Hiam Hiok Lim and Bin Qiu School of Cmputer Science and Sofare Engineering PO Box 26, Monash University, VIC. 3800, Australia Abs/ract-Trafc pricions bave beeD demonstrate "jtb tb cpbil ity 10 imprve Detwon .fceDe and QS in brdband AT n_r Rent rnb sbows tbat rur logic pricon outpmrs cnventona autorsion pricions [11 [21_ Te appliction of fzzy logic 0"0 bas a potential t coDtrl trafc mor efectey_ In this papr, wo prps tbe use ortbe frl logc prdiction on cnnecion admisson contrl (CAC) and con gestion contrl 00 bigh spee oetwor. We frt modeled trafc cbarctor isties using an oo-lioo fzy logc prictor on CAC. Simulation rslts sbow tbat rrl logc priction imprves tbe efcieney of bt conentiooal aod measurmen-base Clo adiioo, tbe measurmen-base appracb i cocporating fuzzy logc infernce and using fuz y logic pricion is sbown to achiee bigber netwon utilization wbile maintaining QoS. We tben ap plie tbe fzy logc predictor to Cag.StOD contl i wbicb lbe AR queue is esmated one round-trp in advance. Simulatioo "ults sbow that tbe fzzy logic contrl scbeme sigifcntly rduc conerence time and over all bufer ruirements as compar wih conventional schemes Kewrr-fz logic prictioo, asyncbrnous transfer moe (AT) network, quality of serice (QoS), connection admission conlrl (CAC), congestion contrl. I. INTRODUCTION According to the ITU-T Recommendation 1.371 [3], there are to categories of control mechanism for AM networks: pre ventive and reactive cntrols. Preventive control is open-loop, meaning the control algorithm makes decision based on its local node, such as link capacity and bufer availability in the system. Reactive control is close-loop in which network situations are feedback from the network to the sources. CAC is an example of preventive control. Its main role is to determine whether a new connection request should be accepted or rejected. The decision is based on the availability of resurces (bandwidth and bufer) and whether the OoS can b gaanteed for a new connection without degrading the promiserl OoS of ex-. isting connections. Conventional CAC relies on analytica[ mod els in which parameters are based on a priori or user-specifed trafc descriptors [4] [5] [6]. However an analytical model may not be able to capture the statistical nature of trafe sources. This is because VBR trafc characteristics such as Sustainable Cell Rate (SCR), Intrinsic Burst Tolerance (IBT) and Cell Delay Variation (COY) are difcult for users to specify precisely. Im proper specification of trafc characteristics could lead to either degraded OoS or lin under-utilization. Explicit rate (ER) switch algorithm for congestion control is an example of reactive control. II uses ER feld of the Re SUTce Management (R) cells to convey network information to sources. The main challenge of a feedback control is long propagation delay which can cause bufer overown and QoS degradation. High utilization, low queueing delay and fairness are thc main goals of a switch scheme. Switch algorithms use either static threshold values [7] [8] or complex mathematical expressions [9] for congestion control have an inherent disad vantage. They cannot adapt well to traffc fuctuations, resulting in high cell loss or low link utilization in some situations. This paper proposes two traffc controllers based on fzzy logic prediction. Firstly, based on the concepts of free band- 0-7803-7293-XOI/$17.0 2001 IEE 99 width and the adaptive weight factor in [10], this paper proposes a novel measurement-based CAC using fuz y logic approach. Secondly a novel closed-loop fuzzy logic congestion controller is proposd. It main aim is to improve the bufer requirements and convergence time on two switch algorithms - Explicit Rate Indication for Congestion Avoidance (ERICA+) [9] and Bi directional Explicit Rate Marki ng Optimizer (BEMO) [11]. In ths study, our main goal is to exploit the avantages of fzy logic prediction and control on broadband cmmunication net works. The remainder of this paper is organized as follows. A de tailed deription of the proposed trafc cntrollers is presented in Section I. Network simulations and results are discussed in Section III. A summary which concludes the paper is presented in Section I. II. Fuzzy LOGIC ApPLICATION IN BROADBAND NETWORKS A. Connection Admision Contrl Fig. l(a) shows the proposed measurement-based fuzy logic connection admission cntroller. It consists of a fuzzy logic traf fc predictor, a fzy logic adaptive weight factor estimator and a connection admission decision controller. When a new connection request arrives, the new trafc trace is predicted using the multi-step ahead predictor. Predicted Peak Cell Rate (POR) is obtained at the end of a measurement inter vaL This predicted PCR is compared with the mini mum avail able free bandwidth on all relevant [inks which the new connec tion will traverse. I it is higher than the minimum available free bandwidth, the new connection is rejected; otherise, the con nection is accepted. A shown in Fig. l(a), the arrival rate, trafc load and queue length are red into a fuzzy logic inference system which pro duces a weight factor (Fuzy Logic Adaptive Weight Factor - F AW F. This weight factor adjusts the available free band width adaptively according to the trafc conditions as expressed by Equation (1). Available Free Bandwidth = Adaptive Weight Fac/or (1) x Aggregate Free Bandwidth Triangular and trapezoidal shapes are adopted for both input and output membershp functions and are shown in Fig. 2. Table I shows the linguistic rles for the fzzy logic adaptive weight factor estimator. The design of the fuzzy linguistic rles takes into consideration of the following conditions: Under severe congestion: the arrival rate of the input trafc increases rapidly, the queue length and trafc load at the output link are normally high, then the adaptive weight factor should be low or moderateJow, to reduce the available free bandwidth. Under a light load condition: the queue length is low, the input traffc and the outgoing link load are low, the adaptive weight 201 IEEE Intemational Fuzzy Systems Confernce ! - (a) The Cnnection Amission ControllI (b) The Congestion Controller. Fig. L Application of lbe Fuzzy Logic Prediction and conlrol. #I"w ,lmI AdIliv WIPar(F A W F) Fig. 2. Membrhip fnctions of Ihe linguislic values for representing the lin glistic variables "current YR queue length", "measred trafc loa" (same for "arrival rate of the input trafc") and adaplive weighl faclor (F AW F). factor should be high, so that the proposed CAe can increase the amount of available free bandwidth in order to accept more cnnections. Under moderate lnad condition: the weight factor is medium s-that the. proposd CAe keeps accepting new connections a long as the QoS requirements are satisfed. B. Congestion Contrl Fig. 1() shows the prposed fzy logic congestion con troller. It consists of a fuzzy logic predictor and a fuzzy logic target utilization factor generator. Separate output queues for diferent categories of serices, i.e. the CBR/BR queue and the ABR queue, are considered. Since CBR and VBR trafc have distinct tolerances for delay, jitter and cell loss ratio, they are sered at a higher prorty than ABR traf fc. The predictor predicts the AR queue one round-trip delay in advance. This predicted queue value, together with the total queue growth rate and curent ABR queue length are provided to the fuzzy logic factor generator which produces a target uti lization factor (T f). This factor varies the target ABR capacity dynamically according to the bufer cndition one round-trip de lay in advance. TABLE I LrNGUISTIC RULES FOR THE Fuzzy LOGIC ADAPTIVE WEIGHT FACTOR ESTIMATOR. Irq i. and loi is I aDd Ar Rite is /I then FAWF is I - lo o ow high low ow me-um moo. rate_hIgh low lo high meium lo medium lo or medium medium lo medium high roerateJow low high lo or medium mediulIl lo hIgh high moeraEeJow medIUm tow low moeraEe_high medium lo medium or hIgh medIum meIUm medium meOlum medium hIgl moderaEeJow high low roeraleJow high medium lo moerale-.lo high medium medium or high lo hIgh high low mod.rate.o htgh high medium or high lo q:current VBR queue; loa : trafc load; Arr cRate: Ara rate of the input Irafc Target ABR Capacit = Target Utilization Factor x Total ABR Capacit (2) When a BRM cell (and FRM cell in the BEMO case) is re ceived by the switch, the explicit rate switch algorithm allocates the explicit rate fairly among AR sources based on this dynamic capacity. Hence the computed ER value has been taen into cn sideration of bufer situation one round-trip delay in advance. A a consequence, the bufer is less likely to be overflown or empty. A shown in Fig. 1(), the trget utilization factor (T/) is gov ered by three input variables, i.c. the current and predicted ABR queue lengths and the total queue growth rate; and the lingistic information stored in the rule base. Triangular and trapezium shapes are adopted for both input and output membership fnc tions. Due to space limitations, detailed design of the member ship fnctions and fuzzy linguistic rles can be found in [12]. Ill. SIMULATION RESULTS AND DISCUSSION A. Connection Admission Contrl Fig. 3 shows the simulation model used in the study. There are n elisting connections, feeding into a single server queue. The specifed QoS requ irements (e.g. CLR) of these n connections are curently maintained by the system. To ensure that the same QoS for the n + 1 (including the new arival) connecti ons are met, this new connection wilt be admitted if the predicted PCR 100 Fig. 3. Simulation Model fo Ihe CAe. TABL !I TRAFC TYPES PARAMETER SETTINGS. is lower than the minimum available bandwidth of all relevant links. Three types of VBR trafc trace are used (II - D). All VBR sources are modeled as a Poisson distributed on-of source which alternate between active and idle periods, and transmit cells with an independent exponentially distributed mean bit rate during ac tive periods. Table II shows the trafc types used in the simula tion. Note that type 3 (3) trafc is a combination of type 1 (II) and type 2 (I2) trafc. The remaining parameters, CLR is fixed at 10 - 6, the link rate is set at 155.52 Mbps. Three methods are used to obtain the trafc descrptors of a new connection, namely, a priori, and measurement-based with and without fz y logic prediction. These methods are applied to a conventional equivalent capacity CAe [4J and the proposed measurement-based CAe. Note that for the conventional CAC, trafc descriptors are composed of peak cell rate, mean bit rate and burst length, whereas for the fuzzy logic CAC, peak cell rate is the only trafc descriptor required. PCR is considered as it is simpler to determine compared with mean bit rate or burst length. A.l Performance Evaluation Fig. 4 illustrates the efciency achieved for the conventional and the fuzzy logic CAC with various traf c descriptors and a buf er capacity of 1000 cells under diferent trafc situations. There are no cell loss for both types of CAe. The Efects of Fuzzy Logic Prediction In the conventional CAC, traf c dcriptors predicted by fuzzy logic achieve a distinctive high utilization compared with the a priori and measured non-predictive trafc descriptors. As shown in Fig. 4(a), the predictive traf c descriptors improve efciency by about 27% and 19% compared to the a priori and measured non-predictive trafc descriptors respectively. Fig. 4(b) shows that the improvement is less impressive but stilt distinctive on the fzy logic CAC as cmpared with the conventional CAe. The overall imprvement in efciency when using the fuz y logic predictive traffc descriptor is about 10% and 4% compared to the a priori and measured non-predictive trafc descriptors respectively. This is mainly because statistical multiplexing gain has fully exploited in the fuzzy logic CAe. In summary, the fuzzy logic predictor provides a sof and ac curate estimation of traffc parameter(s) for new connections. A a result, the perforance of the CAC is signifcantly improved. 101
N_.bmJ..a../kJ " ..\ 1 " ' '- '-- . I ,e., -,.Y Fig. 5. Network Moel for the Generic Fairess Congration-2 (GFC). TAL 1lI MAX-MIN RATES FOR THE GFC2 NETWORK, I Grup I No otVCs I Rt. (bp) I A 3 10.0 B 3 5.0 J J:.Q D 1 35.0 E 2 3).0 F 1 10.0 G 7 5,0 H Z 52.' Efciency Evaluation Fig. 4(b) shows that the proposed fuzzy logic CAC produces consistently higher efciency for all trafc compared with the conventional CAe. From the fgure, it can be seen that efciency of over 90% is achieved for types 1 (II) and 2 trafc () and 80% for type 3 trafc (I3). Compared the efciency improvement between the fuzzy logic CAC and the conventional CAC using predictive traf c descrip tor(s), with type 3 (3) trafc, for example, efciency is im proved by 19%. ,- This is due to the fact that network performance parameters are used as input linguistic variables. A a result, the fzy logic CAC uses more information for making connection admision decisions than the conventional CAe. B, Congestion Contrl In this section, we present the simulation results to illustrate the performance of the fzzy logic congestion control scheme and its advantages over the conventional scheme. We applied the fzzy logic congestion control scheme and in cmparison to con ventional congestion control scheme on the ERICA+ and BEMO algorithms. Fig. 5 shows the network model used in our simulation. It is one of the benchmark confguration recommended hy the ATM Forum [13) and is also known as Generic Fairess Conguration- 2 (GFC2), We used this network model to study the fairess, convergence time and robustness of the schemes. The GFC2 network consists of seven switches and 22 connec tions. These connections are grouped into eight classes (A-H) in a Wide Aea Network (WAN) confguration, The link distance, D, is assumed to be 100 km, and all links connecting sources or destinations to switches have length equal to 10 Km. It is given initial cell rate, ICR=PCR=150 Mbps. All sources are assumed to be AR source. The expected max-min fairness rate for the diferent group of ve are tabulated in Table Ill. From the simulation, all shemes are able to converge to the expected max-min fair share rates as shown in Table III. Ta ble I shows the convergence time of the allowed cell rate (ACR) and queue length for both schemes. It can be obsered that the fuzy logic control scheme has a faster convergence time com- (a) Colenlionai CAC. (b) Fuzy logic CAe. Fig. 4. The efcienc achieed for the conentional and Ihe fz logic measurement-based CAC with various trafc descriptors and a bufer capacit of 100 cells under diferent trafc traces. TABLE [V COMPARISON OF CONVERGENCE TIME FOR TiE GFC2 NETWORK. Cleec \1 EC+ I I BEMO Tme (mBec) , @y , q TALE V COMPARISON OF QUEUE LENGTH FOR THE GFC2 NETWORK. Qeue ER+ BEMO (clls) Cmrettuqu 'uz convenbona Fzy SWI 41 7 9 t 4175 4116 4110 125.15; 57.03 110.73 49.29 3.4* 3. 16 2.45 2.13 SW2 4023 4013 3898 3887 265.62 57.6 237.12 51.16 27.69 5.1B 2.26 1.84 SW3 6771 4746 6205 5035 134.50 60.6 120.36 5199 9.62 8.44 2.69 2.60 SW4 1424 11B3 1320 1154 445.02 253.90 348.78 97.14 50.67 41.83 6.11 2.94 SW5 1677 109 1169 1 40.90 214.11 181.48 97.18 49.97 12.70 23.69 2.52 SW6 673 6129 6233 6163 137.58 58.62 121.97 58.56 15.39 7.94 3.32 3.30 tmaximum; lmcan; .queue deviation pared to the conventional scheme for the respective explicit rate algorithms. Table V shows the maximum, average and deviation of queue length of switch 1 to switch 6 for both schemes. There are slight diferences in the maximum queue length since it depends on the ICR, the distance of the feedback path and the queue control functions. Once the feedback is established, the fuzzy logic con trl scheme is shown to have a lower maximum queue length, es p ecially the SW3, SW4, SW5 and SW6 queues. The average queue length and queue deviation are computed after 400 msec when the rates have converged. Generally, the fuzzy logic control scheme has a much lower and stable queue compared to the conventional scheme. Hence the fuzzy logic cntrol scheme is shown to be able to optimize both the transient and steady state responses, i.e. con verges quickly to a slid steady state from the initial conditions, and drains the queue produced in the transient phase rapidly. IV. CONCLUSIONS Ths paper has presented two efcient trafc cntrol schemes based on fuzzy logic prediction. We have demonstrated that fzzy logic prediction improves the network efciency of the conventional and measurement-based CAe. We also show that Qo measurements are improved with the application of the fzzy logic predictor on the AR queue. In addition, trafc con trollers incorprating fuzzy logic inference are proved to be more efecive than conventional schemes. REFERENCES 11] B. Oiu. A Pred,ctic Fuzy Lgic Congestion Avoidance Scheme. Proc. IEEE GLOBECOM'97, Ariwna, US, voL 2;pp. 967971, Nov 1997. 12] B. Oiu, S. Sutharan, ad H. R. Wu. Neuro-fuzy Modeling and Prediction of VBR MPEG Video Soures. Journal ofR.al'Tune Imging, vol. 5;pp. 359-363, 199. [31 !-T Recommendation 1.371. Traf;Control and Congestion Control in B-ISDN,1996. (4] R. Guerin, H. Ahmadi, and M. Naghshineh. Equivalent Capacity and Its Appliction to Bandwidth Allocation in High-Speed Networks. IEEE Jour nalon Selected Areas i Communications, vol. 9(no. 7);pp. 968--981, Sop 1991. [5] R. J. Gibbens and P. J. Hunl. Efective Bandwidth, for the MulIi-typ UA Channel. Queuing Systems, vol. 9;pp. 1727, 1991. 161 F. P. Kly. Efolie Bandwidths at Multi-class Queue,. QueuingSystems, vol. 9:pp. 5-15, 1991. {7] L Roberts. Enhanc PRCA (Proportional Rate-Control Algorithm). ATM Form conrri/tion 94.75RI, Aug [994. [8] F. M. Chiu .. i, A. Arulamhal,m, Y Xia, and X-Q. Chen. E.plicit Rate ABR Schemes Using Trfc Load as Congestion [ndicato< Comu"r Colli catins and Netork. Prc. Sith Interational Confernce, pages 764 197. (9J R. Jain, S. Kalyanaraman, R. Goya, S. Fahmy, and R. Viswanathan. ER IC Switch Algorithm: A Cmplete Description. ATM FDrum contrition 9-Il7, Aug 1996. (10) M. Zukerman and T. K. Lee. A Framework for Real-time Measuremenl based Connection Admission Control in Multi-serLce Networks. Prc. IE GLOBECOM'98, vol. 5:pp. 2983--2988,1998. (11) J. J. Tirtoidjojo and B. Qiu. An Improved Fa,r Share Calculation and Updating Scheme for ABR Se!Vice. Prc. IEEE ISPACS'98, vol. Z;pp. 933--938, Nov 1998. (12) H. H. Lim and B. Qi . Fuzy Lgic Target Utilialion ,nd Prediction for Trafc Control. Prc. IEEE GLOBTOM 200, San Frncisco, USA, vol. 3;pp. 1641648, Nov 200. [13) R. J. Simco. Test Congurations for Fairess and OtherT.sts.ATM Forwn contribtion 94.f557, Ju11994. 102