Design of Static Resilient WDM Mesh Networks with Multiple Heuristic Criteria

Andrea Dacomo, Simone De Patre, Guido Maier, Achille Pattavina, Mario Martinelli

AbstractWe present a heuristic method to optimize the capacity of a WDM mesh network assigning working and protection capacity to a set of static optical-connection requests under the linkdisjoint constraint. The method can be applied to networks supporting either dedicated or shared path-protection. Routing and ber and wavelength assignment (RFWA) are jointly performed for all the lightpaths with the aim of minimizing the number of bers in the network. The employment of multiple prioritized heuristic criteria and an efcient link disjoint paths search algorithm are proposed. With less computational effort, the proposed heuristic approach allows to obtain good suboptimal results compared to the exact integer-linear-programming optimization. After introducing our design approach, we discuss the optimization of particular case-study networks under various condition and compare the results. Index Terms Wavelength Division Multiplexing, optical ber communication, planning, protection

I. I NTRODUCTION AVELENGTH Division Multiplexing (WDM) is today the established standard transmission technique for large bandwidth telecommunication trafc. WDM networks are being deployed at an extremely rapid rate, for wide-area transport applications as well as in the metro and regional areas. WDM network design is a primary problem which new and old operators must continuously solve, planning new installations or expanding existing infrastructures [1]. Optical network research has been investigating for the last years design and optimization techniques in order to provide operators efcient and exible procedures. In the past, WDM networks used to be based on the ring topology: simple switching functions in the optical layer were performed by Optical Add-Drop Multiplexers (OADM). Optical technology evolution recently made new switching devices, i.e. Optical Cross Connects (OXCs), available on the market, thus making complex mesh WDM networks feasible. The increase in WDM network complexity brought the need for suitable control and management strategies into foreground. Problems such as routing and resource allocation for optical connections can no longer be manually solved. Automatic planning tools and procedures are needed for the future which can achieve an efcient utilization of network resources in a reasonable computational time. In this paper we present a novel heuristic method to perform full planning of a mesh
A. Dacomo, S. De Patre, G. Maier, M. Martinelli are with CoreCom, Via Ampere, 30 - 20131 Milan, Italy. E-mail: maier, depatre, martinelli @corecom.it. A. Pattavina is with Dept. of Electronics and Information, Politecnico di Milano, P.za Leonardo da Vinci, 32 - 20133 Milan, Italy. E-mail: [email protected] M. Martinelli is also with Dept. of Electronics and Information, Politecnico di Milano, P.za Leonardo da Vinci, 32 - 20133 Milan, Italy.

WDM network: given the physical topology and the set of Optical Connections (OCs) that must be setup (OC-layer topology), network capacity dimensioning and resource allocation are solved simultaneously, minimizing a chosen cost function. The method we are proposing is also able to plan the spare capacity in an optimal way, in order to guarantee network survivability against a link failure. The issue of survivability of optical connections has become of outstanding importance today: a loss of a high-speed optical connection operating at such bit-rates as 10 Gbit/s or higher, even for few seconds, means a huge waste of data. For this reason, various possible protection techniques acting directly in the optical layer have been proposed and standardized [2], [3] for WDM mesh networks. In this paper, we have chosen to consider path protection, in the two alternative implementations which are dedicated and shared. Shared pathprotection offers the opportunity to limit the amount of spare resources; however the recovery operation can be complex and requires OXC reconguration in the intermediate nodes. Dedicated path-protection schemes are simpler and faster and do not require reconguration of transit OXCs. In both implementations channel assignment is completely failure independent: this is clear advantage in a context of network planning under static trafc. The paper summary is the following. In section II the basic terminology for WDM networks and the protection and restoration techniques are reviewed. Section III introduces and explains our heuristic optimization approach, which is based on the use of a particular network model called multiber layered graph. Such a model allows to apply complex routing algorithms by combining different heuristic criteria. Section IV presents the comparison of our approach to other optimization solutions proposed in literature. Finally, section V is dedicated to the analysis of the results obtained by applying this approach to some case-study networks. The study-case analysis has been performed not only to evaluate the performance of our multiplecriteria heuristics, but also to investigate some properties of resilient WDM networks loaded with static trafc, such as the advantage of wavelength conversion and the relative costs of different protection techniques. II. R ESILIENT-WDM- NETWORK MODEL DEFINITION Let us describe the general network scenario to which the design method we are proposing refers. Our model of WDM network is dened by a physical and an OC-layer topology. The physical topology is composed of WDM transmission links and WDM switching nodes arranged to compose a given graph,

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

having pre-assigned topology (either ring or mesh); geographical node locations and the physical link lengths are known. A WDM link is a multiber cable (equipped with all the optical ampliers necessary to compensate transmission loss [4], plus other transmission components such as dispersion compensation devices or regenerators). In each (bidirectional) link some bers are used in one propagation direction and some others (not necessarily the same number) in the opposite direction. In the simplied model adopted in this work we will not consider any physical transmission phenomena affecting the signal quality (attenuation, noise, dispersion, non-linear effects, etc.). Each ber of the link carries a given number of WDM channels. We assume for simplicity that all the WDM channels in the network are characterized by the same bit-rate of the carried digital ow (e.g. 2.5 and 10 Gbit/s). The capacity of a physical link is the number of WDM channels it carries, that is the number of bers in the two directions by the number of wavelengths per ber. While the physical topology in terms of links is given, in our approach the capacity of each link is a variable of the dimensioning problem. We will explain again later that in particular we chose to pre-assign the same number of wavelengths in all the bers of the network, while we left as a variable the number of bers per link. The nodes of the physical topology are the OXCs, which perform all-optical switching (space and sometimes wavelength switching) on the WDM transit lightpaths, without the intervention of the electronic layers. In addition, they behave as terminal equipment for some lightpaths, performing the add and drop functions. In our model we will regard each OXC as a potential source and destination of WDM trafc. In this work we are considering a static OC trafc scenario. This means that the optical connections requested from the upper transport protocol layers are permanent and known a priori; moreover, each connection requires exactly the capacity that can be carried by a WDM channel. We will consider only unidirectional point-to-point trafc 1 : therefore each connection request is characterized by a pair of source and destination OXCs. It is understood that each node pair may request more than one connection if the total needed bandwidth exceeds the capacity of a single WDM channel. The OC-layer topology (alias logical topology or offered trafc) is the set of all the optical connections that the network must establish: it can be represented by a graph having all or a subset of the OXCs as nodes and a direct link between each pair of nodes for which a connection has been requested. The nal aim of planning is to activate all the requested optical connections or, in other words, to perform the mapping of the OC-layer over the physical topology (OC-P mapping) (Fig. 1). In doing this, two problems of capacity dimensioning and capacity allocation are simultaneously solved. In WDM networks, trafc is carried by means of circuit-switched transport entities, optically routed on the basis of their wavelength. These basic entities are the lightpaths (elsewhere also optical paths, or optical-channel sections). A lightpath is composed of a sequence of WDM channels, one per each crossed link, connecting the source OXC to the destination OXC. In the present work we will consider two types
Connections among the same pair of nodes in the two directions are treated as independent. This makes the our problem formulation more general.

of WDM networks, according to their wavelength conversion capability [5], [6], [7]:

Wavelength Path (WP): no conversion capability is provided; Virtual Wavelength Path (VWP): every node is fully equipped with converters (opaque OXC) so that any incoming optical signal can always be converted on any idle output wavelength. In the WP case, all the WDM channels composing a lightpath must be at the same wavelengths (wavelength continuity constraint). Connections must be established by suitably conguring switching resources and allocating transmission resources in order to setup the requested lightpaths according to the OC-layer topology. A fundamental choice in our approach is that a solution is found only when ALL the optical connections have been set up (rejecting or blocking a connection request is not contemplated): the capacity of the physical links is dimensioned in order to fulll this requirement under no additional constraint; of course we will seek the allocation that minimizes dimensioning. Dimensioning and allocation strictly depend on the protection strategy that must be adopted in the WDM network. We assumed for simplicity that all the connections are protected with the same strategy and, as already mentioned above, we considered only path-protection. Therefore three options must be considered: unprotected connections, dedicated and shared protection. If WDM survivability is not supported by the network, an OC is activated by setting up one single lightpath (called working lightpath) from the source to the destination node. If the connection is path-protected in the WDM layer, the network must be able to provide spare resources that in case of failure guarantee the recovery of the failed working path. Therefore for each connection two lightpaths, a working and a protection pair (w/p pair) must be set up. In the dedicated case each protection lightpath is exclusively used to backup one single working lightpath. For path-protection to be effective, the two lightpaths must be setup in physical route diversity. If the main purpose of the survivability technique is link failure recovery, then it is sufcient that the two lightpaths are routed in such a way that they do not share any physical link (link-disjoint paths) [8]. If also node failures have to be recovered, then the linkdisjoint condition is not sufcient any more and also sharing of physical nodes has to be prevented (node-disjoint paths). In the present paper we assumed recovery of link failures as our objective and therefore we considered link-disjoint routing for the w/p pairs, leaving node-disjoint path-protection for future developments. Shared path-protection requires the setup of a linkdisjoint w/p pair for each requested connection, exactly as in the dedicated case. The difference is that a single WDM channel may be shared by more protection lightpaths. Sharing is possible only between protection lightpaths that are associated to working lightpaths which are mutually link-disjoint: otherwise a single failure simultaneously affecting the two working paths would cause a shared WDM channel of the protection paths to be overloaded with trafc, making failure recovery impossible [9], [10].

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

Physical Topology

OC-Layer Topology

Network Design

OXC

Connection Request

Create the idle network with unlimited fiber amount Sort connection requests Set-up w/p lightpath pairs sequentially

multifiber link

Mapping Design

Working LP

Spare LP

i j x y

j = y
Optimization cycle
D C

Fig. 1. Mapping of the OC-layer over the physical topology in a WP-WDM network supporting link-disjoint dedicated path-protection.

Fig. 2. Flow-chart describing the operation of the design and optimization tool for WDM networks with path-protection.

III. H EURISTIC OPTIMIZATION METHOD Static optimization of a WDM network (with or without protection) can be so summarized: given a static trafc matrix describing the OC-layer topology, nd the optimum values of a set of network variables that minimizes a given cost function, under a set of constraints. Our approach to WDM static optimization we are now going to describe belongs to the deterministic heuristics. In this class of methods, a rst OC-P mapping is obtained by adopting a set of heuristic criteria. In this phase the physical topology is dimensioned so that all the lightpaths (or w/p pairs in case of path-protection) can be set up and all the OC requests are satised. Then, the network is globally optimized by trying to improve the rst mapping. As we have already pointed out, we chose to consider a multiber network in which all the bers in the network host the same number of wavelengths , chosen a priori. The variable capacity of the physical topology is represented by the number of bers in each link. Therefore our cost function to be minimized is the total number of bers in the network. Most of the cost of optical networks comes today from transceivers, transponders and switching equipment: the number of ber is therefore more a theoretical than a realistic measure. Nevertheless, we chose this particular function for the present work because we intended to test our design approach in simple conditions: other more realistic options will be tested in future developments. The scheme of the network planning procedure is represented in Fig.2. It is composed of two main sections: the initial greedy OC-P mapping and the subsequent optimization cycle. In the following the two sections will be described in details in the cases of unprotected network and of dedicated path-protection. For shared path-protection the procedures are slightly different and they will be described separately afterwards. In the whole design procedure a particular representation of the network has been adopted which is the multiber layered graph (MLG). This is a working auxiliary representation of the network state derived from the layered graph (elsewhere called wavelength graph). This representation, introduced by Refs. [11], [12], has been often used for dynamic trafc analysis [11], [10] or for static optimization in monober networks

[12]. In Ref. [13] we have extended the use of the layered graph to multiber static networks, leading to the multiber layered graph. The construction of this graph is obtained in the following way. Let us assume that all the links in the network have the same number of ber . The graph representing the physical topology is replicated in parallel identical planes, the rst representing ber 1 in all the links, those from to representing ber 2, and so on. Each OXC is replicated in images; horizontal links in the graph corresponds to WDM channels, vertical links connecting OXC images at the same wavelength represent space switching and vertical links connecting OXC images at different wavelengths represent wavelength conversion (in VWP networks). Each OXC is further replicate in one additional image to represent lightpath adding and dropping. The network state is represented by marking as disabled an horizontal link of the MLG whenever the corresponding WDM channel is busy because assigned to some connection. The use of the MLG and its advantages will be described in the next sections. The MLG dened as above is a very general model since it allows to obtain the most complete set of information regarding each lightpath, i.e. the exact wavelength and the exact ber reserved to it on each link. All these data are required for a centralized network control system to actually setup the lightpath. If the purpose of design is only network dimensioning and not lightpath control, there are some cases in which a simplied LG can be adopted. If all the bers in each link are identical or if the network is WP or if both conditions hold, an LG can be used having , or 1 plane, respectively; each plane will have links with capacity , or , respectively. In some real situations the use of the MLG can become necessary, e.g. when an operator has to lease a specic wavelength channel from another operator or when a link is equipped with bers with different physical characteristics. However, the advantages of reducing the number of planes of the LG are modest in terms of computational complexity (in turn, links with capacity greater than one must be managed by the RFWA algorithms).

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

Greedy OC-P mapping

A. Initial greedy OC-P mapping The initial greedy OC-P mapping is obtained starting from the idle physical topology. All the connection requests of the OC-layer topology are set up in sequence one after the other until all have been satised. To do this the initial idle physical topology is oversized: each link initially contains a number of bers so great to be considered innite. In this way we are guaranteed that the greedy OC-P mapping leads to a feasible solution. When path-protection is adopted a further necessary condition to guarantee the solution is that the physical topology is at least bi-connected, i.e. each OXC terms minimum two adjacent links. The initial oversized idle MLG is built as described above by choosing in a way that the subsequent mapping can not be constrained by the lack of capacity. Clearly, must not be too large to limit unnecessary memory occupation. An appropriate value of can be estimated by inspection, also with the help of an easy-to-compute lower bound (introduced further on). The initial OC-P mapping requires that all the requested connections of the OC-layer topology are sorted. There is no way to decide a priori how to sort connections in an optimal way. Therefore we proceeded in a heuristic sense by trying different sorting rules and deducing the best rule from the results. In particular, we took into account the following rules: longest rst: priority is given to connections whose source and destination nodes are farthest apart; busiest pair rst: the node pair requesting the maximum number of lightpaths that have yet to be established is selected; the pair is served by establishing one of its lightpaths; balanced: the previous rules are combined: the node pair to be served is selected according to the product of the distance and the number of lightpath requests and not yet served; random: included as a benchmark for comparison. After the initial sorting, the connection requests are processed. The rst request in the sorted list is satised by setting up a single lightpath or a w/p pair in a greedy way, i.e. by exploiting the available capacity at the best according to the protection strategy. Then the second is satised with the same procedure: the resources available for the second setup comprise the whole network minus those WDM channels already allocated to the rst connection. Then the next connection request is done and so on until all the connections are activated. Setting up a (working or spare) lightpath implies rst of all the accomplishment of the routing function, i.e. the selection of the sequence of physical links (path) that a lightpath must cross, as in traditional circuit switched electronic networks. In the case of WDM, routing is coupled with wavelength assignment, which consists in selecting the WDM channel that must be allocated to a lightpath in each link [11], [14]. The combination of these two functions is called Routing and Wavelength Assignment (RWA), and can be interpreted as the mapping of an arc of the OC-layer topology over the physical topology [15], as displayed in Fig. 1. In the case of multiber links RWA becomes Routing, Fiber and Wavelength Assignment (RFWA), as also a particular ber must be selected on each link for a given lightpath.

Let us rst explain how an unprotected connection would be managed. The impact on the cost function of the network (in our case, the total number of ber) of the criteria chosen to perform RFWA of a single lightpath is not known: the cost of the network is a global parameter which depends on the entire OCP mapping, and not just on one single greedy allocation. Therefore RFWA has to be performed following heuristic criteria dictated by the good sense. Many of these criteria have been proposed and employed in literature [13], [16]. As for routing, Shortest Path Routing (SPR) and Least Loaded Routing (LLR) are the most used. The rst chooses the shortest path available according to a length metric such as the number (minimum hop - mH) or the sum of the physical lengths (minimum length - mL) of the crossed links. The second chooses the path with the highest most-idle-link. By most-idle-link we mean the link on the path with the greatest idle capacity when the lightpath is routed 2 . For wavelength assignment, the First Fit criterion (FFW) has often been used [17]: wavelengths are equally sorted in all the bers of the network and the rst available is always chosen according to this sorting 3. The First Fit criterion can also be applied to ber selection (FFF) in a multiber network: even if all the bers are identical, FFF can have the benecial effect to reduce the amount of partially used bers [14]. We decided to explore the four heuristic criteria mentioned above (SPR, LLR, FFW and FFF), adding the random assignment of wavelength and ber (RNW and RNF). Moreover, in the cases presented in this paper we adopted the mH for SPR; results obtained with the mL metric have been presented in Ref. [13], [19]. Other proposed criteria (most used routing, pack and spread wavelength assignment, and so on) can be easily implemented in our design approach, though they have not been considered in this work. As for the initial connection-sorting rule, in a heuristic approach the possibility of testing many RFWA criteria and then comparing the optimization results is fundamental in order to improve the solution. Therefore the RFWA criteria are selected among those mentioned above at the beginning of the optimization process and then applied in every lightpath setting, both in the rst mapping and in the optimization phases. A powerful advantage of the MLG is that it allows an easy implementation of each of the above criteria. A criterion is implemented by assigning a suitable set of weights to arcs and nodes of the MLG before setting up a new lightpath. Once weights are associated, the Dijkstra algorithm nds the route connecting the source to the destination on the MLG with the least total weight. In some cases a more efcient algorithm called SPAWG [11] (a modied Dijkstra) can be applied instead4 . This route is nally mapped on the physical topology, thus obtaining the optimal lightpath setup according to the chosen criterion.
Unlike SPR, LLR requires an updated knowledge of the network state. In the VWP case wavelength assignment is performed for all the WDM

channels, while in the WP case it is related to the whole lightpath[18]. The total weight is often the (linear) sum of the weights of the elements along the MLG route (e.g.for the SPR), but it not always like that. For example in the LLR case the total weight is the maximum weight along the MLG route (a non-linear operation). SPAWG can be adopted with criteria based on linear weights.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

To obtain a complete RFWA, more criteria must be applied together. We have made this possible by associating to each MLG element an array of weights, instead of a single weight. Each weight of the array is determined by a specic criterion. Then we have modied the Dijkstra algorithm so that the criteria can be applied in a prioritized sequence. Each time several alternative MLG routes have an equal total weight according to the primary criterion, they are compared according to the secondary criterion, and so on. A very interesting option offered by this approach is that more criteria concerning the same aspect of RFWA can be applied in sequence, taking advantage of the best heuristic quality of each one. For example, SPR favor the set up of short lightpaths, while LLR tends to avoid link congestion: we can expect that cascading LLR and SPR or viceversa can improve the heuristic. In section V we will show the benets of criteria prioritized combinations by comparing some case-studies results. We shall nally note that in the heuristic approach we are proposing RFWA is performed in an unconstrained mode, that is all the possible routes on the MLG connecting source to destination are scanned in setting up a new lightpath. In other similar optimization techniques, in order to simplify calculations, routing is performed in a constrained mode [20], [21], considering only a pre-determined subset of paths among each node pair. The Dijkstra algorithm applied to the MLG allows us to compute unconstrained RFWA in a reasonable time despite its higher complexity, giving the opportunity to explore more potential solutions than the constrained mode (this is not always an advantage, as we will point out in section V). When dedicated path-protection is supported, the problem of preallocating the w/p pair to a connection request on a mesh WDM network is quite complex [22]. The RFWA of the working lightpath is coupled to the RFWA of the protection lightpath by the route-diversity constraint. In order to focus on the linkdisjoint constraint and to easily explain this aspect, let us take a mesh network as an example, having one single channel per link (one wavelength and one ber) and assume the SPR is adopted as routing criterion. There are two main techniques to nd two link-disjoint paths connecting two nodes of a mesh topology. The simplest one is called two-step search. The shortest path in the network connecting the two nodes is found (e.g. applying the Dijkstra algorithm) and it is allocated to the working lightpath. Then the shortest-path algorithm is run again to route the protection lightpath, which will be assigned, in this way, the second-shortest path. This technique is greedy: there are situations in which it fails to nd a solution even if a w/p linkdisjoint pair actually exists. The network represented in Fig. 3 is the simplest topology in which such situation does occur [23]. Fig. 3a represents the rst step of the two-step search assuming mH metric (each link has weight one). Fig. 3b shows that, once the shortest path has been allocated to the working lightpath between nodes A and D, there is no chance to setup the protection path because the network is disconnected. For this particular property this has been called trap network [24]. Finally, Fig. 3b shows that a pair of link-disjoint paths actually exists between A and D, but the two-step search can not nd it. A technique able to overcome the two-steps limitations has been proposed by Bhandari [23]. It is called one-step search,

spare C A working F D A F C D

(a)
B spare C A working F D

(b)

(c)
Fig. 3. Routing of a link-disjoint path pair in a trap network: (a) working-path routing with the two-step algorithm; (b) the two-step algorithm fails to nd the protection path; (c) with the one-step algorithm the link-disjoint pair can be routed.

since the two lightpaths of the w/p pair are not routed separately, but they are jointly routed by performing a suitable algorithm (the modied Dijkstra algorithm). Besides being able to solve trap-network topologies, it can also nd better solutions than the two-step search. This is because the total length of a w/p pair composed of the rst- and the second-shortest paths between two nodes is not necessarily the minimum-length pair connecting the nodes. In our WDM design method we developed both the two-step and the one-step techniques: in the dedicated path-protection case one of the two can be chosen at will, though the one-step search is the best choice; we will explain later that only the two-step search can be used in our method for shared path-protection. In the two-step case RFWA is performed by applying the procedure described above for the unprotected connections, rst on the working and then on the protection lightpath, as they were two different unprotected lightpaths. During the second step the MLG must be suitably modied to keep into account the link-disjoint constraint while routing the protection lightpath. We dened a new one-step search procedure by adapting the Bhandari algorithm to the MLG. This extension does not conict with Bhandaris limiting assumptions on the graph 5 . In particular, a set of rules has been added to the modied Dijkstra algorithm to control the edge inversion operation [23] in the MLG environment with and without wavelength conversion. Moreover, the new rules also allow to nd the least-loaded link-disjoint pair in a single-step search, when LLR is considered as a criterion. B. Optimization cycle Let us now complete the description of the operation of the design tool. After the greedy OC-P mapping has been completed (starting from the idle network), all the empty bers are pruned from the physical topology; correspondingly, all the MLG horizontal links representing WDM channels of the pruned bers are disabled. Then the optimization iteration begins. An optimization counter is dened and initially set to 1. Any ber containing only occupied and unused

The MLG is a directed graph and does not contain looping, negative or multiple edges (see Ref. [23]).

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

wavelengths (let us name it a -ber) is detected. The lightpaths routed on a -ber are sequentially considered. In the unprotected case each of them is temporary deallocated and the -ber is temporary disabled. Each deallocated lightpath is reallocated by performing RFWA with the same criteria adopted for the greedy mapping. In this case there is no guarantee of success of the reallocation since the physical network has a constrained capacity. If all the deallocated lightpaths can be rerouted then the ber is permanently disabled; otherwise, initial lightpaths are restored and the ber is preserved. In the case of dedicated path-protection the procedure is similar. The -ber contains WDM channels assigned either to working and protection lightpaths, each one belonging to a specic w/p pair. The deallocation and the subsequent tentative reallocation (applying the RFWA criteria and link-disjoint search chosen for the optimization) in this case are performed on each identied w/p pair, instead of a single lightpath. If all the w/p pairs can be rerouted without meeting blocking conditions, the tentative reallocations are conrmed; otherwise the old lightpaths are maintained. When operations on a -ber are completed, another -bers is processed, until all the -bers have been processed. Then all the empty bers are pruned again, is incremented and another iteration begins. The whole is repeated until . The rational of this optimization description is the following. Given our cost function, the inefcient spots of the network are the partially empty bers; we attempt to solve rst easy situations, i.e. bers loaded with a small number of lightpaths (low ), and then more difcult bers loaded with many lightpaths (high ).

channels). In the dedicated case the number of lightpaths using a ber was exactly equal to the number of its busy WDM channels; in the shared case this is not true any more since a single busy WDM channel can be shared by more than one lightpath. As we have stated above, the difculty of the attempt of pruning a partially used ber depends on the number of lightpaths that must be rerouted. To follow the same optimization strategy dened above it is thus sufcient to change the denition of a ber as described: all the rest of the optimization cycle remains the same. IV. R ELATED WORKS Let us identify the approach we are presenting in the framework of the research on path-protected WDM-network optimization, by comparing it to related works. One of the main features that characterize the various proposed optimization techniques is the way by which a working lightpath is associated to its protection resources. Global and local design approaches may be distinguished. According to a global approach the primary network is designed rst, i.e. only working lightpaths are established, and then protection lightpaths are dimensioned [18], [21], [25]. In Refs. [18], [21] protection lightpaths are planned according to the failure dependent design: all the possible link failures are considered one by one and for each of them a residual network is identied able to recovery connectivity. In Ref. [25] another dimensioning technique for protection resources is proposed beside the previous one. By this technique, called coordinated design, the protection network is planned by considering all the possible failure scenarios together. A characteristic of both failure dependent and coordinated designs is that working and protection lightpaths are not necessarily link disjoint. A local approach, instead, preallocates working and protection resources individually per each requested connection, and it is based on the link-disjoint constraint. This is one of the major approaches for path protection selection and it is called failure independent design since the same protection path is used to recovery a failure that occurred to any link crossed by the working path. The exact solution of failure independent WDM optimization with static trafc, leading to the absolute minimum of the cost function, can be found by applying the Integer Linear Programming (ILP). In ILP link-disjointness is guaranteed by the constraints introduced by the mathematical formulation, as in Ref. [22] or in Refs. [26], [27]. In these papers different protection schemes are considered assuming network dimensioning problem in WP monober scenario. In particular dedicated and shared path protections are compared. The same problem is examined in Ref. [20] minimizing the transmission cost in a multiber network where wavelength conversion is available in each OXC or in Ref. [28] where the network cost is evaluated using a more complex function that contains transmission and terminal and in line equipment costs (i.e. mux, demux and amplier). The ILP method requires quite a complex mathematical formulation and is computationally intensive when applied to large networks. This is why heuristic approaches have been introduced which usually provide good sub-optimal but computationally simpler solutions. As far as failure indepen-

C. Heuristic optimization for shared path-protection The procedure described above must be slightly modied to be adapted to WDM network design when the shared pathprotection strategy is adopted. A rst change regards the w/p pair allocation. In section II we explained that WDM-channel sharing is allowed only for protection lightpaths: while performing RFWA for a new lightpath, knowing whether it is working or protection becomes thus essential. This excludes the possibility to employ the one-step search which set up a w/p pair jointly, without distinguish between working and protection mate. Therefore, only the two-step search can be employed with shared protection, even if this inevitably degrades the heuristics performance [9] (improvements to this approach are currently under study). Moreover, in setting up a new spare lightpath WDM channels that have been previously allocated to other lightpaths can be used provided that some conditions are satised. The WDM channel can not be used if: 1) it is already used by a working lightpath; 2) it is already used by one or more spare lightpaths whose working lightpaths are not link-disjoint to the working lightpath that we are trying to protect. Both Dijkstra and MLG have been consequently modied to allow checking of this conditions while performing RFWA of a protection lightpath. The two changes above involve the greedy OC-P mapping as well as the optimization section with shared protection. An additional change is required in the optimization section. The denition of a -ber given above for the dedicated case must be replaced. In the shared protection case a -ber is a ber used by lightpaths (and not a ber with busy

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

dent design is implemented, these approaches can be classied according to how the link-disjoint constraint is enforced. The two-steps and the one-step search that we have already explained above, are the two possible alternatives. Most of the papers in literature rely on the two-step search [22], [29], [30], almost always based on the application of the Dijkstra algorithm. Ref. [31] is the only paper, to our knowledge, that exploits the one-step search for dedicated path-protection. However, it employes the Suurballe [32] algorithm in monober networks and for dynamic trafc. We propose a different approach considering the shortest pair of disjoint paths between two nodes using Bhandari. Moreover we propose a prioritized multiple criteria in performing RFWA. The concept of using multi criteria heuristic was introduced in Refs [33] for OCH-SPRings dimensioning: this approach is able to consider multiple criteria having different objective functions and to choose the best ones. In our solution this approach is possible but we can evaluate also a combination of criterion giving each one a different priority as we can see later. We will demonstrate that combining together more than one criterion allows us to perform best results.

V. H EURISTIC METHOD APPLICATION In the following of this paper we present and comment some of the results we obtained by applying our optimization method to realistic network topologies. Two well-known case-study networks were considered: the National Science Foundation Network (NSFNet) and the European Optical Network (EON). Data regarding their physical topologies, represented in Fig. 4(a) and 4(b), were taken from Ref. [20] and Ref. [34], respectively. NSFNet has 14 nodes and 44 links, while EON has 19 nodes and 78 links. The OC-layer topologies are based on the static (symmetric) trafc matrices derived from real trafc measurements which are reported in the same references. The two OC-layer topologies comprise 360 and 1380 unidirectional connection requests for NSFNet and EON, respectively.

(a): NSFNet
Fig. 4. Considered network topologies.

(b): EON

We performed different sets of experiments aimed to point out various aspects of our heuristic approach. Each experiment comprises a series of optimizations of the same network with the same initial sorting rule, RFWA criteria and protection strategy and with the pre-assigned number of wavelengths assuming the values: (in some cases, also has been tested). We monitored the following output design parameters upon which we measured the performance of our tool:

total ber number ( ): is the total number of bers which have at least one active WDM channel and can not therefore be pruned from the network; deployed channels ( ): is the total number of WDM channels in the network, given by ; assigned channels ( ): is the total number of WDM channels of the network which have been assigned to some lightpath and are therefore in use. As mentioned above, the method we proposed is able to optimize the total length of the bers in the network (mL metric). However, for brevity, we do not report here results obtained with this cost function (such results are reported in Refs. [13], [35] for unprotected networks). In the following cases the cost function is thus always (mH metric). The rst set of experiments was aimed to compare the initial sorting rules we listed in section III-A. By optimizing the NSFNet we found that balanced sorting outperformed the other rules in the 37% of the experiments, both for unprotected connections and for dedicated path-protection. The percentages of cases in which longest rst, busiest pair rst and random lead to the best results are 26%, 26% and 11%, respectively. Therefore all the results we are going to display next have been obtained by adopting balanced sorting. We shall also note that the effect of the initial sorting rule on the nal result is quite weak (differences among the various rules are below 10% of the nal value): this is quite expected as an effect of the optimization cycle. The second set of experiments was carried out to compare different RFWA criteria combinations. The results reported in Fig. 5a (VWP case) and 5b (WP case) are obtained optimizing the NSFNet when dedicated path-protection is adopted for all the connections. The parameter is used for the vertical axis in place of in order to make the differences between the curves more visible, especially for high values of . The acronym describing each criterion has been dened in section III-A. The order in which the acronyms appear in the graph legends represent the priority by which they have been applied. The worst results are obtained when a single routing criterion is coupled with random ber and wavelength assignment (e.g. LLR-RNF-RNW). This is especially true with LLR, which tends to dramatically increase the length of the lightpaths. The heavy performance degradation is also a consequence of unconstrained RWFA [17]: constrained routing would be more appropriate in this case. The adoption of First Fit for ber and/or wavelength assignment leads to some improvement. An improvement is also obtained combining two routing algorithms (LLR and SPR). The best results are given by dual routing combined with First Fit ber and wavelength assignment. On average, SPR-FFF-FFW-LLR is the best combination, while LLR-SPR-FFF-FFW is the second best combination. The two graphs also indicate that the choice of the RFWA criteria is more important when the network is not equipped with wavelength converters than when wavelength conversion is implemented, though, as expected [1], the behavior in the VWP and WP scenarios is quite similar. The differences between RFWA criteria in terms of optimized costfunction value are not very large (except when compared to LLR-RNF-RNW). These differences become much more evident when the optimization duration is considered. In Fig. 6

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

NSFNet, WP
3000 LLR-RNF-RNW SPR-LLR-RNF-RNW SPR-RNF-RNW SPR-FFF-RNW LLR-SPR-FFF-FFW SPR-FFF-FFW-LLR

2800 Number of deployed channels, D

2600

2400

2200

2000

8 10

30

50

70

is again the worst combination. The results displayed above regard a single design experiment, but the other experiments performed conrm the behaviors of the tested multiple criteria. Finally let us mention that we also tested other criteria combinations, namely those with Pack and Spread wavelength and ber assignment and those giving ber/wavelength assignment higher priority than routing, but we obtained worse results than those presented here. A third set of experiments was carried out, aimed to compare our heuristic optimization method with other design strategies. For this set we employed the best performing initial sorting rule (balanced) and RFWA criteria combination (SPR-FFF-FFW-LLR) identied as reported above. We applied our heuristic method to both NSFNet and EON, assuming dedicated path-protection as survivability strategy and in a VWP scenario. Results are shown in the table I in terms of .
TABLE I
T OTAL NUMBER OF FIBERS

Number of wavelengths, W

(a)
NSFNet, VWP
2700

FOR VWP NETWORK SCENARIO

( DEDICATED PATH PROTECTION).

2600

NSFNet W 2 4 8 16 32 64
ILP Heu

Number of deployed channels, D

2500

ILP

Heu

2400

2300

2200

2100

LLR-RNF-RNW SPR-LLR-RNF-RNW SPR-RNF-RNW SPR-FFF-RNW LLR-SPR-FFF-FFW SPR-FFF-FFW-LLR

983 492 246 123 62 31

983 492 248 125 65 -

995 505 256 134 73 41

3692 1846 923 462 231 116

3693 1848 927 475 239 125

3704 1864 949 488 258 142

2000

1900 2 4 6 8 10 30 50 70 Number of wavelengths, W

(b)
Fig. 5. Optimized value for the different RFWA criteria combinations in the WP (a) and VWP (b) case.

we represented the evolution of during the optimization cycle for the RFWA criteria considered above in a VWP NSFNet when . On the horizontal axis the number of steps measures the number of partially used bers that the optimization algorithm have tried to prune. The value at the optimization step 0 is the result of the initial greedy OC-P mapping after pruning idle bers. The horizontal extension of the tracks in the graph is the total number of steps performed by the optimization process before the end, i.e. before the counter dened in section III-B reaches the value . The total number of steps is roughly proportional to the computation time required by the optimization. The graph shows the very remarkable reduction of the computation time obtainable with the dual-routing/First Fit criteria: in particular they make the initial heuristic set up of the connections extremely good, with very close to the nal optimized value of the cost function. A similar behavior (though less efcient) is displayed by SPR-FFF-RNW. Criteria combinations not based on First Fit starts form much higher values of and converge more slowly to the nal value, while LLR-RNF-RNW

Heuristic results are compared with data obtained by solving the same optimization problem with the Integer Linear Programming (ILP) technique. This data represent the absolute minimum of the cost function, the best any optimization method can give. Due to the high computational complexity of ILP, only cases with could be solved 6 , thus highlighting the importance of the scalability of heuristic methods in realistic planning situations. The comparison with ILP shows that our heuristic method leads to suboptimal results that are very close to the absolute minimum. The gap between heuristic and ILP is less than 8% when dedicated path protection and are considered. The gap tends to increase with : the large granularity by which capacity can be added to or subtracted from the physical network (bers with many wavelengths) amplies any small difference between heuristic and ILP mapping, making the former less effective. The granularity effect on the heuristic performance is also conrmed by measuring ber utilization: for the NSFNet, the ratio , equal to 99.1% with , uniformly decreases down to 80.0% with (from 99.8% to 86.5% for the EON). In the table I we also reported a lower bound on the total number of bers as a function of . Such a bound can be evaluated in several ways. We have decided to consider the so called capacity bound [17] as it is quite easy to evaluate. It is obtained by assigning to each connection the w/p link-disjointed pair routed using SPR, neglecting any wavelength conict and assuming
In the cases with , as well as in the WP scenario, ILP solver had to be stopped before reaching the absolute minimum of the cost function to limit memory occupation and run time.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

innite capacity on each link. This ideal OC-P mapping gives the minimum possible number of assigned channels necessary to set up all the connections. The corresponding bound of the total ber number is . In both the studied networks, the heuristic optimization curves are very close to the capacity bound, indicating a good topological matching between OC-layer and physical topology. The last set of experiments concerns the comparison between heuristic optimization of unprotected and dedicated and shared path-protected networks. As in the previous set, we considered NSFNet and EON, either in the VWP and WP scenarios, and solved heuristic optimization using balanced initial sorting rule and SPR-FFF-FFW-LLR. Figs. 7 and 8, represent the percentage extra number of ber necessary to implement protection in optimized networks, where protected unprotected unprotected . Dedicated path protection requires a great redundancy. For NSFNet is, in the worst case ( ) about 160% of the corresponding unprotected , while for EON redundancy is about 140%. Similar results are reported in [36] and obtained using a 17-nodes and 24-links network. Shared path-protection, allows a significant saving of bers, leading to the maximum extra-ber cost to 62% and 77% for NSFNet and EON, respectively. For the NSFNet shared protection in comparison with dedicated allows a saving in the number of ber and it varies between 17% and 40%; for the EON is about 25%-32%. A nal remark regards WP networks adopting dedicated path-protection. This protection strategy can be performed in two modes: protection switching (1+1) and protection transferring (1:1) [37]. Our design method does not distinguish between protection switching and transferring in the VWP case: the w/p pairs setup in dedicated protection can be indifferently used in the 1+1 and 1:1 modes. The 1:1 mode is instead treated differently in a network without converters. In a WP scenario the wavelength-continuity constraint may be imposed on the w/p pair as a whole or individually to each one of the two lightpaths of the w/p pair. Being in rst or in the second situation depends upon the fact that network transmitters and receivers are not provided or are provided with the wavelength agility function, as dened in Ref. [17]. In Fig. 7 is also represented in the case of 1+1 dedicated path protection without wavelength agility. When bers have large granularity, increases since more wavelength channels are not used.

NSFNet, W=16 VWP

140 136 132 0 80 160 240 Optimization step 320 400 480

Fig. 6. Optimization step adopting different RFWA combination criteria in the VWP network scenario and W=16.
NSFNet, SPR-FFF-FFW-LLR
200

Percent extra fiber number, M %

150

100

dedicated VWP shared VWP dedicated WP shared WP dedicated WP without w. agility

50

0 2 4 6 8 10 30 50 70 Number of wavelengths, W

Fig. 7. Growing of the number ber protection for the NSFNet.

in the dedicated and shared path

SPR-FFF-FFW-LLR

VI. C ONCLUSIONS In the work we have presented we focus on the issue of planning and optimization of WDM networks providing dedicated or shared path-protection. The design method we have developed is able to perform the automatic planning of the working and protection lightpaths while optimizing the amount of bers that must be installed to satisfy all the connection requests. The optimization procedure follows a heuristic approach which allows to attain good sub-optimal results with a moderate computational effort. It can be applied to VWP and WP networks under the assumption that the requested connections are static and known a priori. The method guarantees that working and protection lightpaths are link-disjoint by the exploitation of two alternative path-search algorithms. The two-step algorithm is

Percent extra fiber number, M %

100

50 dedicated VWP shared VWP 0 2 4 6 8 10 30 50 70 Number of wavelengths, W dedicated WP shared WP

Fig. 8. Growing of the number ber protection for the EON.

in the dedicated and shared path

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

used for shared protection, while a more efcient one-step algorithm can be applied to the dedicated case. We have developed an original one-step algorithm deriving it from the known Bhandari algorithm by extending its usage to the WDM environment, thanks to a powerful network-state representation: the multiber layered graph. Exploiting this graph we are able to solve link-disjoint routing, ber and wavelength assignment in an integrated way. The results presented in the paper, obtained using NSFNet and EON as case-study topologies, give an interesting insight regarding the sensitivity of an optimized WDM network to such parameters as number of wavelengths, wavelength conversion capability, and RFWA criteria. In particular, we show the effectiveness of multiple prioritized RFWA criteria in obtaining good sub-optimal mappings of the OC-layer over the physical topology and in abbreviating the computational time of the optimization procedure. R EFERENCES
[1] T.E. Stern and K. Bala, Multiwavelength Optical Networks: A Layered Approach, Addison Wesley, 1999. [2] ITU-T Intern. Telecom. Union Telecom. Standard. Sector, Architecture of Optical Transport Networks, Number G.872. 1999. [3] G. Maier, S De Patre, A. Pattavina, and M. Martinelli, Optical network survivability: protection techniques in the WDM layer, Photonic Networks Communications, accepted for publication, 2002. [4] M. Chbat, E. Grard, L. Berthelon, A. Jourdan, A. Leclert, et al., Toward wide-scale all-optical transparent networking: the ACTS Optical Pan-European Network (OPEN) project, IEEE Journal on Selected Areas in Communications, vol. 16, no. 7, pp. 12261244, Sept. 1998. [5] N. Wauters and P. Demeester, Design of the optical path layer in multiwavelength cross-connected networks, IEEE Journal on Selected Areas in Communications, vol. 14, no. 5, pp. 881892, June 1996. [6] Ken-Ichi Sato, Advances in Transport Networks Technologies, Artech House Publishers, 1996. [7] S. Chaudhuri, E. Bouillet, and G. Ellinas, Addressing transparency in DWDM mesh survivable networks, in OFC 2001, 2001, vol. 2, pp. TuO51TuO53. [8] J. Strand, A. Chiu, and R. Tkach, Issues for routing in the optical layer, IEEE Communication Magazine, vol. 39, no. 2, pp. 8187, Feb. 2001. [9] R. Ramamurthy, S. Sengupta, and S. Chaudhuri, Comparison of centralized and distributed provisioning of lightpaths in opticla networks, in OFC 2001, 2001, vol. 1, pp. MH4.1MH4.3. [10] R. Ramamurthy, Z. Bogdanowicz, S. Samieian, D. Saha, and B. Rajagopalan, Capacity performance of dynamic provisioning in optical networks, Journal of Lightwave Technology, vol. 19, no. 1, pp. 4048, Jan. 2001. [11] I. Chlamtac, A. Farago, and T. Zhang, Lightpath (Wavelength) Routing in Large WDM Networks, IEEE Journal on Selected Areas in Communications, vol. 14, no. 5, pp. 909913, June 1996. [12] C. Chen and S. Banerjee, A new model for optimal routing and wavelength assignment in wavelength division multiplexed optical networks, in Proceedings, IEEE INFOCOM 96, 1996, vol. 2a.4.1-8, pp. 164171. [13] G. Maier, A. Pattavina, L. Roberti, and T. Chich, Static-lightpath design by heuristic methods in multiber WDM networks, in Proceedings, Opticomm 2000 SPIE Conf., Oct. 2000, pp. 6475. [14] R. Ramaswami and K. N. Sivarajan, Routing and wavelength assignment in all-optical networks, IEEE/ACM Transactions on Networking, vol. 3, no. 5, pp. 489500, Oct. 1995. [15] B. Mukherjee, Optical communication networks, McGraw-Hill, 1997. [16] G. Maier, M. Martinelli, A. Pattavina, and M. Pierpaoli, Performance analysis of wavelength division multiplexing mesh networks, in Optical Networking, A. Bononi, Ed. Sept. 1999, pp. 5263, Kluwer. [17] S. Baroni, Routing and Wavelength Allocation in WDM Optical Networks, Ph.D. thesis, Department of Electronic and Electrical Engineeing, UCL London, 1998. [18] K.-I. Sato, Advances in Transport Network Technologies - Photonic Networks, ATM, and SDH, Artech House, Norwood, MA, rst edition, 1996. [19] G. Maier, A. Pattavina, L. Roberti, and T. Chich, A heuristic approach for the design of static multiber WDM networks: principles and applications, Optical Network Magazine, accepted for publication, 2002.

[20] Y. Miyao and H. Saito, Optimal Design and Evaluation of Survivable WDM Transport Networks, IEEE Journal on Selected Areas in Communications, vol. 16, no. 7, pp. 11901198, Sept. 1998. [21] S. Baroni, R.J. Gibbens P. Bayvel, and S.K. Korotky, Analysis and Design of Resilient Multiber Wavelength-Routed Optical transport networks, IEEE Journal of Lightwave Technology, vol. 17, pp. 743758, may 1999. [22] C. T. Hiang, S. Gangxiang, S. K. Bose, and L. Chao, Designing WDM Optical Network for Reliability: Routing Lights Paths Efciently for Path Protection, Optical Fiber Conference 2000, vol. 3, pp. 5052, 2000. [23] R. Bhandari, Survivable networks, algorithms for diverse routing, Kluwer Academic Publishers, 1999. [24] D. A. Dunn, W. D. Grover, and M. H. Gregor, Comparison of k-shortest paths and maximum ow routing for network facility restoration, IEEE Journal on Selected Areas in Communications, pp. 8889, 1994. [25] M. Alanyali and E. Ayanoglu, Provisioning Algorithms for WDM Optical Networks, IEEE/ACM Transactions on Networking, vol. 7, no. 5, pp. 767778, oct 1999. [26] S. Ramamurthy and Biswanath Muckherjee, Survivable WDM Mesh Networks, part. I - Protection, in Proceedings, IEEE INFOCOM99, 1999, vol. 2, pp. 744751. [27] S. Ramamurthy and Biswanath Muckherjee, Survivable WDM Mesh Networks, part. II - Restoration, in Proceedings ICC99, 1999, vol. 2, pp. 202330. [28] B. Van Caenegem, W. Van Parys, F. De Turck, and P.M. Demeester, Dimensioning of Survivable WDM Networks, IEEE Journal on Selected Areas in Communications, vol. 16, no. 7, pp. 11461157, Sept. 1998. [29] V. Anand and C. Qiao, Static versus Dynamic Establishment of Protection Paths in WDM Networks, part. I, in IEEE ICC00, 2000. [30] H. Zang, J.P. Jue, and B. Mukherjee, A Review of Routing and Wavelength Assignment Approaches for Wavelength-Routed Optical WDM Networks, Optical network magazine, 2000. [31] Weifa Liang, Robust Routing in Wide-Area WDM Networks, in Parallel and Distributed Processing Symposium, Proceedings 15th International, 2001. [32] J. W. Suurballe, Disjoint paths in a network, Networks, vol. 4, pp. 125145, 1974. [33] R. Meersman, W.Van Parys, P. Arijs, P. Demeester, E. Iannone, M. Pierpaoli, and F. Bentivoglio, Multiple Objective Heuristic for Ring Loading and Logical Wavelength Assignment in OCH-SPRings, in ONDM 2001 Preliminary Proceedings, 2001. [34] A. Fumagalli, I. Cerutti, M. Tacca, F. Masetti, R. Jagannathan, and S. Alagar, Survivable Networks Based on Optimal Routing and WDM SelfHealing Rings, in Proceedings, IEEE INFOCOM 99, 1999. [35] A. Pattavina, G. Maier, and M. Tornatore, WDM network optimization by ILP based on source formulation, IEEE INFOCOM 2002, accepted for publication. [36] Georgios Ellinas, Mesh Restoration for Link and Node Failures in Optical Networks with Shared Risk Groups, in Proceedings of Gigabit Networking Workshop, 2001, www.comsoc.org/tcng/conference/gbn2001. [37] Tsong Wu, Emerging Technologies for Fiber Network Survivability, IEEE Communication Magazine, pp. 5874, feb 1995.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.