 3, when the traffic is heavy, using the non-fuzzy approach, the low priority packets are choked out by higher priority packets-- from
approximately 120000 packets to 68000 packets. However, when the fuzzy approach is used, there is a small drop in the high priority throughput but a large increase in the lower priority packets. The overall throughput of higher priority packets is still much higher than the lower priority packet but without choking out the lower priority packets. VI. Conclusion The simulation results showed that the proposed FuzAQM model achieved its objectives. It is clear from the simulation results that FuzAQM improves the total throughput by 0.79% to 6.46% depending on the traffic load and maintaining fair treatment for lower priority packets while maintaining at least 87% of throughput for higher priority packets in the network. It can be concluded that FuzAQM, an active queue management for DiffServ traffic, works better than classic type of non-responsive queue management model. Future enhancement includes expanding the research to include various types of packet metering scheme such as time sliding window with 3 color marking, token bucket and etc. The dynamic scheme can also be used to reset the parameter of the token, leaky bucket and other metering mechanism in conjunction with the network congestion level. [2] Blake, S. et al. An Architecture for Differentiated Services, RFC 2475, December 1998. [3] Floyd, S., and Jacobson, V.,Random Early Detection gateways for congestion avoidance IEEE/ACM Trans. on Networking, 1(4) August, 1993, pp. 397-413. [4] Zhang, R. and Ma, J. Congestion control using fuzzy logic in differentiated services networks. IN: Fourth International Conference, Proceedings of the Computational Intelligence and Multimedia Applications, 30 October-1 November 2001, Yokusika City Japan, IEEE. pp. 288 292. [5] Chrysostomou, C. et al. Fuzzy Explicit Marking for Congestion Control in Differentiated Services Networks. IN: Eighth IEEE International Symposium on Computers and Communications June 30 - July 03 2003, Kemer-Antalya, Turkey, IEEE, pp. 312-319. [6] Yanfei, F., Fengyuan, R., and Chuang, L. Design of an active queue management algorithm based fuzzy logic decision. IN: Communication Technology Proceedings, 2003. International Conference on, 9-11 April 2003, China, IEEE, pp. 286 289. [7] Karthik, S., Venkatesh, C., and Natarajan, A.M. Congestion control in ATM networks using fuzzy logic IN: Proceedings of the18th International Parallel and Distributed Processing Symposium, 26-30 April 2004, New Mexico, USA, IEEE Computer Society Press. pp.162. [8] The ns Manual, The VINT Project, December 2003. References [1] Orda, A, QoS Routing: Challenges and Solution Approaches, Second International Conference on Quality of Service in Heterogeneous Wired/Wireless Networks, 2005.