Optimal reconfiguration of provisioning oriented optical networks

Optimal reconfiguration of provisioningoriented optical networksTamás KÁRÁSZ, Zsolt PÁNDIBudapest University of Technology and Economics, Dept. of TelecommunicationsMagyar tudósok körútja 2.Budapest, H-1117, HungaryE-mail: {karasz,pandi}@hit.bme.huAbstract:Today and in the foreseeable future optical networks are and will be used to implement the transport layer in telecommunications networks. If in a distributed environment the arrivals of permanent (long lasting) optical channel demands are distributed and independent in time and space, the service must also be demand-driven and real-time. The applicable online provisioning algorithms, despite their efficiency, lead to a suboptimal network configuration, where the network performance strongly depends on the random arrival sequence of demands.To improve the capacity efficiency of the evolved network configuration a consolidation step must be applied after a certain number of demands arrived. The consolidation is a limited network reconfiguration to improve the capacity efficiency of the network configuration by means of changing pre-planned or on-line defined architectural parameters. The consolidation process can be based on the results of an off-line design algorithm, which may be either an exact or a heuristic method.The present paper summarizes the measures and algorithms of reconfiguration that can be found in the literature. Then the grounds and applicability of the consolidation approach is illustrated over two on-line optical channel provisioning strategies. The presented numerical results demonstrate the capacity gain achieved by means of the network consolidation step.Index terms:optical networks, on-line (OCH) provisioning, capacity efficiency, network consolidation.1Introduction and motivationsIn next generation networks automated switching flexibility and signaling intelligence are becoming fundamental features, which enable fast provisioning capabilities in the optical transport layer.The change in the networking paradigm from off-line design and configuration to on-line provisioning may be considered as an advancement to cope with problems originated from the modelling and forecasting difficulties of complex IP traffic patterns and to support the “pay as you grow” preferences of IP based service and content providers.Although in the recent years numerous research and development projects like IST Next Generation Optical Network for Broadband in Europe (NOBEL) and Multi-Partner European Test Beds for Research Networking (MUPBED), standardization activities e.g. Optical Internetworking Forum: UNI specification [1], and a large number of publications have been devoted to fast provisioning enabled optical networks, less attention has been paid to the network state evolution during the provisioning processes.The present paper gives an initial overview on the efficiency issues of network configurations evolved in on-line provisioning processes. The main scope of this paper is to identify the inherent lack of capacity efficiency of provisioning processes and to propose a consolidation based lifecycle for network design and development of provisioning oriented optical networks in order to improve the overall performance.Section 2 describes a potential approach to design provisioning oriented optical networks, Section 3 sums up the reconfiguration measures and algorithms that can be found in the literature. Then Section 4 briefly introduces the SWAP and the threshold based provisioning strategies anddiscusses how the efficiency of these techniques can be improved by applying optimal reconfiguration. Finally, the presented work is summarized and concluded.2 A potential approach to design provisioning oriented optical networks Traditionally, transport network design and dimensioning was based on traffic forecasts and networks were designed and pre-configured to meet these forecasts on the given time horizon. In practice, the experience was that the traffic – mainly that of voice services – can be modelled and forecasted with acceptable confidence.By now the hardly predictable and permanently increasing data traffic of IP services and broadband applications has become the largest traffic component in transport networks. This traffic structure transformation together with major market changes and the layered structure of the service market players (content providers, network service providers and transport service providers) made modelling and forecasting of services and traffic growth significantly more difficult. This evolution disabled the traditional forecast driven off-line design and pre-configuration of network resources. The implementation of the next generation network concept, e.g. the realization of a unified all-IP service platform will strengthen these trends in the foreseeable future.To follow unexpected changes in the traffic pattern under demands of uncertain capacity and to prevent network bottlenecks and blocking of transport capacity requests intelligent configuration flexibility or inefficient capacity over-dimensioning must be applied in the network. The availability of fast provisioning capabilities promotes the former, more efficient solution.Fast provisioning of permanent optical channels, that is, soft permanent optical channel services [9] can be interpreted, as follows. The clients generate optical channel requests spread in time and space. Based on the distributed signaling and switching intelligence of the optical network nodes the routing and wavelength allocation (RWA) problem is solved on-line, and the appropriate network elements are configured to accommodate the given optical channel request [8]. Once an optical channel is set up to accommodate a request it remains unchanged, assuming a simple incremental traffic model.There are numerous novel approaches in the literature to solve the on-line RWA problem under different resilience constraints, e.g. the method proposed in [1] uses proportionally weighted path selection instead of blind path alternation, [2] applies protection domains to improve the efficiency of failure recovery, [3] introduces shared wavelength pools for multiple class optical channels, and [4] describes an availability threshold based on-line RWA algorithm.The lifecycle of a provisioning oriented network can be divided into two repetitive basic phases: •The provisioning phase: arriving requests are served sequentially. During this process some network resources may become saturated and capacity extension may be necessary for continued provisioning.•Network capacity extension phase: additional resources are designed and installed to remove network bottlenecks.This basic cycle leaves the configuration and setup of network resources unchanged, thus, the applied provisioning process determines the capacity efficiency of the network.Because in case of on-line provisioning the demands are not known in advance and the service of the arriving optical channel requests happens in succession, the decisions are inherently sub-optimal. In other words, the accommodation of the current request is optimal according to an objective function of the given network state, but not for the whole provisioning process. Assuming a set of successive requests, the measure of this kind of sub-optimality can be evaluated by comparing the network state to that yielded by traditional off-line design methods, which may then be considered as a simple theoretical upper bound on efficiency (see Figure 1).Optical channel requestsOptical channel demands Consolidation:Network EfficiencyLowHigh arriving spread in time andspace are served one byone by a distributed andflexible network intelligenceapplying on-line provisioningalgorithms.assumed to be known in advance (based on a proper forecast) and an optimal network configuration is designed to meet the demands.Practical CaseTheoretical Lower Bound rearrangement of already arrived and served requestsFigure 1: Inefficiency of on-line provisioningThe channel setup of the output of off-line design is optimal with respect to the total set of demands in the network and may be realized by rearranging the optical channels. The objective of this optimization can be to minimize the blocking rate or to maximize the resource utilization. This problem is similar to the traditional short term network planning approach [5], and the planning problem to be solved is a special extension of the multi-commodity flow problem [6].However, the gain in capacity efficiency that may be realized by the consolidation is limited bythe extent of practically feasible rearrangements. The consolidation phase should be repetitive since the requests arriving after a consolidation phase are served according to sub-optimal decisions again.This extension of network phases results in a three-phase lifecycle (Figure 2) including:• Provisioning,• Consolidation,• Extension of network capacities.2.1 ProvisioningThe aim of the provisioning process is to set up optical lightpaths performing on-line decisions and configuration actions. A logical network topology is given with the states (idle, occupied) of network resources and the arriving optical channel requests should be served by setting up appropriate lightpaths.The decisions cover both path selection and wavelength assignment, and different strategies can be based on preselected fixed or alternate routes and on different details of the current network state. Separate and joint solutions of the RWA problem may be applied. The decisions are made by the distributed intelligence of the network nodes, request by request. Different resilience options specified for the optical channel requests to be served can be taken into account during the process. Gradual saturation of network resources may trigger the network management to warn for the need of additional resources. Additional resources may be freed due to a consolidation process when limited rearrangement and reconfiguration of the network is performed to improve capacityefficiency. If the consolidation does not provide enough idle resources the network capacity should be extended by installing new network elements.2.2ConsolidationNetwork consolidation is a limited reconfiguration of the network. The aim of the process is to improve the capacity efficiency achieved during the on-line provisioning process by refining online defined or pre-defined architectural parameters.The basic idea is that configuration decisions based on the knowledge of a certain group of optical channel requests (already in the network) is definitely more efficient than those resulted from a given realization of the provisioning process (a given demand sequence). More optimal resource allocation may be obtained by performing ‘traditional’ network design including all demands currently being served. A sequence of reconfiguration actions is then needed to set up the obtained optimal network state.The feasible rearrangements are limited by practical considerations, e.g. optical channels carry live traffic and the rearrangement itself has certain costs, as well. The type and amount of changes may also restrict the extent of the rearrangements, as well. Different basic rearrangement strategies can be defined enabling different types of rearrangements and different target functions. Obviously, the originally specified service requirements (e.g. resilience options, physical characteristics of the optical channel, etc.) should be met in the new configuration. For example, having a network configuration evolved during a provisioning process with predefined routes, the routing can be left unchanged, and only modifications in the wavelength assignment may be necessary. However, it is also possible to change both the set of routes and the wavelength assignment in place. As a consequence, the trade-off between the increase in capacity efficiency and the extent of rearrangements should be considered in the consolidation process. The amount of changes can be limited explicitly, or by the target functions specified to optimize the rearrangements.2.3Extension of the networkThe aim of network extension is to decide where and how to install new network resources in order to eliminate bottlenecks in the network. Since the appearance of bottlenecks is inherent in the provisioning process and depends on the behaviour of the provisioning process, the optimal extension of the network resources can not be separated from the provisioning and consolidation processes.Consolidation can be performed prior to the extension or the consolidation and the extension can be planned and executed jointly.Prior to or in combination with the capacity extensions network consolidation can be performed to improve capacity efficiency. Both the combined and the sequential approach to the consolidation and extension are well known from traditional network design as the optimal capacity extension and the optimal network rearrangement problems [7, 8].2.4Formalization of the three-phase lifecycle:For better understanding the formalization of the tree-phase lifecycle is given in what follows. NOTATIONd - arrival demands that are unknown in advance; the arrivals are distributed andindependent in time and space{d} - sequence of arrival demands, i.e., n demands one after the other, where the order of demands is important[d] - set of arrival demands, where the order of demands is not importantst curr - current network state, the state of network elements (idle/occupied)st opt - optimal network state, the state of network elements (idle/occupied)st subopt - suboptimal network state, the state of network elements (idle/occupied)PROVISIONING PHASEInput: - {d}- st currProcess: n provisioning stepsInput:- st curr- next d in the sequenceProcess:- accommodate the current demand according to an objectivefunction; the next state of network will be only sub-optimalst curr* = st subopt after serving a demandOutput: -Output: -st curr* after serving the sequence of demandsCONSOLIDATION PHASEInput: - [d]- st currProcess: - accommodate the set of demands in one step according to the objective functionst opt after serving the set of demands, and by means of reconfiguration st curr* = st opt Output: -NETWORK EXTENSION PHASEInput: - st currProcess: - extend the network capacities to be able to serve the blocked demands and the forecasted traffic as wellst curr*Output: -3Measures and algorithmsThe reconfiguration problem of virtual network topology has been studied widely in research works. In this section we summarize these measures and algorithms and classify them based on different aspects.3.1Measures:The works of recent years apply many kinds of measures according to the specification level of network models and the employed network technologies. Under the reconfiguration the measures can be divided into two basic classes. The first class contains measures that consider the optimal reconfiguration of occupied network resources, e.g. minimize average hop-numbers [11], the number of occupied wavelengths [10], the number of occupied physical links [10], the average propagation delay over a ligthpath [12] or the maximum link load [12]. At the same time the second class of measures describe the number of changes necessary to reach the optimal network state from the current one. Various types of changes may be considered according to the network model, e.g. the number of wavelength changes [10], the number of wavelength path changes [11] or in [13], where the network nodes are modelled with transmitters and receivers, the mean number of disrupted transceivers or the maximum instantaneous number of disrupted transceivers.These measures can be applied individually or combined, but too many measures make the objective function unnecessarily complicated and thus probably increase the solution time significantly. In addition to this, the measures must be normalized and weighted to facilitate meaningful comparison.3.2Algorithms:To solve the virtual topology reconfiguration problem a lot of algorithms and heuristics are applied in the literature. In what follows the algorithms are classified according to different aspects. With respect to accuracy, two groups may be distinguished.•Methods giving exact results: In this case the searching of the optimal solution is a multi-commodity flow problem. The problem is formalized in a mathematical representation and then it is solved as a linear programming (LP) or a non-linear programming (NLP) problem. In the literature several publications can be found with LP or NLP formalizations, where the objective functions contain the measures listed in the previous section [10-12], [15], [17]. However, solving NLP problems is NP-hard, and, in order to avoid this several heuristic methods exist,e.g. LP relaxation [16]. A comprehensive summary of heuristics can be found in [17]. •Methods giving approximate results: The poor scalability of exact solutions to real-size problems justifies the existence of these methods. The applied algorithms include metaheuristics, such as genetic algorithms [14], [12], Lagrange decomposition [21] and simulated annealing. Simulated allocation also gives good results to the capacity allocation problem [6].Another classification criterion is the relation between the current and the optimal network state in terms of the extent of the network configuration that may be changed during the optimization. According to this there are three classes of algorithms [19].•Direct approaches: In this case the new network configuration is independent from the current one. The target is to find the globally optimal solution; however, it is a hard problem, as discussed earlier, and time and space complexity of the solution increases rapidly with the size of problem. Another issue of direct approaches is the transition between the current and the new network states. One possible approach is to totally cancel the current state and configure the new one, but it may be inefficient from several aspects. Another approach is an incremental switching where the starting point is the current network state and the target is the new configuration (Incremental Reconfiguration Migration). Several algorithms have been proposed to minimize the number of disrupted demands during the configuration changes [18].•Partial reconfiguration approaches: Instead of searching for an optimal solution over the whole network, the optimization may also target at a subset of the network. Various solution methods have been proposed for this problem as well, e.g. the Most and Least Loaded Channel Balance algorithm, or the Merge-Split Reconfiguration algorithm for ring topologies [19]. These algorithms reach solutions only 10-15% worse than those computed by exact methods.•Local search approaches: These methods always choose the most optimal configuration from the set of neighbouring network states. Firstly, the steps that define allowed transitions in between neighbouring network configurations have to be defined. These solutions are usually applied for reconfiguration of networks of ring topology [19-20].4Illustrative consolidation examples for provisioning algorithmsIn what follows, the sub-optimality of the fast provisioning process and the grounds and applicability of consolidation are illustrated using two examples.The first example is the SWAP (Separate Wavelength Pool) technique. SWAP is a strategy for multiple-class optical channel, i.e., lightpath provisioning. The novel idea of this solution is to logically separate the available wavelengths into pools and optical channel demands belonging todifferent service classes are served using different pools. Thus, by applying SWAP different blocking probabilities can be guaranteed for different classes of lightpath requests [3].In this study two optical demand classes are considered. The connections of the high class require resilient optical channels, i.e. these channels should survive any single link failure. The utilized resilience scheme is shared (backup) path protection (S(B)PP), where each lightpath in this class is assigned two lightpaths: a working lightpath and a Shared Risk Link Group (SRLG) disjoint backup lightpath [9]. High class lightpaths are used to carry circuit-oriented services (voice, leased lines, virtual private networks) that request lower blocking probabilities and are long lasting. The demands of the low class, on the other hand, are assumed not to require any fault tolerance. The blocking probability is also allowed to be higher than that of high class connections. Moreover, in case of a failure resources may be taken away from low class connections to recover interrupted channels.The definition of the pool sizes may only be based on a rough estimation of the channel requests. An inefficient separation of pools may lead to penalty in blocking probability, thus the potential gains of the SWAP method significantly depend on the accuracy of the estimation of the pool sizes. If the primary pool is too small, then the primary blocking (a high class connection request is rejected because a working lightpath cannot be found) will be higher. If the secondary pool is too small, then the secondary blocking (a high class connection request is rejected because a backup lightpath cannot be found, even though a working one is available) will be higher. This observationis illustrated in Figure 3, where the networkis a 4x4 mesh-torus with 40 wavelengths on each link and the offered load consists of 50-50% high class and low class uniform traffic.The figure depicts the sum of primary and secondary blocking. 2P[w1-w2] denotes thepool size ratios, where w1 and w2 are the number of wavelengths of the primary andsecondary pool. In the example the optimal separation is2P[32-8], where the blocking probability wasthe best. A decreasing trend of blocking isvisible as the size of the primary pool increases, however, as it crosses 32 wavelengths the trend changes. The reason of this phenomenon is the secondary blockingthat becomes more significant as the size of the secondary pool decreases.Figure 3 depicts simulation results without the consolidation phase. After the consolidation, which was carried out by an LP based offline algorithm where only wavelength and route assignment changes were allowed, blocking could be eliminated from the 2P[32-8] pool size setup. With other pool size setups the gain in blocking after consolidation was 2-13%. The fluctuation of results is due to the different arrival sequences of demands in the simulation cases.Nevertheless, blocking probability may improve only up to a certain degree that depends on the goodness of the separation of wavelengths into pools. In order to further decrease blocking probability the dividing line between the pools can be shifted in addition to reconfiguration of paths. If the primary blocking is high, then the pool size ratio must be changed to increase the size of primary pool. If the secondary blocking is high, then the size of secondary pool must be increased. However, we can only shift the dividing line if there are free wavelengths in the pools and it is more likely to happen right after the consolidation phase. Moreover, an additional objective of consolidation can be, besides decreasing the blocking rate, to free wavelengths for shifting the dividing line towards the optimal separation. Thus, the consolidation, in addition to reconfiguration of the paths, may be extended to a higher architectural level of networks.The second example is an unavailability threshold based provisioning strategy. In order to establish connections with end-to-end reliability guarantees in the presence of multiple simultaneous failures while utilizing network resources in an efficient way a provisioning strategy was proposed in [4]. The complexity of availability computations is eliminated by introducing the concept of sharing unavailability, defined as the probability that a shared backup resource is activated and thus becomes unavailable to a demand. The extent of allowed sharing is determined by introducing a threshold of sharing unavailability. Shared backup path protection is combined with this threshold in [4] to propose an on-line provisioning strategy.To examine the effect of the differentorder of demand arrivals an ILP designprocess was run to determine linkcapacities necessary to accommodate allof a set of demands. Then, the capacitiesobtained with the ILP design processhave been assigned to the network linksand each potential demand arrivalsequence was then fed to the simulator ofthe on-line provisioning process. Due tothe high number of permutations a smallexample was used for this purpose (6nodes, 14 links, 8 demands).The histogram of blocking rates achieved during the provisioning process executed with the different demand arrival sequences is depicted on Figure 4. The figure gives an illustration of the sub-optimality of the provisioningprocess with respect to the possibleoptimal capacity obtained using theknowledge of all the demands to beserved.The results of another experiment aredepicted on Figure 5. A European opticalnetwork topology (19 nodes, 78 directedlinks, 4 wavelength/link) was used foraccommodating a certain random demandsequence. After 10, 20 and 30 demandarrivals a snapshot was taken of thenetwork state and resource consumptionyielded by the on-line provisioningalgorithm was compared against the output of the ILP design process run with the data of the respective snapshot. Resource usage is expressed in terms of wavelength times link, i.e. the unit is one wavelength on one link.Figure 5 demonstrates clearly that the deviation of the provisioning process from the optimal solution increases over time, and so does the potential resource utilization gain due to consolidation. On the other hand, by merely intuitive reasoning one may safely assume that the costs of reconfiguration increase with the number of admitted demands. The observed change of sub-optimality altogether with the growing costs of reactive measures suggests that an optimal point may be sought for, when such measures have to be taken. This could be subject of future research.5Summary and conclusionsThe present paper illustrates a potential approach to manage provisioning oriented optical networks. A consolidation stage is introduced as part of a three-phase lifecycle in order to improve the inherently limited capacity efficiency of on-line provisioning strategies and to facilitate moreeffective network extension. A summary is given of the measures and algorithms published for reconfiguration and the potential gains from consolidation are also illustrated over two provisioning strategies. Moreover, the concept of a possible extension of the consolidation to higher level architectural parameters is also introduced. As the deviation of the provisioning process from the optimal configuration increases over time as well as the costs of reconfiguration, the problem may be treated as an optimization task.AcknowledgementsThe research work presented in this paper has been initiated in the framework of Italian-Hungarian Inter-Governmental Project I-17/04 titled “Planning and Implementing Reliable IP over WDM Optical Networks”, and partially granted by OTKA 048985 Project titled “Dimensioning and reliability analysis of fault-tolerant networks in Differentiated Reliability (DiR) environment”. References[1] L. Valcarenghi and A. Fumagalli, “Implementing Stochastic Preplanned Restoration with Proportional Weighted Path Choice in IP/GMPLS/WDM Networks”, Photonic Network Communications --- Special Issue on “Routing, Protection, and Restoration Strategies and Algorithms for WDM Optical Networks”, Kluwer Academic Publishers, vol. 4, no. 3/4, July/December 2002.[2] Pin-Han Ho and Hussein T. Mouftah, ”Allocation of Protection Domains in Dynamic WDM Mesh Networks”, Proceedings of the 10th IEEE International Conference on Network Protocols (ICNP’02)[3] Nicola Andriolli, Tivadar Jakab, Luca Valcarenghi, and Piero Castoldi, “Separate wavelength pools for multiple-class optical channel provisioning”, Proceedings of 11th International Telecommunications Strategy and Planning Symposium - NETWORKS 2004, Vienna, Austria, June 2004, pp. 379-384.[4] Zs. Pándi, M. Tacca and A. Fumagalli, “A Threshold Based On-line RWA Algorithm with Reliability Guarantees”, ONDM 2005 Conference, February 2005, Milan, Italy[5] EURESCOM P709 Project “Planning of full optical networks Project, Deliverable 3: Optical Network Planning”, see URL: http://www.eurescom.de/public/projectresults/P700-series/709d3.htm[6] M. Pióro, D Mehdi, “Routing, Flow and Capacity Design in Communication and Computer Networks”, ELSEVIER, 2004, ISBN: 0-12-557189-5[7] S. Sengupta, R. Ramamurthy, “From Network Design to Dynamic Provisioning and Restoration in Optical Cross-Connect Mesh Networks: An architectural and Algorithmic View”, IEEE Network, July/August 2001, pp. 46/54.[8] EURESCOM P1012 FASHION Project- Flexible, Automatically SwitcHed, client Independent Optical Networks, see URL: www.eurescom.de/public/projects/p1000-series/P1012/default.asp[9] S. Ramamurthy, L. Sahasrabuddhe, and B. Mukherjee, “Survivable WDM mesh networks” , J.Lightwave Technology, vol. 21, no. 4, pp. 870-883, Apr. 2003.[10] B. Ramamurthy and A. Ramakrishnan, “Virtual topology reconfiguration of wavelength routed optical WDM networks“, Proceedings of IEEE GLOBECOM, 2000, pp. 1269-1275.[11] Dhritiman Banerjee , Biswanath Mukherjee, "Wavelength-routed optical networks: linear formulation, resource budgeting tradeoffs, and a reconfiguration study", IEEE/ACM Transactions on Networking (TON), v.8 n.5, p.598-607, Oct. 2000.[12] Bin Zhou, Jun Zheng, Hussein T. Mouftah, "Dynamic reconfiguration based on balanced alternate routing algorithm (BARA) for all-optical wavelength-routed WDM networks", GLOBECOM 2002 - IEEE Global Telecommunications Conference, vol. 21, no. 1, November 2002 pp. 2713-2717。