1130

IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 17, NO. 5, MAY 2005

Performance Analysis of a Limited Number of Wavelength Converters in an Optical Switching Node

Hailong Li and Ian Li-Jin Thng

AbstractIn this letter, we present studies of an optical switching (OS) node utilizing a limited number of wavelength converters (WCs). The study stems from practical observation that WCs are expensive. Consequently, each output wavelength may not necessarily have its own WC and has to share a limited pool of WCs with other output wavelengths. In order to improve the utilization of the limited number of WCs, a share-per-ber (SPF) mode is proposed for the OS node. Subsequently, a two-dimensional Markov chain model of SPF is presented to evaluate its performance. Numerical results are presented to verify the accuracy of the analytical model. Index TermsOptical burst switching (OBS), optical packet switching (OPS), optical switching (OS), share-per-ber (SPF), wavelength conversion, wavelength routed networks.
Fig. 1. Structure of optical switch with SPF WCs.

I. INTRODUCTION

ANY PROMISING optical ber networks, e.g., optical burst switching (OBS) [1], optical packet switched (OPS) [2], wavelength routed [3] networks utilize optical switches to switch trafc through a node. Optical switches use wavelength converters (WCs) rather than opticalelectronicoptical elements to achieve statistical wavelengths multiplexing performance. Consequently, a bufferless, all-optical switching path can be achieved end to end. However, many of the abovementioned optical switching (OS) scenarios assume that WCs abound in the system so that any optical input can be converted to any output wavelength without blocking if needed. It is possible that, in order to lower the cost, an optical switch may be equipped with a reduced number of tunable optical WCs. In such a scenario, apart from output wavelength contention, a WC contention problem also needs to be addressed. This letter, thus, presents analytical solutions to addressing the WC contention problem and answers an important question of whether it is possible to use a limited number of WCs to achieve similar performances to a system with a full number of WCs. When limited WCs are used, some form of sharing policy must be implemented. In this letter, we focus on the simplest sharing policy which we refer to as sharing per ber (SPF). In SPF, as shown in Fig. 1, WCs are only shared within one particular output ber. If a new optical input needs to use a WC to convert itself to another wavelength on this output ber, one available WC from the limited WCs bank will be assigned to

this new optical input. If either WC or wavelength is not available, the optical data is dropped. More detail about structure of SPF can be found in [4] and [5]. A related work on limited WCs is found in [4] and [5]. Eramo et al. proposed a mathematical method (not based on Markov chain analysis) to evaluate the number of WCs required in a synchronous slotted OS network under SPF mode. However, there is currently little or no theoretical analysis for the performance evaluation of limited WCs in the case of asynchronous trafc in OS network. In this letter, the case of asynchronous Poisson trafc with variable optical data length is considered. A new mathematical analysis, modeled upon a two-dimensional Markov chain, is presented to evaluate the performance of a bufferless OS node employing a limited number of WCs. The main motivation for considering Poisson trafc is because recently more and more independent studies have shown that asynchronous trafc on the internet is very similar to Poisson or short-term Poisson [6][8]. More specic analysis on OS [9], [10] systems also shows that the trafc in such systems is similar to Poisson. II. ANALYSIS OF A LIMITED NUMBER OF WAVELENGTH CONVERTERS It is well known that for the case of an ideal number of WCs, the number of WCs is equal to the number of output wavelengths , and the drop probability of the system can be determined by the ErlangB formula. For the case of limited WCs, the scenario where is now considered. Consequently, the ErlangB formula gives the lower bound of the drop probability of all SPF scenarios where , since optical data may be dropped due to the lack of WC. We also assume that optical data arrive on each wavelength with equal probability, i.e., uniformly distributed amongst the wavelengths. No assumption is required to be made on the size

Manuscript received October 21, 2004; revised December 7, 2004. The authors are with the Electrical Computer Engineering Department, National University of Singapore, Singapore 119260, Singapore (e-mail: engp2363@nus.edu.sg; eletlj@nus.edu.sg). Digital Object Identier 10.1109/LPT.2005.845663

1041-1135/$20.00 2005 IEEE

LI AND THNG: PERFORMANCE ANALYSIS OF A LIMITED NUMBER OF WCs IN AN OS NODE

1131

distribution of the optical data. Size distribution is not required here because the analysis does not involve queues. This is similar to the ErlangB formula where it is also applicable to a queuesystem [11]. Recent researches have shown less that the optical data size distribution in OBS networks is either Gaussian or xed [9], [10] (i.e., not exponentially distributed). However, our analysis is still applicable since there are no constraints on the size distribution. We denote and to be the arrival rate of optical data on the ber and the service rate of each wavelength. Therefore, the trafc load on each wavelength is . The twodimensional state indicates that there are wavelengths in use by optical data, and WCs in use by some of these wavelengths at the same time. It is clear that and . We now determine the state probability of the . The state probabilities provide the elementary state building blocks to obtain all other probability measures related to the overall performance of the SPF node, for example, the overall drop probability (see Corollary 3). We now present the analysis for determining . Lemma 1: The state probability (for all valid states and ) can be obtained from the following simultaneous equations:

for all

and

(1) where , , , and are transition speeds for various scenarios to be described later. Proof: In Markov chain analysis, the state transition probability in/out of each valid state is required. To simplify our analysis, we will only present the transition which is outgoing from , since any incoming transition is also an outgoing state transition from some other state . to , for . This scenario indicates that the wavelength of the incoming optical data can be scheduled on an available wavelength of the output ber. The incoming optical data does not require any WC to nd a suitable output wavelength. Thus, the wavelength of the new optical data must correspond to one of the currently unused wavelengths. Thus, the transition speed is . to , for and Case 2: . This case indicates that the wavelength of the incoming optical data correspond to one of the wavelengths currently in use. Thus, the optical data has to use one WC to nd a suitable output wavelength. Thus, the transition speed is . to , for and . This Case 3: case indicates that a optical data not using any WC has just been sent out completely. As there are optical data not using WCs, the transition speed is, therefore, . to , for and . Case 4: This case indicates that an optical data using one WC has just been sent out completely. As there are Case 1:

Fig. 2. Markov chain state transition diagram of SPF. (a) State transition for state (i; j ). (b) Entire state transition diagram.

optical data using WCs, the transition speed is, therefore, . From the description of the four transition cases, the state transition for state is shown in the Fig. 2(a). It can be seen that there are at most eight transitions in/out of the state . Including the fact that the sum of all state probabilities is equal to unity, the simultaneous equations in (1) are valid. The entire state transition is shown in Fig. 3(b), which is a trapezium. End proof. We now analyze the solvability of (1) via the number of variables and equations. Corollary 1: The number of states in the simultaneous equations of (1) is . Corollary 2: The simultaneous equations in (1) in are solvable. There are state equations (one equation per state) and an additional sum-to-unity equation in (1). Noting that for Markov transition diagram, there is always one redundant state equation in (1), which (any one) can be deleted. Therefore, combining with the unity equation, there are enough equations to solve for . the state probabilities Corollary 3: From state probabilities , many useful parameters can be obtained as follows.

1132

IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 17, NO. 5, MAY 2005

Fig. 3. Tail distribution function of SPF with different number of WCs. Both theoretical and simulation values are plotted with Gaussian, Exp, Fix optical data size distributions with K = 16,  = 0:8, M = 8; 12; 16.

Fig. 4. SPF drop probability versus number of WCs. Both simulation and theory results are plotted with different data size distributions for K = 16,  = 0:4; 0:8.

We denote random variable as the number of WCs being used. Thus, the tail distribution function of used WCs can be written as (2) and drop probability as (3) (2)

of WCs gives similar drop probability performance as operating with an ideal number of WCs. This is due to the effect of statistical multiplexing for both wavelength and WCs in the SPF model. IV. CONCLUSION

(3) In (3), the rst term on the right side is the drop probability due to lack of WC. For this probability, we have to consider those new optical data arrivals whose wavelength corresponds to one of the wavelengths currently in use. This explains the factor. The second term is the drop probability due to lack of available output wavelength. III. NUMERICAL RESULTS In this section, we compare the theoretical results obtained by simultaneously solving (1) (using Matlab software) with numerical results obtained through simulation. The numerical simulations include optical data with distributions that are exponential, Gaussian, and xed. Fig. 3 illustrates the theoretical [see (2)] for a value of the tail distribution function different number of WCs , and for and . The simulation results obtained for exponential, Gaussian, and xed optical data size distribution are the same. The results clearly show that the SPF analytical model is in agreement with the simulation results. The function decreases faster than exponentially with increasing . Thereby, after a certain point, the usage probability of these WCs are negligible. This means it is not necessary to provide the ideal number of WCs for the SPF model. After some point, as illustrated in Fig. 4, a limited set of WCs is able to achieve almost identical performance compared to having an ideal number of WCs. Fig. 4 illustrates the drop probability versus , for and for . Similar to Fig. 3, the simulation results apply also for exponential, Gaussian, and xed optical data size distribution. Similar to Fig. 3, the simulation results veries the theoretical results obtained through Lemma 1. It is also noted increases, the drop probability decrease dramatithat when cally. After a certain point, the drop probability decreases very slowly and then levels out. The leveling of the drop probability indicates that after some point, operating with a limited number

In this letter, we have proposed a two-dimensional Markov chain analysis to model the performance of SPF with a limited number of WCs. Our theoretical analysis gives an accurate math model to evaluate the performance of a limited number of WCs. Therefore, given a loading value, the analytical model is useful for network operators to accurately predict the drop probability number of WCs. This enables the netof the system given work operator to balance performance versus cost without the need for conducting needless rounds of numerical simulations. In addition, the model is useful for different optical data distribution and is also useful for systems which employ WCs like OBS, OPS, and wavelength routed networks. REFERENCES
[1] C. Qiao and M. Yoo, Optical burst switching (OBS)A new paradigm for an optical internet, J. High Speed Networks, vol. 8, pp. 6984, 1999. [2] M. Renaud et al., Network and system concepts for optical packet switching, IEEE Commun. Mag., vol. 35, no. 4, pp. 96102, Apr. 1997. [3] D. Banerjee and B. Mukherjee, Wavelength-routed optical networks: linear formulation, resource budgeting tradeoffs, and a reconguration study, IEEE/ACM Trans. Netw., vol. 8, no. 5, pp. 598607, Oct. 2000. [4] V. Eramo and M. Listanti, Packet loss in a bufferless optical WDM switch employing shared tunable wavelength converters, J. Lightw. Technol., vol. 18, no. 12, pp. 18181833, Dec. 2000. [5] V. Eramo, M. Listanti, and P. Pacici, A comparison study on the number of wavelength converters needed in synchronous and asynchronous all-optical switching architectures, J. Lightw. Technol., vol. 21, no. 2, pp. 340355, Feb. 2003. [6] J. Cao, W. S. Cleveland, D. Lin, and D. X. Sun, Internet Trafc Tends Toward Poisson and Independent as the Load Increases, in Nonlinear Estimation and Classication. New York: Springer, 2003, pp. 83109. [7] T. Karagiannis et al., A nonstationary poisson view of internet trafc, presented at the 23rd Ann. Joint Conf. IEEE Computer and Communications Soc. (INFOCOM 2004), vol. 3, Mar. 711, 2004, pp. 15581569. [8] R. Morris and D. Lin, Variance of aggregated web trafc, in IEEE Proc. INFOCOM 2000, vol. 1, Mar. 2630, 2000, pp. 360366. [9] X. Yu, Y. Chen, and C. Qiao, A study of trafc statistics of assembled burst trafc in optical burst switched network, in Proc. Opticomm, vol. 4874, 2002, pp. 149159. [10] H. Li and I. L.-J. Thng, Edge node memory usage in optical burst switching networks, Photon. Netw. Commun., submitted for publication. [11] S. M. Ross, Stochastic Processes. New York: Wiley, 1983, pp. 168171.