Performance advantages of resource sharing in polymorphic optical networks

R. J. Durán, I. de Miguel, N. Merayo, P. Fernández, R. M. Lorenzo, E. J. Abril, I. Tafur Monroy, Performance advantages of resource sharing in polymorphic optical networks, Proc. of the 0th European Conference on Networks & Optical Communications (NOC 05), pp. 249-256, 2005. Performance advantages of resource sharing in polymorphic optical networks Ramón J. Durán, Ignacio de Miguel, Noemí Merayo, Patricia Fernández, Rubén M. Lorenzo, Evaristo J. Abril, Idelfonso Tafur Monroy 2 Dpt. of Signal Theory, Communications and Telematic Engineering, University of Valladolid, Campus Miguel Delibes, 470 Valladolid (Spain). Tel: +34 983 423665, Fax: +34 983 423667, E-mail: {ramon.duran, ignacio.miguel}@tel.uva.es 2 COBRA Institute, Eindhoven University of Technology, P.O. Box 53, 5600 MB Eindhoven, The Netherlands, E-mail: i.tafur@tue.nl A polymorphic network simultaneously supports several optical switching paradigms in a single physical topology. In this way, efficient and flexible optical networks can be built as resources are shared, as long as it is possible, by the different paradigms supported by the network. We show that the integration of semi-static and dynamic wavelength routing paradigms in a polymorphic network brings advantages in terms of resource utilisation when compared with two networks one semi-static and one dynamic operating in parallel. We have made a simulation study, which shows that a polymorphic network has a lower blocking probability for dynamic connections (up to four orders of magnitude) while keeping the same performance for semi-static connections than the parallel networks.. Introduction A polymorphic multiservice optical network (PMON) [, 2] is an integrated architecture that combines several switching paradigms in a single physical network, allowing resource sharing, as long as possible, among all the supported paradigms. In this way, not only does the network optimally support different traffic types and services, but it also uses its resources effectively. The framework of polymorphic networks has its origin in the work by Qiao et al. [3]. These authors propose a network architecture sliced into several virtual optical networks (VON), each one designed to support a different class of service. For each VON, a dedicated set of resources is allocated as well as a different type of switches (slow/fast) depending on the service to be provided. In contrast with that proposal, the PMON is an integrated architecture which allows resource sharing. We have proposed two polymorphic networking architectures, the Optical Circuit- Switched Polymorphic Network (OCSPN) and the Labeled Optical Burst-Switched Polymorphic Network (LOBSPN) [, 2]. In this work, we focus on the former one. We briefly review the fundamentals of the Optical Circuit-Switched Polymorphic Network and then study, by means of simulation, whether it brings performance advantages when compared to parallel networks one semi-static and one dynamic. When using parallel networks, both the semi-static and the dynamic one must be independently overdimensioned in order to be able to accommodate future traffic growing and to allow smooth logical topology reconfiguration when the

network faces traffic changes. On the other hand, dimensioning a polymorphic network with extra resources is more effective as these are shared by all the switching paradigms supported by the network. Hence, the polymorphic optical network is expected to bring performance advantages. The evaluation of such advantages is the aim of this study. 2. Optical Circuit Switched Polymorphic Networks An OCSPN transmits data over lightpaths established in a static or dynamic way. It combines semi-static wavelength-routed optical networks (WRONs), dynamic WRONs, and wavelength-routed optical burst-switched networks (WR-OBS). The architecture of this polymorphic optical network consists of two layers, an access layer and a core layer. In the access layer, traffic aggregation and classification is performed by edge routers. The core layer is composed by a number of core routers, which transport data transparently in the optical domain. In semi-static WRONs, a limited set of lightpaths is established between pairs of edge routers, thereby embedding a virtual (or logical) topology in the physical topology. Moreover, if the traffic demand associated to this scheme changes, the virtual topology can be reconfigured in order to adapt efficiently to the new scenario. In the OCSPN, the semi-static WRON is chosen as an efficient solution for supporting bursty traffic such as best-effort Internet traffic. In dynamic WRONs, lightpaths between any two edge routers are established and released on user-demand on real time. This is the paradigm employed for services requesting dedicated circuits on user-demand in the OCSPN. In Wavelength-Routed Optical Burst-Switched Networks (WR-OBS) [4], packets are electronically aggregated into bursts and sent to the destination edge router once a lightpath has been established. The WR-OBS paradigm is used in an OCSPN for the provision of services requiring bounded delays due to the capability of this architecture to provide such guarantees. For each optical paradigm described above, a certain amount of network resources (wavelengths, fibres, transmitters, receivers, etc) are either statically or dynamically assigned. As next generation networks are characterised by a dynamic behaviour and variable traffic patterns, dynamic resource allocation is the most effective mechanism because it allows a seamless adaptation to traffic pattern changes. 3. Performance advantages using polymorphic optical networks We have made a simulation study to compare the performance of a polymorphic network with parallel networks, when the same amount of resources is used. In this study, we consider an Optical Circuit-Switched Polymorphic Network that only combines a semi-static and a dynamic WRON (with resource sharing), and compare with the utilisation of two parallel networks a semi-static and a dynamic WRON (without resource sharing). In particular, we analyse the traffic congestion when the semi-static paradigm is used, and the blocking probability for the dynamic paradigm. WR-OBS networks have not been explicitly included in this study, since they are also based on dynamic lightpath establishment, and our aim in this work is to analyse only the blocking probability of such dynamic paradigms.

3. Simulation scenario We have implemented both, an OCSPN architecture and a set of two parallel networks, following the NSFNet topology [5]. Thus, the network consists of 4 nodes, each one combining the functionality of an edge and a core router. A cable between two network nodes, say A and B, is assumed to consist of two unidirectional fibres (one for transmission from A to B, and another for transmission from B to A). Although we have performed simulations equipping the nodes with different numbers of transceivers, and equipping the fibres with different numbers of wavelengths, we will only show the results for the case of eight wavelengths per fibre and 3 transceivers per node (that is, the edge router section of each node is equipped with 3 transmitters and 3 receivers). A uniform traffic matrix has been used to obtain the logical topology employed in the semi-static paradigm. For dynamic traffic, we assume that lightpath requests arrive at the network according to an independent Poisson process with arrival rate λ. The holding time of a lightpath is exponentially distributed with mean T, and the source and destination nodes are randomly chosen according to a uniform distribution. Thus, the dynamic traffic load is obtained as ρ = λ T/(N (N-)), where N is the number of network nodes. The simulator has been implemented in the OMNeT++ platform [6], with the help of AKAROA-2 [7] in order to make a correct statistical treatment of data. Simulation results will be shown with a 95% confidence interval. 3.2 Control algorithms We have assumed that the networks employ centralised control. HLDA (Heuristic Logical topology Design Algorithm) [8] has been used to design the logical topology, and so to determine the set of resources required for the semi-static WRON, and AUR-EXHAUSTIVE [9] has been selected as the method for dynamic lightpath establishment. HLDA designs the logical topology with the aim of minimising network congestion (the traffic associated to the semi-static paradigm that is carried by the most loaded link in the virtual topology), and it takes into account both the traffic matrix and the availability of physical resources. Hence, HLDA not only provides the set of semistatic lightpaths to be established in the network, but also the routes and wavelengths that they must use. This property is very useful in the OCSPN since the logical topologies obtained can always be embedded in the physical topology. One drawback of HLDA is that it does not ensure that the virtual topology obtained is connected. This means that two or more disconnected virtual subnets could be formed within the network, so that traffic may not be routed from the source to the destination node if they belong to different subnets. However, all nodes should be able to communicate with all other network nodes. Therefore, in order to solve this problem, we have designed a correction algorithm. If the virtual topology obtained is not connected, we check whether splitting in two the virtual link (the semi-static lightpath) that carries the lowest load solves the problem. If it is not solved, we restart the design of the virtual topology by first establishing one Hamiltonian circuit [0] (as long as it is possible), and thus ensure that the virtual topology is connected. Then, we use HLDA to complete the design of the virtual topology.

In the parallel networks, we use the original definition of HLDA. The algorithm has two phases. In the first one, logical links are established attending to the traffic matrix, and in the second one, it randomly assigns free resources (as long as it is possible) in order to establish additional virtual links. However, even with this random assignation, some resources (wavelengths and transceivers) cannot be assigned and are therefore wasted. In the polymorphic network, we do not perform this random assignation but directly assign all the idle resources to the dynamic paradigm, so that we improve resource utilisation by employing resources that otherwise could be wasted. Once the logical topology has been designed and established, the traffic associated to the semi-static paradigm is routed from the source to the destination node following the shortest-paths in the virtual topology. We have used AUR-EXHAUSTIVE to solve the dynamic routing and wavelength assignment problem with the objective of minimising the call blocking probability, since it has proved to be an efficient algorithm regarding blocking performance [9]. 3.3 Performance advantages in the OCSPN In order to compare the performance of the OCSPN to that of the parallel networks, we have equipped both architectures with the same amount of resources, and have analysed the congestion obtained when using the semi-static paradigm, and the blocking probability when the dynamic paradigm is used. In all cases, the congestion was the same when using the OCSPN and the parallel networks. Hence, we will focus on the results for the blocking probability. In the first set of tests, we have fixed the number of wavelengths used for each paradigm, and have varied the number of transceivers in each node assigned to the semi-static paradigm (Tx static ). First of all, Figure shows the blocking probability obtained for the OCSPN and for the parallel networks, when three wavelengths have been reserved for the semi-static paradigm and the other five wavelengths have been assigned to the dynamic one. 0. Blocking Probability 0.0 = 0, Tx dynamic = 8, Tx dynamic E-3, Tx dynamic = 8, Tx dynamic = 0 = 0, Tx dynamic E-4 = 8, Tx dynamic, Tx dynamic = 8, Tx dynamic = 0 E-5 0.0 0.2 0.4 0.6 0.8.0 Dynamic Traffic Load Figure : Blocking probability comparison when three wavelengths are reserved for the semi-static paradigm.

As shown in Figure, the OCSPN gets lower blocking probability than the parallel networks in all cases. The reason is that the polymorphic network also uses additional wavelengths and transceivers, initially assigned to the semi-static paradigm, but that finally were not used for establishing the virtual topology. In contrast, the parallel network devoted to dynamic circuit establishment could not use those resources, thereby leading to higher blocking probabilities than the OCSPN. Due to resource sharing in the polymorphic network, it is clear that an OCSPN will always get better or at least equal results in terms of blocking probabilities than the parallel networks. The question is under what circumstances is the gain in terms of blocking probability significant. In Figure 2, we show the results obtained when five wavelengths are used for the semi-static paradigm and three for the dynamic one. Again, the OCSPN obtains lower blocking probabilities than the parallel networks. However, in this case, when eight transmitters are assigned to the semi-static paradigm, the difference between the polymorphic and the parallel networks is less significant than in other scenarios. Blocking Probability 0. 0.0 = 0, Tx dynamic = 8, Tx dynamic, Tx dynamic = 8, Tx dynamic = 0 E-3 = 0, Tx dynamic = 8, Tx dynamic, Tx dynamic = 8, Tx dynamic = 0 E-4 0.0 0.2 0.4 0.6 0.8.0 Dynamic Traffic Load Figure 2: Blocking probability comparison when five wavelengths are reserved for the semi-static paradigm. The explanation for this behaviour is as follows. If we assign three transceivers to the semi-static paradigm, the virtual topology established employs nearly all the transceivers reserved (40 out of 42) but only requires at most three wavelengths in any link. This means that there is a significant amount of idle wavelength channels (since five wavelengths were reserved in all the links). When the parallel networks are used, the dynamic paradigm cannot take advantage of these idle resources, but when the OCSPN is used, the dynamic paradigm has 74% additional wavelength channels (for a dynamic traffic load of 0.), leading to an improvement of more than four orders of magnitude in blocking probability when compared to the parallel networks. However, if we reserve eight transceivers for the semi-static paradigm, the virtual topology employs 9 out of the 2 available transceivers, and uses the five reserved wavelengths in many of the network links. Hence, in this case, the number of idle resources is lower than in the previous case, so that the dynamic paradigm in the OCSPN only receives 4% additional wavelength channels (when a dynamic

traffic load of 0. is considered). For this reason, the blocking probability for the OCSPN is very similar to that of the parallel networks. We have also performed a set of simulations fixing the amount of transceivers per node reserved for the semi-static paradigm, and varying the number of wavelengths used for each paradigm (W static and W dynamic ). Figure 3 shows the simulation results when eight transceivers per node are reserved for the semi-static paradigm and five for the dynamic one. Blocking Probability 0. 0.0 Parallel W static = 7, W dynamic = Parallel W static, W dynamic Parallel W static, W dynamic Parallel W static =, W dynamic = 7 E-3 OCSPN W static = 7, W dynamic = OCSPN W static, W dynamic OCSPN W static, W dynamic OCSPN W static =, W dynamic = 7 E-4 0.0 0.2 0.4 0.6 0.8.0 Dynamic Traffic Load Figure 3: Blocking probability comparison when eight transceivers per node are reserved for the semi-static paradigm. As we have seen before, the OCSPN always obtains lower blocking probabilities than the parallel networks, but like in Figure 2, there are some configurations of wavelength reservation where the difference between the polymorphic and the parallel networks is less significant. If we assign one wavelength to the semi-static paradigm, the virtual topology established employs all wavelength channels reserved in the network, but there is a significant amount of idle transceivers. When the parallel networks are used, the dynamic paradigm cannot take advantage of these idle transceivers, but when the OCSPN is used, the dynamic paradigm has 20% additional transceivers (for a dynamic traffic load of 0.), leading to an improvement of more than four orders of magnitude in blocking probability when compared with the parallel networks. While in the scenario analysed in Figure 2 the OCSPN took advantage of idle wavelengths, in this case, it takes advantage of idle transceivers initially assigned to the semi-static paradigm. However, if we reserve five wavelengths for the semi-static paradigm, the virtual topology employs nearly all the wavelength channels (90%) and most of the assigned transceivers, for establishing the logical topology. Hence, in this case, the number of idle resources is lower than in the previous case, so that the dynamic paradigm in the OCSPN only receives 30% additional transceivers (for a dynamic traffic load of 0.). For this reason, the blocking probability for the OCSPN is very similar to that of the parallel networks. These results are summarised in Figure 4. It represents, for a dynamic traffic load of 0.2, the improvement in terms of blocking probability in the OCSPN when compared

to the parallel networks. This parameter is defined as OCSPN-Parallel = log 0 (p b_parallel /p b_ocspn ), where p b_parallel is the blocking probability in the parallel networks, and p b_ocspn the blocking probability in the OCSPN. This parameter represents how many orders of magnitude the blocking probability in the OCSPN is lower than that obtained in the parallel networks. Wavelengths reserved for the semi-static paradigm 7 6 5 4 3 Scenarios with excess of wavelengths Scenarios 2 with excess of trasceivers 2 3 4 5 6 7 8 9 0 2 Trasceivers per node reserved for the semi-static paradigm Best balanced resource assignation 5.000 4.500 4.000 3.500 3.000 2.500 OCSPN - Parallel 2.000.500.000 0.5000 0 Figure 4: Improvement in terms of blocking probability between the OCSPN and the parallel networks for a dynamic traffic load of 0.2, and a uniform traffic matrix associated to the semi-static paradigm. As shown in Figure 4, if the number of wavelengths and transceivers assigned to the semi-static paradigm is well balanced, the improvement in terms of blocking probability between the OCSPN and the parallel networks is little or moderate (between zero and one order of magnitude). However, if the reservation of wavelengths and transceivers for the semi-static paradigm is not well balanced (there is an excess of one kind of resource when compared to the other one), as we have explained in previous examples, a lot of resources will be wasted if the parallel networks are used. In contrast, these resources are utilised by the dynamic paradigm in the OCSPN, leading to significant improvements in terms of blocking probability when compared to the parallel networks (several orders of magnitude). 4. Conclusion An Optical Circuit-Switched Polymorphic Network is a novel network architecture that provides service differentiation at the optical layer by supporting simultaneously three optical circuit-switching paradigms with different grades of dynamism in the same physical network. In this paper, we have made a simulation study to analyse the performance advantages obtained by using an OCSPN when only the semi-static and the dynamic paradigms are used, in comparison with the utilisation of two parallel networks. An OCSPN always obtains lower blocking probabilities (or at least equal) for the dynamic paradigm than employing parallel networks, while getting equal or very close performance in terms of congestion for the semi-static paradigm. The

improvement in terms of blocking probability is especially significant if the reservation of wavelengths and transceivers for the semi-static paradigm is not well balanced (there is an excess of one kind of resource when compared to the other one). In that scenario, a lot of resources will be wasted if the parallel networks are considered, while they are reused by the dynamic paradigm in the polymorphic network thereby leading to significant improvements in blocking probability. While a network could be initially properly designed, in a way that the number of wavelengths and transceivers assigned to the semi-static paradigm were wellbalanced, traffic changes may lead to an unbalanced situation. Hence, in parallel networks (since the semi-static network and the dynamic one are independent) a set of resources would be wasted, while the OCSPN would take advantage of these idle resources by using them for establishing dynamic connections. However, even in a balanced scenario, the OCSPN can get a lower blocking probability than the parallel networks (up to one order of magnitude). Moreover, as dynamic resource allocation is used in polymorphic networks, if the traffic pattern changes, the OCSPN can easily adapt to the new situation by reassigning the set of resources devoted to the semi-static and dynamic paradigms, as well as facilitating the process of virtual topology reconfiguration. Future research on polymorphic networks will analyse the performance advantages of the OCSPN when traffic pattern changes and the reconfiguration of the virtual topology are considered. Acknowledgements This work has been funded by the Spanish Ministry of Science and Technology (Ministerio de Ciencia y Tecnología) under Grant TIC2002-03859. References [] I. de Miguel, J.C. González, T. Koonen, R.J. Durán, P. Fernández, I. Tafur Monroy, Polymorphic Architectures for Optical Networks and their Seamless Evolution towards Next Generation Networks, Photonic Network Communications, vol. 8, no. 2, pp 77-89, Sept. 2004. [2] I. de Miguel, F. González, D. Bisbal, J. Blas, J.C. González, J.C. Aguado, P. Fernández, J. Durán, R.J. Durán, R.M. Lorenzo, E.J. Abril, I. Tafur Monroy, Nature-inspired routing and wavelength assignment algorithms for optical circuit-switched polymorphic networks, Fiber and Integrated Optics, vol. 23, no. 2-3, pp 57-70, March-June 2004. [3] C. Qiao, Polymorphic control for cost-effective design of optical networks, European Transactions on Telecommunications, vol., no., pp. 04-4, Jan./Feb. 2000. [4] M. Düser, P. Bayvel, Analysis of a dynamically wavelength-routed optical burst switched network architecture, Journal of Lightwave Technology, vol. 20, no. 4, pp. 574-585, April 2002. [5] S. Baroni, P. Bayvel, Wavelength requirements in arbitrarily connected wavelength-routed optical networks, Journal of Lightwave Technology, vol. 5, no. 2, pp.242-25, Feb. 997. [6] OMNeT++ discrete event simulator system, http://www.omnetpp.org [7] Project Akaroa, http://www.cosc.canterbury.ac.nz/research/rg/net_sim/simulation_group/akaroa/ [8] R. Ramaswami, K.N. Sivarajan, Design of logical topologies for wavelength-routed optical networks, IEEE Journal on Selected Areas in Communications, vol. 4, no. 5 pp. 840-85, June 996. [9] A. Mokhtar, M. Azizoglu, Adaptative wavelength routing in all-optical networks, IEEE/ACM Transactions on Networking, vol. 6, no. 2, pp. 97-206, April 998. [0] R. Johnsonbaugh, M. Schaefer, Algorithms, Pearson Education, Inc., pp. 75-77, 2004.