Discrete Event Simulation of a large OBS Network Stein Gjessing Simula Research Laboratory & Dept. of Informatics, University of Oslo, P-O. Box 1080, N-0316 Oslo, Norway steing@simula.no Arne Maus Dept. of Informatics University of Oslo, P-O. Box 1080, N-0316 Oslo, Norway arnem@ifi.uio.no Abstract – Optical Burst Switching (OBS) is a much researched paradigm for the next generation optical Internet. We have made a detailed discrete event simulation model of OBS networks. Among other things our model includes self similar traffic sources, burst assembly with fixed and variable length bursts, burst scheduling, wavelength conversion and fiber delay lines. In this paper we simulate a large optical burst-switched backbone network using the COST 239 network topology that connects 11 European cities. The performance of this realistic network is investigated, mainly by varying the load and the number of channels (lambdas). We propose and evaluate a new method for the utilization of otherwise unused network capacity by very low priority traffic. Other interesting results include how total burst loss increase when high priority bursts are used. Keywords: Optical burst switching, discrete event simulations, traffic modeling, burst loss, burst scheduling. 1 Introduction and motivation Optical Burst Switching (OBS) is a much research paradigm for the next generation optical Internet [1]. OBS is performed on Wavelength Division Multiplexed (WDM) fibers, is more fine-grained than optical line switching, and more coarse-grained than Optical Packet Switching (OPS). The main motivation behind optical switching is to transport the data with minimal delay by keeping it in the optical domain. A control packet precedes the data burst in the network and reserves resources on the links and in the switches for the data burst. Only the control packet is con- verted from optical to electrical (and back) in each switch. Analyses of OBS network have been performed mainly by analytical models, or relatively simple simulation models often looking at the traffic over a singe link with a few traffic sources. However, also some more detailed simulation models have been developed [2]. Schlosser at al. have simulated a two node OBS network with Poisson and Pareto distributed traffic [3], and Ahmad and Malik have built an OBS simulator using the Ptolemy framework [4]. In [5] a two node OBS network is simulated in order to study blocking probabilities, and the NSF network in the USA has been modeled by several researchers [6,16]. Most of the analytical work reported uses Poison arrival rates, e.g. [1,5,8,16]. More complex traffic has been used in some papers, e.g. by analyzing OBS edge routers using traffic generated by Markovian processes [7]. Blocking probabilities and QoS have been researched by many; notable research are reported in [5,9,10,15]. The main contribution of this paper is a detailed realistic large scale discrete event simulation and traffic model of a core OBS network and some initial results obtained from running this simulator on a large network. As will be detailed in the next section, we model a large number of traffic sources with variable (IP) packet sizes, burst assembly (fixed or variable sized), burst scheduling with the possibility of wavelength conversion and fiber delay lines, QoS using longer control packet lead time (CPT), and deflection routing. Few such detailed and realistic discrete event simulation models have been published before. The most detailed model we know of uses much shorter bursts [2]. By running our simulator we find performance properties that are not revealed by analytic models or simpler simulations. Initial findings reported in this paper includes the effect on regular bursts from high priority (QoS) bursts, optimal bursts sizes, and a new method to send very low priority bursts that utilize otherwise unused bandwidth and hence do not affect regular bursts at all. The traffic load onto an OBS core network comes from IP- subnets and Ethernets. It is well known that Ethernet and IP traffic exhibit self similar properties. The statistical properties of the bursts in an OBS core network however are not well known. A contribution of our work is to implicitly find these parameters by making a detailed simulation model of the burst assembly process, using synthetically generated Ethernet and Internet self similar traffic. This paper is organized as follows. In the next section we present and discuss our simulation model. In section 3 we present a case study involving a large European network, and section 4 presents and discusses results obtained from running traffic in this network. Finally in section 5 we conclude and point at possible further work.