the network configuration and traffic routing. The proposed global provisioning solution can be easily combined with dynamic routing solutions, providing the ...
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
1141
A Multilayer Solution for Path Provisioning in New-Generation Optical/MPLS Networks Roberto Sabella, Senior Member, IEEE, Marina Settembre, Gianpaolo Oriolo, Federica Razza, Fabio Ferlito, and Giulia Conte
Abstract—This paper proposes an offline solution for global path provisioning in new-generation optical networks based on the generalized multiprotocol label switching (GMPLS) paradigm. This solution is based on a multilayer approach, which involves both the optical and the electrical layers and optimizes the network configuration and traffic routing. The proposed global provisioning solution can be easily combined with dynamic routing solutions, providing the network with the possibility of reacting promptly to traffic changes. Data flows are assumed to be structured into label switched paths (LSPs), which represent the connection in a GMPLS-based network, at any hierarchical level. The global provisioning issue is a difficult optimization problem. As a solution, we propose a new heuristic algorithm based on the shortest path computation and a mathematical programming approach, which makes use of the optimization solver CPLEX. A large computational study shows the effectiveness of the former, in terms of quality of the solutions. The advantages of the multilayer provisioning strategy are analyzed in a relevant case study by evaluating the network congestion. Index Terms—Generalized multiprotocol label switching (GMPLS), heuristic approach, integer linear programming (ILP) formulation, logical topology, multilayer routing, new-generation optical networks, path provisioning algorithm.
I. INTRODUCTION
T
HE REMARKABLE growth of Internet traffic during the past years witnesses the success of Internet protocol (IP)-based networks in providing services of different kind. There is a wide consensus among market analysts that IP will constitute the common platform for both traditional data and new real-time and multimedia services. As a consequence, new-generation infrastructures are going to handle huge IP-based traffic, supporting different types of service, with several levels of quality of services (QoS). Moreover, the network infrastructure will have to provide a more flexible and dynamic utilization of the network resources to handle the consistent variation of traffic distribution, both due to the unpredictability of Internet traffic and traffic demand varying with time and depending on the peculiarity of the different services [1], [2]. In this framework, traffic engineering (TE) is becoming an increasingly indispensable network function Manuscript received August 13, 2002; revised February 18, 2003. R. Sabella and M. Settembre are with Ericsson Lab Italy, R&D Division, 00133 Rome, Italy. G. Oriolo, F. Razza, and F. Ferlito are with the Dipartimento di Informatica Sistemi e Produzione (DISP), Università degli Studi di Roma “Tor Vergata,” 00133 Rome, Italy. G. Conte is with the CoRiTel—Consorzio di Ricerca sulle Telecomunicazioni, 00040 Morena, Rome, Italy. Digital Object Identifier 10.1109/JLT.2003.811424
to map efficiently traffic demands onto the network topology in order to make an effective and dynamic utilization of the network resources, while meeting QoS requirements (i.e., delay, packet loss, and protection) [3]–[5]. The ATM model allows some of these traffic engineering functions. However, the success of the IP paradigm is so widespread that TE functions have to be redesigned in an IP context. For this reason, the Internet Engineering Task Force (IETF) introduced the multiprotocol label dwitching (MPLS) technique [6], [7]. Roughly speaking, MPLS reproposes some concepts, such as virtual connection, i.e., label switched path (LSP), previously introduced through asynchronous transfer mode (ATM) and frame relay, and harmonizing the connectionless nature of the Internet and the intrinsically circuit-switched transport layer, e.g., synchronous digital hierarchy/synchronous optical network (SDH/SONET) and wavelength-division multiplexing (WDM). In particular, MPLS can direct a flow of IP packets along a predetermined path across the network, calculated by the ingress router taking into account user and network constraints. Moreover, MPLS has the ability to reserve link bandwidth and modify link attributes, allowing handling of QoS and protection/restoration of data flows. Generalized multiprotocol label switching (GMPLS) [8], [9] extends the features of the MPLS technique to different technologies, network layers, and architectures, providing a gradual and future-proof approach toward new-generation networks. In practice, the GMPLS control plane manages heterogeneous network elements (e.g., IP/MPLS routers, ATM switches, SDH/SONET devices, and even optical elements) in order to efficiently and cooperatively respond to traffic demands and network reconfiguration. Using a common control plane, which is IP-based, facilitates the actualization of cross-layers TE, optimizing network configuration and traffic routing. In general, in order to route the LSPs entering the network, both offline and online approaches can be considered. Specifically, a global path provisioning can be accomplished offline any time an appreciable change of traffic is recognized, involving a global routing algorithm to accommodate a set of requests in an optimal way on a calculated logical topology, while a dynamic routing approach can be applied to accommodate online a new request by means of signaling protocols. A combined use of both online and offline approaches allows the network to react adequately to traffic changes and to realize the “bandwidth on demand” service. In [10] and [11], we introduced such a hybrid approach: a global path computation according to an expected traffic matrix is accomplished offline through a path provisioning function, which aims at
0733-8724/03$17.00 © 2003 IEEE
1142
minimizing the use of network resources, while actual traffic requests are accommodated online by means of a dynamic routing function in order to reply timely to traffic changes due to the unpredictable component of IP traffic. In this framework, we propose an offline configuration strategy (global provisioning) that assigns the routes for every requesting LSP, minimizing the network resource usage and, thus, facilitating the online accommodation of new requests. The proposed procedure plans the network configuration following a multilayer routing approach. This means that the procedure designs a logical topology by selecting the optical paths (lightpaths) and their routes on the physical topology and concurrently routes the connection requests (LSPs) onto such a logical topology. Since, in our multilayer approach, the routing of the lightpaths and the LSPs is performed simultaneously, the complexity of the problem is higher when compared with a simple multihop approach. We propose two different approaches for solving the multilayer provisioning problem: a heuristic combinatorial algorithm based on the shortest path computation and an integer programming approach based on the use of the optimization solver CPLEX. Since the size of the real problems advises against exact solutions, we mainly use CPLEX on a suitable relaxation of the problem in order to get estimates on the value of optimal solutions. These estimates have then been used in a large computational study, which has certified the quality of the solution found by the heuristic algorithm. Given its efficiency and simplicity, we propose it as an effective solution to the provisioning problem. The paper is structured as follows. In Section II, the network scenario and the concept of multilayer routing are detailed, while in Section III, the proposed strategy and related algorithms for the global provisioning are presented, considering both a heuristic and an exact approach. In Section IV, the comparison between the two solving approaches and the performance analysis of the proposed strategy are presented and discussed for different network topologies. Finally, some conclusions are derived. II. REFERENCE SCENARIO AND MULTILAYER PROVISIONING Our reference scenario is based on a GMPLS network consisting of an optical transport (WDM) layer, whose nodes are represented by optical cross-connects (OXCs) that perform wavelength routing operations, and an MPLS layer, whose nodes are represented by MPLS routers, namely label switching routers (LSRs). In this framework, different network scenarios can be envisaged. In fact, GMPLS can support either a client/server relationship among the layers, as in the overlay architectural model, or a peer architectural model, where all the network devices act as peer devices sharing a common and complete topological view of the whole network. In such a scenario, the overlay model assumes that the optical network interacts with an MPLS layer by means of a separate control plane. It means that the MPLS network asks the optical network for a connection, and the optical network performs the routing, handling its own resources. As a result, since different control planes perform routing functions separately, the utilization of network resources is suboptimal. A single control plane
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
Fig. 1. Connection request can be routed onto a single lightpath or on a concatenation of lightpaths.
aware of the whole network resources can perform routing function in an optimal way. That is the case of the peer model scenario, which takes advantage of both MPLS and WDM routing capabilities. The provisioning procedure can be performed either considering only the WDM layer or the WDM and the MPLS layers jointly. In the first case, the LSPs are aggregated at the edge devices of the optical domain in order to have wavelength requests entering the WDM network. This approach is called the single-layer approach to indicate that the routing and wavelength assignment issue is considered only at the WDM layer on the basis of a traffic matrix expressed in terms of number of wavelengths per any pair of nodes [12]–[15]. The most recent literature considers the routing problem in a two-layered network architecture [16]–[18], treating jointly the optical and the electrical domain. In this case, the provisioning problem can be schematized in two main steps: 1) the design of a logical topology of the optical network layer and 2) the routing of the data flows at the IP/MPLS layer (namely the LSPs) onto the logical topology. This means that a connection request (i.e., LSP) can be routed on a logical concatenation of optical paths (multihop routing), and it can be aggregated with other connections at the intermediate nodes, sharing the same optical path (Fig. 1). Actually, the path provisioning aims at supplying the lightpaths and LSP routes configuration at a minimum cost according to the objective function. Generally, the objective functions are derived from a criterion that is in agreement with the network operator policy. In this paper, the adopted criterion is to find configurations able to support potential variations in the offered traffic demands without requiring established routes to be reconfigured. Intuitively, for a given set of traffic demands, the fewer amounts of resources a configuration solution uses, the more resources are available for allocating new routes in response to future fluctuations of the traffic with no rearrangement of the established configuration. In terms of minimization of the employed resources, a multihop approach is more suitable than the single-layer one. Actually, the multihop routing makes independent the two subproblems: the lightpath selection and the LSP routing. In this sense, the multihop routing is more flexible than the single-layer approach. Moreover, this paper faces the global multihop provisioning problem with a multilayer approach that consists of the res-
SABELLA et al.: MULTILAYER SOLUTION FOR PATH PROVISIONING IN NEW-GENERATION OPTICAL/MPLS NETWORKS
olution of the routing problem by simultaneously solving the two subproblems. This multilayer routing approach allows an LSP to be routed on a concatenation of optical connections in a single routing instance, leading to a more effective use of network resources. The paper presents two approaches to solve the path provisioning problem: a heuristic combinatorial algorithm and an algebraic approach using an optimization solver (CPLEX). Specifically, the former solves all the instances of the provisioning problem efficiently, while the latter is mainly employed to find either exact solutions (to small instances) or good lower bounds, and it can therefore be used to assess the quality of the heuristic algorithm. The details are reported in the following section.
Fig. 2.
1143
Input and the output of our multilayer provisioning.
III. PROPOSED SOLUTION FOR GLOBAL MULTILAYER PROVISIONING Fig. 2 depicts the provisioning problem, showing its inputs and outputs. Basically, a solution to the problem provides both the design of the logical topology, i.e., the set of lightpaths to be set up and the route for each lightpath, and the routing (in the logical network) of each LSP, on the basis of a traffic matrix and a given physical network topology. Specifically, the offered traffic can be described by a suitable is a vector of LSPs traffic matrix, whose generic element between node and node
(1)
, where and are the minwith imum and the maximum possible bandwidth of an LSP, respectively. Each LSP is assumed to be unsplittable, which means that an LSP is not allowed to be split between different paths. The physical topology of the optical network, assumed to be set during the network planning phase, is composed of a set of nodes connected by a set of arcs in a given mesh topology, as sketched in Fig. 3. Each arc stands for a bundle of fibers between two adjacent nodes, and a single fiber can support a fixed number of wavelengths, as shown in Fig. 4(a). Each node can consist of either an LSR integrated with an OXC or a stand-alone OXC, depending on whether or not the node is capable to originate and terminate LSP traffic. Two types of ports can be recognized in each OXC: 1) interoffice ports, supporting fibers coming from or going to adjacent OXCs and 2) intra-office ports, connecting the OXC to the MPLS routers in the same node [see Fig. 4(b)]. A lightpath is established by setting up a sequence of OXCs connected by fiber links and wavelengths so that a continuous physical path exists from the ingress to the egress nodes. The OXC’s are assumed to have full wavelength conversion capability, and therefore, the wavelengths are assigned link by link to each lightpath. The outputs of the algorithm consist of the set of lightpaths representing the logical topology of the optical network, the
Fig. 3. Reference network.
(a)
(b) Fig. 4. Network elements: (a) optical link and (b) OXC.
routes for those lightpaths, and the routes of the LSPs onto those lightpaths. In the proposed multilayer approach, the combined problem of designing the logical topology and routing all the LSP requests onto it is simultaneously addressed. Different objective functions can be defined for the provisioning problem according to the network operator policy. For instance, the objective function could be the maximization of the efficiency of network resources consumption (optical resources, electrical resources, or both), the minimization of the traffic lost, or the average packet hop distance [17], [18]. The specific objective function, which has been considered, is the minimization of the congestion on the network resources. Formally, it is defined as the maximum ratio between used and
1144
available resources over all the optical resources, that is, wavelengths on each optical link, intraoffice ports incoming to each LSR node, and intraoffice ports outgoing from each LSR node. The rationale is that the less congestion arises in the network from serving the foreseen traffic matrix , the more likely we will be able to accommodate online new requests of connections or to handle fluctuations of the traffic demand [10], [11]. A. Linear Programming Formulation The problem of routing a set of commodities over a given network so as to minimize the congestion (minimum congestion problem) is known to be NP-hard [19], and in fact, there are no known algorithms solving it in polynomial time. It is straightforward to check that our problem, which deals with unsplittable flows and other additional constraints, is, to a greater extent, a member of the NP-hard problems class. For its solution, we propose two different approaches: the first solves an integer linear programming formulation of the problem by the optimization solver CPLEX, and the second is heuristic and is based on shortest path computation. The two approaches are complementary. In fact, while CPLEX is able to find exact solutions to the problem only for small instances, the heuristic approach is able to efficiently solve large instances of the problem, but it does not provide any guarantee on the quality of solutions. We have defined a suitable (linear) relaxation of the integer linear program (ILP) so as to always and efficiently get a lower bound on the value of an optimal solution for real-life instances. Then, we used this lower bound for assessing the quality of the solution found by the heuristic approach. Before delving into the details of these approaches, we state the integer linear programming formulation of the problem. We use the following notations. is the physical network (directed graph) constituted by nodes and physical arcs. is the logical network (directed graph) constituted by nodes and logical arcs (i.e., lightpaths). is the set of network nodes, with each node corresponding to either an LSR integrated with an OXC or a stand-alone OXC. is the set of physical arcs, with each arc corresponding to a bundle of parallel fibers. is the set of lightpaths, with each lightpath being defined by a path on . In general, is a multiset, since several (parallel) lightpaths may correspond to a same path on . is the set of commodities to be routed. Each commodity , defined by a triple ( , , ), its volume, origin, and destination, respectively, corresponds to an LSP with bandwidth going from to [hence, for some between 1 and , see (1)]. is the subset of lightpaths of using the physical arc ( , ). is the subset of lightpaths of starting from node .
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
is the subset of lightpaths of ending in node . is the capacity of a wavelength. is the total number of wavelengths available on the physical arc ( , ). is the number of intraoffice ports outgoing from node (not a stand-alone OXC node). is the number of intraoffice ports incoming to node (not a stand-alone OXC node). The decision variables used in the formulation are if the commodity is routed through lightpath otherwise if lightpath is selected otherwise is the congestion, defined as the maximum ratio between used and available capacity over the following resources: wavelengths on each physical arc; intraoffice ports incoming to each node (not a stand-alone OXC); and intraoffice ports outgoing from each node (not a stand-alone OXC). The formulation of the global provisioning problem can be modeled as shown at the bottom of the next page.The objective function is to minimize the congestion , subject to the equations (1a)–(8a),where 1a) for each lightpath, the total amount of traffic passing through the lightpath does not exceed the capacity of the lightpath itself if the lightpath is selected, otherwise it is 0; 2a) congestion is no less than the congestion on each arc of the physical network, defined as the ratio between the number of selected lightpaths using that arc and the number of available wavelengths on it; 3a) congestion is no less than the incoming congestion on a node (not a stand-alone OXC), defined as the ratio between the number of selected lightpaths starting at that node and the number of intraffice ports entering that node; 4a) congestion is no less than the outgoing congestion on a node (not a stand-alone OXC), defined as the ratio between the number of selected lightpaths terminating at that node and the number of intraoffice ports exiting from that node. 5a) commodity has to leave from node ; 6a) commodity has to arrive at node ; 7a) flow conservation of commodity is at each intermediate node; and , making the flows 8a) integrality is on unsplittable. B. Heuristic Approach The heuristic algorithm (HA) consists of the following two phases: 2) paths are set up by means of a successive shortest path algorithm; 3) a local search procedure is used for improving the solution.
SABELLA et al.: MULTILAYER SOLUTION FOR PATH PROVISIONING IN NEW-GENERATION OPTICAL/MPLS NETWORKS
Successive Shortest Path Algorithm: First of all, the set of LSPs (commodities) is sorted according to the minimum number of hops between the source and the destination; then, the LSPs are sequentially routed, starting with an LSP with a minimum number of hops. In order to keep the congestion as low as possible (Fig. 5), at each step, we try to route the LSP on paths made only of lightpaths, which have been established in the previous steps; otherwise, if such a path does not exist, we try to set up a new lightpath between the end nodes of the LSP. For LSP sorting, different possibilities have been tested: the sorting based on the minimum number of hops always resulted in the one leading to the best performance. Intuitively, it can be explained as follows. By starting the sequential routing algorithm with the shortest commodities, in terms of minimum number of hops, the first setup lightpaths are between near LSRs. The successive commodities find a logical topology whose links directly connect adjacent nodes: this topology is suitable to route LSPs among multihop paths where no new lightpaths have to be established. In particular, at the th step, we consider the th LSP with source , destination , and bandwidth . In order to route the of lightpaths, LSP on the logical network, we consider a set so far established, whose spare capacity is at least . The cost is equal to its residual bandwidth. Then, of each lightpath of if paths exist from to using only arcs from , we choose the shortest one (e.g., by Dijkstra’s shortest path algorithm) and path in the route the LSP on it; otherwise, we look for an be physical network. For each physical arc ( , ) of , let the number of lightpaths crossing a wavelength of ( , ), which be the set of physical have been established so far, and let . We give to each arc ( , ) of arcs ( , ) so that cost (2)
(1a) (2a) (3a) (4a) (5a) (6a) (7a) (8a)
1145
Then, if there are paths from to using only arcs from , we choose the shortest one (e.g., by Dijkstra’s shortest to on path algorithm), establish a (new) lightpath from this path, and route the LSP on it. Otherwise, we discard by considering the next the LSP and move to the step LSP. Improving the Solution: In the local search phase, the algorithm attempts to improve the solution achieved in the previous phase by employing an iterative procedure (Fig. 6). This is done in three steps. Step 2) We order all the physical links according to the congestion from the most to the least congested link and select the most congested link . Step 3) We consider the lightpath crossing with the maximum spare bandwidth. Step 4) We try to reduce the congestion on by rerouting all of the LSPs crossing on other paths made of (already established) lightpaths. If this is possible, then we remove and go back to Step 1); otherwise, we set equal to the next physical link in the ordered set and go back to Step 2). It is important to highlight that every LSP must be rerouted in order to remove the corresponding lightpath. For that reason, at this step, the commodities are sorted from the one with the highest value of bandwidth to the one with the lowest one. In fact, if it is not possible to reroute the LSP with the highest value of bandwidth at the first step of the rerouting phase; there will be no point in trying to reroute the other LSPs. In order to find a new path for the single LSP, we consider of lightpaths, whose spare capacity is at least the LSP the set bandwidth, and we assign a cost of each lightpath equal to its residual bandwidth. Then, if there exist paths using only arcs from , we choose the shortest one.
1146
Fig. 5.
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
Sequential phase of the HA.
The search stops when all the arcs have been considered and we have not succeeded in reducing the congestion on any. Observe that, since the total number of lightpaths decreases every time we remove a lightpath , the search terminates in a finite number of steps. Observe also that, if the local search procedure produces a new routing solution, further physical resources might have become available, and it might be possible then to accommodate some commodities, which were previously discarded. Hence, we start again our algorithm with the sequential phase, considering those LPSs as the new input traffic matrix. The iteration of the sequential and the local search phases is repeated up to when it is no longer possible to route other LSPs. This happens either when all the LSPs have been successfully routed or when there is no way to route the remaining LSPs.
C. ILP Approach Solving the ILP formulation of a problem through an optimization solver like CPLEX has the advantage of providing either an optimum solution, when the integral program can be solved, or, at least, a lower bound on its value when the size of the integer program advises against an exact solution. Such lower bound is provided by the value of the optimal (fractional) solution to the so-called linear relaxation, that is, the linear program we get by discarding the integrality constraints. Nevertheless, when the size of a problem is large, even solving the linear relaxation may be a challenging task for CPLEX. This is the case with our problem, as it is shown in Table I.
SABELLA et al.: MULTILAYER SOLUTION FOR PATH PROVISIONING IN NEW-GENERATION OPTICAL/MPLS NETWORKS
1147
Fig. 6. Search phase of the heuristic algorithm.
The time needed to get the optimum solution of the linear relaxation is rather short when dealing with small graphs (e.g., 6 min are enough to determine the optimum for a six-node instance), but it grows a lot as the size of the graphs increases. In fact, more than 6 h are needed to solve the linear relaxation for a seven-node instance, while 24 h is not a long enough time to get the optimum solution of the linear relaxation of the problem for an eight-node instance. In such cases, finding an optimal integer solution to the formulation is hopeless. One of the problems we observed was that the number of possible lightpaths for large instances tends to explode. Hence, we decided to employ a suitable technique for reducing the number of possible lightpaths (space reduction technique): in fact, while in large networks thousands or even millions of lightpaths could be potentially set up, the very number of lightpaths, which will be set up in an optimum solution, is far smaller. An effective criterion to achieve this space reduction would be that of choosing only those lightpaths used by the solution of the linear relaxation of the model. Nevertheless, as Table I shows, this criterion was not suitable for our networks, because of the long time needed
to solve the linear relaxation of the model. In order to overcome such a difficulty, we introduced a further relaxation of the model that could be finally solved by CPLEX. We use this solution for making the space reduction. In more detail, we 2) define a further relaxation of the model, which is easier to solve (this is done by discarding other constraints of the model besides integrality ones); 3) use the fractional solution to this relaxation for reducing (substantially) the number of lightpaths for the integer program; 4) solve the integer program by considering only the subset of potential lightpaths defined at step 2). The basic idea of the further relaxation is depicted in Fig. 7. Instead of considering the physical capacity of each single lightpath, it is possible to consider the aggregated capacity of the whole bundle of fibers carrying such a lightpath. In other words, the congestion constraints are surrogated, which means that they are substituted by slightly looser constraints: hence, the new problem is called a surrogate problem (model). Of course, any solution to the original problem will be a solution to
1148
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
TABLE I VARIABLES AND CONSTRAINTS NUMBER INCREASES AS THE NETWORK TOPOLOGY SIZE GROWS
optimum (fractional) solution as well. The effectiveness of the surrogate model reduction is better shown in Fig. 8. As we already stated, the solution to the surrogate model is used to achieve a large reduction in the number of integer variables for the integer program. In fact, the only possible lightpaths considered by the integer program are those that were used by the fractional solution to the surrogate model. A general sketch of our approach is resumed in Fig. 9. IV. PERFORMANCE ANALYSIS Fig. 7.
Basic idea for the surrogate model.
the surrogate problem but not vice versa. Hence, in general, the value of an optimal solution to the surrogate problem will give only a lower bound on the value of a solution to the original problem; on the other hand, since the surrogate problem has less variables and constraints than the original one, it is therefore likely that the solving time will be shortened as well. For the sake of completeness, the equation at the bottom of the next page is the linear integer program that we defined as surrogate model. Observe that constraints (1a), (2a), (3a), and (4a) of the general model introduced in Section III-B have been replaced (“surrogated”) by constraints (1b), (2b), and (3b) of the surrogate model. In practice, the previous model allows reductions in the time needed to get the optimum solution of the linear relaxation: in every simulation made for an eight-node instance, the optimum was found within 1 min, whereas previously 24 h were not enough to determine it. Besides, when the size of the instance allowed us to compute the optimum solution for the general model so that a comparison between the two approaches was possible, the surrogate model always proved to find an
In this section, two kinds of analysis are discussed. Namely, Section IV-B reports the comparison between the heuristic and algebraic solutions for the global multilayer provisioning problem for different network topology sizes. In Section IV-C, the heuristic approach has been adopted to investigate the performance of the proposed multilayer provisioning strategy compared with the single-layer one in a practical case study that cannot be solved by an algebraic solver. The details of the simulation environment are discussed in the Section IV-A. A. Simulation Environment The three network topologies shown in Fig. 10 have been employed for comparing the heuristic and algebraic approaches. In particular, since they consist of a limited number of nodes (five, six, and eight, respectively), it is possible to obtain solutions employing both the proposed approaches. However, in order to show the real advantages of the proposed multilayer provisioning, the performance results are reported for a more interesting case represented by the NSF Network, shown in Fig. 11, which has been extensively used in the literature for testing routing algorithms. For the NSF Network, 30 wavelengths per
SABELLA et al.: MULTILAYER SOLUTION FOR PATH PROVISIONING IN NEW-GENERATION OPTICAL/MPLS NETWORKS
Fig. 8.
1149
Time to solve linear relaxation of the multi-ayer model versus topology dimension.
optical link and per intraoffice port are assumed.In every considered network, the wavelength capacity is assumed equal to 2.5 Gb/s. In the simulations, the LSP traffic matrix has been derived from an aggregated traffic matrix, as described in (1). In particular, for each pair of nodes and , a random number is picked, . The aguniformly distributed between 0 and 1, , gregated demand of all the LSPs from node to node , , where is a scaling parameter. Afterward, a set of is random number, , representing the bandwidth associated with each LSP from node to node , are generated so that and , where and are the minimum and the maximum bandwidth requested
(1b)
(2b)
(3b) (4b) (5b) (6b) (7b)
by an LSP. The sum of the bandwidth requested by all the LSPs for each pair of nodes is the traffic volume TV
(3)
and have been In the simulations, the default values of fixed to 100 Mb/s and 1.5 Gb/s, respectively; otherwise, they are specified. B. Comparison Between the HA and ILP Approaches In order to test the quality of the HA, we evaluated the “distance” between its solutions and the solutions found by CPLEX,
1150
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
In the case of the six-node topology, the agreement between the heuristic approach and the algebraic one is still very good, even if a small difference can be observed at coarse granularity. Regarding the eight-node topology, the solutions found by means of the HA show smaller congestion values than the ones provided by the solver. It is important to observe that the difference between the heuristic solution and the optimal solution to the linear relaxation of the surrogate problem does not overtake the values of 0.1 and 0.12 for the six-node network and for the eight-node network, respectively. Since the latter solution is always a lower bound on the optimal solution, it can be stated that the heuristic algorithm is able to provide widely acceptable solutions for the provisioning problem. C. Advantages of the Multilayer Strategy
Fig. 9. Proposed algebraic approach.
with respect to different network topology sizes and commodity granularities and for a fixed traffic volume. In particular, in order to test the effectiveness of the algorithm in different conditions, three levels of granularities for the bandwidth , requested by each LSP, have been considered, as follows: 500 Mb/s Fine granularity: 100 Mb/s 1000 Mb/s Intermediate granularity: 500 Mb/s 1500 Mb/s. Coarse granularity: 1000 Mb/s In Tables II–IV, a comparison among the congestion value obtained by means of the heuristic approach and the solver (ILOG/CPLEX [21]) are reported for the five-node to the six-node and to the eighth-node network topologies, respectively. As far as the five-node topology is concerned, it is possible to compare the HA with the exact solution of the integer problem. In Table II, the optimal solution to the linear relaxation of the program (lower bound) is also reported for the sake of completeness. It is worth noting that the performance of the HA is very good: the different congestion values obtained for fine granularity correspond to the activation of only one more lightpath with respect to the configuration provided by the algebraic approach so that, even in this case, the performance is satisfactory. In the case of six-node and eight-node topologies, CPLEX is not able to find the optimal solution for the integer problem in a reasonable time. Actually, even the solution of the linear relaxation of the problem takes more than 12 h. Hence, in these cases, we employ the algebraic approach discussed in Section III-C. First, the surrogate model is applied to achieve a reduction in the number of integer variables (lightpaths). Then, the integer program is solved by considering only such a subset of variables: the best solutions found by CPLEX in 12 h of computation are listed in the third column of Tables III and IV. In the same tables, the second column reports the optimal fractional solution to the surrogate problem, which is, in any case, a lower bound to the optimal solution.
The advantages of the proposed strategy are described by showing the values of the average and the maximum congestion on the different network elements versus the traffic volume, and then comparing them with the single-layer strategy. Performance parameters have been chosen in order to analyze the ability of the algorithm in balancing the load and avoiding the bottlenecks. Each value in any curve is averaged among ten different traffic patterns so that the pattern effect can be neglected. The single-layer approach is implemented employing the same algorithm of the multilayer strategy with a slight modification: every time there is a routing on the logical topology, the method allows each LSP to cross only one lightpath. With such a modification, the LSPs are groomed only at the edge nodes and routed completely in the optical domain. Fig. 12 shows the advantage of using the flexible multilayer strategy instead of the single-layer one. In fact, the average percentage of used wavelengths on optical links is always lower by an amount of 10% of the total link capacity in the multilayer case compared with the single-layer one leading to an incremental gain of optical link congestion of about 30%. In Fig. 13, the maximum congestion ratio for the optical links is reported. It can be seen that the benefit of the multilayer strategy is more evident for low values of traffic volume where the maximum congestion is lower by 21% with respect to the single-layer case. The gap is reduced to 15% for the highest simulated value of traffic volume. This states that even though the algorithm goal was to minimize the average congestion, it still behaves quite well even in case of bottlenecks. Figs. 14 and 15 depict the same performance parameters as the ones presented in Figs. 12 and 13, respectively, but referring to the intraoffice ports (the optical ports that connect the generic OXC with an LSR). It can be noticed that the multilayer strategy offers congestion values always smaller than the values achieved in a single-layer approach: from 15% to 13% for the average congestion (Fig. 14) and from 11% to 6% for the maximum congestion (Fig. 15). The incremental gains are 25% and 18%, respectively. The obtained results are not obvious, since a more extensive use of the intraoffice ports could be expected in the multilayer strategy. In fact, even if the multilayer scheme allows processing in intermediate LSRs and consequently favors the occupation of the intraoffice ports, the number of established lightpaths in the
SABELLA et al.: MULTILAYER SOLUTION FOR PATH PROVISIONING IN NEW-GENERATION OPTICAL/MPLS NETWORKS
1151
Fig. 10. Network topologies: (a) five-node topology with an average number of ten wavelengths per link and per intraoffice port; (b) six-node topology with an average number of wavelengths per link and per intraoffice port equal to 15; and (c) eight-node topology with an average number of wavelengths per link and per intraoffice port equal to 25.
Fig. 11.
NSF network. TABLE II ALGORITHM SOLUTIONS FOR THE FIVE-NODE TOPOLOGY WITH A TRAFFIC VOLUME EQUAL TO 15 Gb/s
multilayer case is, as a whole, lower than the number of established lightpaths in the single-layer approach. Actually, in the single-layer strategy, all of the LSPs with the same source destination are first groomed into a wavelength, and then this
wavelength is routed; whereas in the multilayer approach, the grooming and the routing are performed simultaneously. This leads to a configuration where the lightpaths are almost completely filled with the LSPs’ bandwidth, as can be observed in
1152
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
TABLE III ALGORITHM SOLUTIONS FOR THE SIX-NODE TOPOLOGY WITH A TRAFFIC VOLUME EQUAL TO 30 Gb/s
TABLE IV ALGORITHM SOLUTIONS FOR THE EIGHTH-NODE TOPOLOGY WITH A TRAFFIC VOLUME EQUAL TO 60 Gb/s
Fig. 12.
Average congestion on the optical links for the NSF Network.
Fig. 16. The graphic depicts the ratio between the used and the available bandwidth within a lightpath averaged among all the lightpaths constituting the logical topology (average lightpath filling) versus the traffic volume. It is important to notice that while in the single-layer approach, the lightpath filling is always between 34% and 52%, whereas with the multilayer, the lightpaths are filled between 83% and 91%.
V. DISCUSSION AND CONCLUSION This paper proposes an offline path provisioning strategy in a two-layer GMPLS network based on a multilayer approach. In particular, the strategy provides the selection of the lightpaths that constitute the logical topology, the routes of the lightpaths on the physical network and, concurrently, the routes of the LSPs on the logical topology.
SABELLA et al.: MULTILAYER SOLUTION FOR PATH PROVISIONING IN NEW-GENERATION OPTICAL/MPLS NETWORKS
Fig. 13.
Maximum congestion on the optical links for the NSF Network.
Fig. 14.
Average congestion on the intraoffice ports for the NSF Network.
Fig. 15.
Maximum congestion on the intraoffice ports for the NSF Network.
1153
1154
Fig. 16.
JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 5, MAY 2003
Lightapth filling percentage.
In order to implement such a strategy, a new HA has been proposed. The goodness of this algorithm has been tested by means of a comparison with an algebraic algorithm under different conditions. The performance of the multilayer strategy has been evaluated in a relevant case study, quantifying the advantages with respect to a single-layer scheme in terms of minimization of network congestion. The results show that the multilayer strategy is able to optimize the use of the network resources by exploiting the lightpath bandwidth better than the single-layer approach. In fact, the number of the selected lightpaths is lower, and the average filling percentage is higher. A more efficient routing of the commodities on the lightpaths may facilitate the accommodation of new requests by means of online dynamic routing tools. A combined use of both the proposed provisioning strategy and the dynamic routing approach will be the natural evolution of this work.
ACKNOWLEDGMENT The authors wish to sincerely thank Dr. S. Canale and Dr. A. Pacifici for fruitful discussions.
REFERENCES [1] T. W. Chung et al., “Architectural and engineering issues for building an optical Internet,” Draft, 1998. [2] M. Listanti, V. Eramo, and R. Sabella, “Architectural and technological issues for future optical internet networks,” IEEE Commun. Mag., pp. 82–92, Sept. 2000. [3] B. Rajagopalan, D. Pendarakis, D. Saha, R. S. Ramamurthy, and K. Bala, “IP over optical networks: Architectural aspects,” IEEE Commun. Mag., pp. 94–102, Sept. 2000. [4] J. Ash, “Traffic engineering & QoS methods for IP-, ATM-, & TDMbased multiservice networks,” IETF Draft, Oct. 2001. [5] L. Cheng et al., “A framework for Internet network engineering,” IETF Draft, July 2001.
[6] E. Rosen, A. Viswanathan, and R. Callon. (1999, Aug.) Multiprotocol label switching architecture. Internet Draft [7] D. Awduche et al., “Requirements for Traffic Engineering Over MPLS,”, RFC2702, 1999. [8] A. Banerjee, J. Drake, J. P. Lang, B. Turner, K. Kompella, and Y. Rekhter, “Generalized multiprotocol label switching: An overview of routing and management enhancements,” IEEE Commun. Mag., Jan. 2001. [9] P. Ashwood-Smith et al.. (2002, Mar.) Generalized multi-protocol label switching (GMPLS) architecture. Internet Draft [10] G. Conte, P. Iovanna, R. Sabella, M. Settembre, and L. Valentini, “A traffic engineering solution for GMPLS networks: A hybrid approach based on off-line and on-line routing methods,” in ONDM 2003, Budapest, Feb. 2003. [11] P. Iovanna, R. Sabella, and M. Settembre, “A traffic engineering system for multi-layer networks based on the GMPLS paradigm,” IEEE Network, pp. 28–37, Mar./Apr. 2003. [12] M. Alanyali and E. Ayanoglu, “Provisioning algorithms for WDM optical networks,” IEEE/ACM Trans. Networking, vol. 7, pp. 767–778, Oct. 1999. [13] J. F. P. Labourdette, G. W. Hart, and A. S. Acampora, “Branch-exchange sequences for reconfiguration of lightwave networks,” IEEE Trans. Commun., vol. 42, pp. 2822–2832, Oct. 1994. [14] N. Nagatsu and K. Sato, “Optical path accommodation design enabling cross-connect system scale evaluation,” IEICE Trans. Commun., vol. E78-B, no. 9, pp. 1339–1343, Sept. 1995. [15] B. Mukherjee, D. Banerjee, S. Ramamurthy, and A. Mukherjee, “Some principle of design a wide-area WDM optical network,” IEEE/ACM Trans. Networking, vol. 4, pp. 684–696, May 1996. [16] R. Ramaswami and K. N. Sivarajan, “Design of logical topologies for wavelength-routed optical networks,” IEEE J-SAC Commun., vol. 14, pp. 840–851, June 1996. [17] D. Banerjee and B. Mukherjee, “Wavelength-routed optical networks: Linear formulation, resource budgeting tradeoffs, and a reconfiguration study,” in Proc. Conf. Computer Communications (IEEE INFOCOM), Kobe, Japan, Apr. 1997, p. 269. [18] T. Cinkler, D. Marx, C. P. Larsen, and D. Fogaras, “Heuristic algorithms for joint configuration of the optical and electrical layer in multi-hop wavelength routing networks,” in Proc. IEEE INFOCOM 2000, 2000, pp. 1000–1009. [19] D. Bienstock and E. O. Raskina, “Asymptotic Analysis of the Flow Deviation Method for the Maximum Current Flow Problem,” Columbia University, New York, NY, CORC Report 2000–02, 2000. [20] D. Bienstock and O. Günlük, “Computational experience with a difficult mixed-integer multicommodity flow problem,” Math. Programming, vol. 68, 1995. [21] [Online]. Available: http://www.ilog.com/products/cplex/
SABELLA et al.: MULTILAYER SOLUTION FOR PATH PROVISIONING IN NEW-GENERATION OPTICAL/MPLS NETWORKS
1155
Roberto Sabella (M’92–SM’02) was born on December 10, 1962 in Rome, Italy. He received the degree in electronic engineering (Laurea in Ingegneria elettronica) in 1987. He then joined Ericsson, Rome, Italy, where he was involved first in hardware design and subsequently in research activities on advanced fiber-optic communication systems. His research interests have covered the fields of optical device technology, high-speed optical communication systems, wavelength-division-multiplexing (WDM) optical networks, and new-generation Internet networks. In May 1997, he became the Technical Coordinator of the research consortium named CoRiTeL. Since 1999, he has been the Manager of the Research Unit in Ericsson Lab Italy, Rome, Italy. He holds five patents relating to optical networks and traffic engineering strategies, is coauthor of a book on high-speed optical communications, and author and coauthor of approximately 100 papers in either international scientific/technical journals or international conferences. He has been Lecturer (professore a contratto) at the Università degli Studi di Roma “Tor Vergata,” Rome, Italy, and at the Polytechnic of Bari, Bari, Italy. Currently, he is Adjunct Professor of design of telecommunications equipment at the Università degli Studi di Roma “La Sapienza,” Rome, Italy. He is a member of the editorial board of the journal Photonic Network Communications (New York: Kluwer Academic) for which he was twice Guest Editor for special issues on “WDM Transport Networks” and “Routing and Failure Restoration Strategies in Optical Networks.” He was also Guest Editor of a special issue on optical networks for the journal Computer Network (Elsevier). Mr. Sabella has been Guest Editor twice for IEEE magazines, for the special issue “Optical Networking Solutions for Next Generation Internet Networks”in IEEE COMMUNICATION MAGAZINE, and for the special issue “Traffic Engineering in Optical Networks” published in IEEE NETWORK.
Gianpaolo Oriolo received the Dr.Ing. degree in electrical engineering and the Ph.D. degree in operations research from the Università degli Studi di Roma “La Sapienza,” Rome, Italy. In the past, he has been a Visiting Researcher at the Mathematical Sciences Research Center, Bell Laboratories, Murray Hill, NJ; a Research Officer at the Centre for Discrete and Applicable Mathematics, London School of Economics, London, U.K.; and a Visiting Researcher at the Department of Mathematics, Ecole Polytechnique Fédérale de Lausanne—EPFL. Currently, he is a Research Associate in operations research at the Università degli Studi di Roma “Tor Vergata.” His research interest ranges from the theory of combinatorial optimization, above all polyhedral combinatorics and graph theory, to the design, analysis, and experimental evaluation of exact and approximate algorithms for the solution of optimization problems on networks.
Marina Settembre was born in Rome, Italy, on July 6, 1960. She received the Laurea degree in physics from the Università degli Studi di Roma “La Sapienza, ” Rome, Italy, in 1985. In 1985, she was granted a fellowship at the Fondazione Ugo Bordoni, working on innovative materials for optical devices. From 1986 to September 2000, she worked as a Research Scientist in the Optical Communication Division of the Fondazione Ugo Bordoni, where she focused first on optical devices for signal routing/processing and then on physical layer modeling and on high-capacity optical transmission systems. Since October 2000, she has joined Ericsson Lab Italy, Rome, Italy, working in the research department on network architectures, traffic engineering strategies, and related algorithms and control plane definition for new-generation networks based on the GMPLS paradigm. She has been actively involved in several European COST, ACTS, and IST projects (ACTS-Upgrade, ACTS- Esther, Cost 245, Cost 266, IST Atlas, IST Embrace). She published more than 90 papers in international scientific journals and conferences and a book titled Non-linear Optical Communication Networks New York: Wiley, 1998). Her present research interests are in the area of broad-band wireless access technologies and architectures, focusing on traffic management, Media Access Control Protocols, and Quality of Service for packet-based communication in wireless systems.
the GMPLS paradigm.
Federica Razza was born on February 25, 1977, in Rome, Italy. She received the B.Sc. degree in industrial engineering and the M.Sc. degree in industrial engineering from the University “Tor Vergata” in 2001 and 2002, respectively. Her thesis discussed an euristic alghorithm for the optimization of traffic engineering strategies for routing in the TLC networks of new generation. Her present research interests include operations research techniques and related algorithms and network design applied to new-generation networks.
Fabio Ferlito was born on May 11, 1975 in Catania, Italy. He received the B.Sc. and M.Sc. degrees, both in industrial engineering, from the University “Tor Vergata,” Rome, Italy, in 2001 and 2002, respectively. His B.Sc. thesis was on multi-objective optimization, and his M.Sc. thesis was on MILP formulations for routing in new-generation networks. His present research interests include traffic engineering strategies and related algorithms, network design and optimization, and operations research techniques applied to new-generation networks based on
Giulia Conte was born on December 26, 1976 in Rome, Italy. She received the degree in telecommunications engineering from the University “Tor Vergata,” Rome, Italy, in January 2001, with a thesis on routing and failure restoration strategies in the presence of different traffic classes in wavelength-division-multiplexing (WDM) networks. In February 2001, she joined CoRiTeL, Rome, Italy, where she took part in the research activity of the Optical Networks Group. Within this project, she was mainly interested in the system analysis of the routing and protection schemes for the new generation of optical networks. She has been involved with the ASTERIX project to implement and validate hybrid routing strategies and signaling protocol extension for IP/MPLS next-generation networks. Her present research interests include traffic engineering strategies and related algorithms, network architectures, and control plane definition for new-generation networks based on the GMPLS paradigm. She also works on the TANGO project, where she is contributing to the definition of an optimized IP-based control plane development for next-generation IP/MPLS networks.
