You are on page 1of 5

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 14, NO.

5 , JUNE 1996

909

Lightpath (Wavelength) Routing in Large WDM Networks


U

Imrich Chlamtac, Fellow, IEEE, AndrAs Farag6, and Tao Zhang, Member, IEEE

Abstract-We address the problem of efficient circuit switching in wide area optical networks. The solution provided is based on finding optimal routes for lightpaths and the new concept of semilightpaths. A lightpath is a fully optical transmission path, while a semilightpath is a transmission path constructed by chaining together several lightpaths, using wavelength conversion at their junctions. A fast and practical algorithm is presented to optimally route lightpaths and semilightpaths taking into account both the cost of using the wavelengths on links and the cost of wavelength conversion. We prove that the running time of the algorithm is the best possible in the wide class of algorithms allowing linear algebraic operations on weights. This class encompasses all known related practical methods. Additionally, our method works for any physical realization of wavelength conversion, independently whether it is done via optoelectronic conversion or in a fully optical way.
I.

INTRODUCTION

IGHTPATHS provide a powerful approach to tap the vast available bandwidth in wavelength-division multiplexed (WDM) networks [1]-[4]. A lightpath is an all-optical transmission path between two network nodes, implemented by the allocation of the same wavelength throughout the path. Data transmitted through a lightpath requires no wavelength conversion or electronic processing at intermediate nodes. Therefore, lightpaths enable an efficient utilization of the optical bandwidth in WDM networks, reduce electronic processing cost at intermediate nodes, improve reliability and the quality of services provided to data transmitted through lightpaths. While transmitting all traffic between every pair of nodes over lightpaths is desirable, it is not generally feasible to establish lightpaths between every pair of nodes and accommodate all the traffic by lightpaths, due to physical constraints such as limited number of wavelengths, limited number and tunability of optical transceivers at each node, as well as lightwave dispersions that limit the physical length of a lightpath. Additionally, given the network conditions, in general a single optical wavelength may not be available between a given source and destination because some of the resources are already occupied by existing lightpaths. Therefore, we introduce the concept of semilightpaths-a transmission path obtained by establishing and chaining together several lightpaths. Thus, in a semilightpath wavelength conversion is required at some intermediate nodes, but generally not at all
Manuscript received April 5, 1995; revised September 25, 1995. I. Chlanilac is with the Department of EIcCtriCal, Cvmputer and Systems Engineering, Boston University, Boston, MA 02215 USA. A. Faragd is with thc Department of Telecommunications and Telematics, Technical University of Budapest, Budapest, Hungary H-l I1 1. T. Zhang is with Bellcore, Morristown, NJ 07960 USA. Publisher Item Identifier S 0733-8716(96)03679-7.

nodes. (The lightpath is a special case of semilightpath when the number of intermediate nodes on the path with wavelength conversion is zero.) The objective of this paper is to present an algorithm for optimally routing semilightpaths between a given source and destination. The figure of merit we optimize is a combination of the following two costs associated with the path: 1) the cost for traversing a link on some wavelength; and 2) the cost for wavelength conversion when the path has to switch to a different wavelength at some intermediate nodes. Note that if only the first cost factor is present, then the problem reduces to the well solved task of finding a shortest path. The additional cost for wavelength conversion, however, makes the direct application of shortelst path algorithms to the network graph inappropriate. In the proposed solution, we first transform the network graph into an auxiliary graph, called the wavelength graph (WG), then we find a shortest path in the wavelength graph. The found shortest path corresponds to the optimal semilightpath in a natural way. The principle of this transformation is straightforward and provides a practical solution. It is not trivial, however, to find solutions faster than simply applying general shortest path algorithms. In this paper we show that this is possible via making use of the special structure of the WG. We prove that our algorithm is optimal for the considered problem with respect to running time, and, therefore, it provides the best possible lightpath routing solution, in the class of algorithms where only linear algebraic operations are allowed on the weights. Note that all known practical shortest path algorithms belong to this class. In this paper, we concentrate on the algorithmic issues raised by the optimal routing of semilightpaths, discussing and optimally solving the problem in a graph theoretic network model. The different network implementational issues of lightpaths have been discussed in a number of earlier papers, and are also the subject of at least two testbed efforts referenced in [SI, [9], and, therefore, will not be discussed here. In our solution we assume that the cost structure is already known from physical considerations and is given as input to the algorithm. Note that the physical realization of wavlelength conversion, that is, whether it is done via opto-electronic conversion or in a fully optical way, is irrelevant from the viewpoint of path-cost minimization, once the associated co;sts have been specified.
11. NETWORK MODELAND PROBLEM FORMULATION

The network topology is modeled by a directed graph G = (V,E ) , where V stands for the set of network nodes

0733-8716/96$05.00 0 1996 IEEE

910

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 14, NO. 5, JUNE 1996

(vertices in the graph) and E stands for the set of directed links (edges of the graph). Note that the undirected case can be modeled by replacing r;n undirected edge with two oppositely directed edges. We assume that a set A = {XI, . . . , A,} of wavelengths is given in the network. The cost structure of using the resources is represented as follows. For each link e and wavelength X i a weight w(e, Xi) is given, representing the cost of using wavelength X i on link e. If X i is not available on the link then the weight is infinite (or a very large number for practical number representation). It is assumed that all weights are nonnegative. The cost of wavelength conversion is modeled via cost factors of the form (A,, A,). This is the cost of wavelength , conversion at node i from wavelength A to A,. If for certain values of i , p , q the conversion is not available then e, (A, A,) is infinite (or a very large number). If the two wavelengths are equal, then the cost is zero, that is, ci(X,, A = 0. The ), above-defined wavelength conversion costs accommodate the general case where conversion costs depend on the nodes and wavelengths involved, a case often encountered in practical networks. For example, wavelength conversion nodes in critical locations in a network (e.g., nodes used to interconnect different networks) should be used only when necessary and therefore should be assigned higher conversion costs. When wavelengths are grouped into wavebands [lo], conversion between wavelengths in the same waveband is generally less costly than conversion between wavelengths in different wavebands. The conversion costs can also be used to enforce certain rules on how wavelengths should be assigned. For example, the costs can be easily assigned in such a way that the least numbered available wavelengths will be used first. Furthermore, conversion costs can be dynamically adjusted to adapt to changing network conditions. For example, when the traffic going through a wavelength conversion node is getting closer to the nodes capacity, wavelength conversion cost at this node can be increased so that new semilightpaths can be routed through other nodes. A semilightpath is defined as a sequence e l , . . , e, of directed links, such that the start-point of e;+l coincides with the endpoint of e i , i = 1, . . . , T - 1. Further, a wavelength X ( i ) E A is associated with each e;, this is the wavelength used by the semilightpath on link e;. The cost C(P) a of semilightpath P is defined as follows. Denote by w(e) the endpoint of link e . The cost of P is

Let us remark that it is not forbidden in general for a semilightpath to visit a node more than once, on different wavelengths. It is expected, however, that in practical cases the cost structure will exclude such cases from being optimal, but we do not have to exclude them apriori. The case, however, when the same link is traversed on the same wavelength more than once is automatically excluded from the potential optimal solutions, since it can be shortened by an obvious shortcut. (This is why we do not have to exclude it in the definition.)

111. FIRST PHASE: PROBLEM TRANSFORMATION


We first show the principle of the solution. In its first phase, a simple and practical algorithm is obtained by constructing an auxiliary graph, called the WG of the network. The WG is constructed such that a minimum weight path in the WG will correspond to a minimum cost semilightpath. In the next section, we will show how to implement the shortest path search in the WG such that the special structure is utilized and an algorithm of optimal speed is obtained. The wavelength graph WG is defined as follows. 1) Take N = k n vertices, where k is the number of wavelengths and n is the number of nodes in the original network. 2) Arrange the vertices in a matrix-like (or grid-like) structure with k rows and n columns. Each column corresponds to a node of the network and each row corresponds to a wavelength. 3) In the ith row, i = 1, . . , k , draw a directed edge from column j to column h whenever there exists a link e from node j to h in the network and the wavelength X i is available on this link. Assign the weight w(e, Xi) to this edge. The subgraph induced by the ith row of vertices in the wavelength graph is called the &-plane. This represents the network with respect to wavelength Xi, together with the costs associated with X i . 4) In column j , j = 1, . . , n, draw a directed edge from row i to 1, if at node j wavelength conversion is available from X i to Al. Assign the weight cj(X;, A,) to this edge. The subgraph induced by the column represents the possible wavelength conversions at node 2 , together with their costs. The construction of the wavelength graph yields a straightforward 1-1 mapping of semilightpaths into the WG. If we move along a semilightpath from node s to node t , then each time we traverse a link on wavelength A,, this corresponds to moving along a horizontal edge in the &-plane of the WG. Similarly, each time we do a wavelength conversion between A, and Ah at a node j , this corresponds to moving along a vertical edge in column j of the WG, from the Xi-plane to the Ah-plane. From the above mapping it is clear that a minimum cost lightpath from node s to t corresponds to a minimum weight path in the WG from column s to column t . Thus, the problem is reduced to finding a shortest (= minimum weight) path in the WG between two columns. This can be further reduced to finding a shortest path between two vertices, if we add two new vertices s, t to the WG: s is connected to each vertex

Here the first summation gives the cost of traversing the links and the second sum gives the cost of wavelength conversion at nodes. Note that only those summands cmtribute to the second sum where X(7) # X(z 1), otherwise the conversion cost is 0, as defined above. Now the task is the following. Given two network nodes s and t. find a semilightpath P that connects them, such that C ( P ) is minimum.

CHLAMTAC ef al.: LIGHTPATH (WAVELENGTH) ROUTING IN LARGE WDM NETWORKS

911

Step 2 (Designation of a New Permanent Label): 1) Find the minimum of Ri, Cj (Yi; j ) . 2) Find an h E T with minimum uh in the row or Thus, viewing the problem from this high level, we have column which gave the minimum above (ties are broken found a practical solution, since the minimum cost semiarbitrarily). lightpath problem is reduced via a relatively straightforward 3) T := T - { h } ; P := P U { h } transformation to a usual shortest path problem. In the next 4) If T = 0 then STOP. section we show how to make use of the special structure of Step 3 (Updating Row and Column Minimum): the WG to obtain the solution faster than by simply using a 1) If h , found in Step 2, is in row i and column j , standard shortest path algorithm. 2) then R; := min {uj : j is in the ith row, j E T } 3) Cj := min{u; : i is in the j t h , i E T } I v . SECOND PHASE: MAKING USE OF THE 4) (The minimum over an empty set is taken to be 00). SPECIAL STRUCTURE TO OBTAIN OPTIMAL SPEED Step 4 (Revision of Tentative Labeh): If the shortest path computation is done in the WG by one of 1) If h, found in Step 2, is in row i and column j , the fastest known practical algorithms, such as Dijkstras [5], then, for all 1 E T in row i and in column j , set then we can achieve a running time of O ( N 2 )= 0 ( k 2 n 2 ) . uz := min{ul, uz a ~ } Go to Step 2. . Now we show that the special structure of the WG allows an Algorithm SPAWG, just as Dijkstras algorithm, computes implementation of Dijkstras algorithm such that it yields a the minimum weight distance from the root 1 to all other lower complexity 0 [ ( k n )kn] , which is, in fact, proven to vertices. With an obvious modification the path itself can be be the best possible achievable complexity for the minimum found by keeping record of the minimizing elements, as it cost semilightpath problem, in the so called algebraic decision is usually done in shortest path algorithms. Using the above tree class of algorithms where only linear algebraic operations algorithm, the minimum cost semilightpath (MCSLP) problem are allowed on the weights. is solved as follows. We next present the algorithm for finding a shortest path between two given vertices of the WG. Recall that in the previous section the problem of finding a shortest path between B. Algorithm MCSLP two columns was reduced to this case. Note that the two 1) Construct the wavelength graph of the network. additional vertices, used for this reduction, do not disrupt the 2) Find a shortest path between the source and terminator structure of the WG, as they can be viewed as members of columns using the SPAWG algorithm. an additional (degenerated) wavelength plane that contains no 3) Map the found path back to the network to obtain the edges. minimum cost semilightpath. For easy notation we label the vertices of the WG simply by Before analyzing the solution, we first specify precisely the numbers 1, . . , N and the weight of edge ( i ,j ) is denoted (If by uZ3. the edge does not exists then it has infinite weight.) that in which sense we are going tio prove the optimality If two vertices, say i and 3 , are either in the same column of the running time. This is needed because the complexity or in the same row, this relation is denoted by i N 1. In the of a shortest path algorithm may depend on what algebraic algorithm we use three sets of numerical variables: a variable operations are allowed on the numerical weights, regarding the algebraic operations as elementary steps of the algorithm. That U , is associated to each vertex; a variable R, is associated to is, an algorithm may be optimal for a certain set of allowed each row and a variable C, is associated to each column. The R, , C, variables will represent the row and column minimum algebraic operations, while not optimal for another class. An important and practical class is the so called algebraic values, respectively. For simplicity, we assume that the startpoint of the searched path is labeled by one. We also use two decision tree model of computation (for the shortest path set-valued variables P and T for permanently and tentatively problem, see [6]). This can be characterized by allowing labeled vertices, respectively (permanently labeled vertices are only linear algebraic operations on the weights. More exactly, arbitrary sums of edge weights can be calculated and compared those for which the computation has already terminated, that to each other (but no nonlinear operations are allowed on is, the variable U% contains the weight of a shortest path from the weights, such as taking squares or finding the roots of 1 to i . For the tentatively labeled vertices the computation is polynomials, etc.). Although, in theory, more general models still going on). Now we can present the special shortest path are possible with sophisticated nonlinear operations on the algorithm for the WG (SPAWG). weights, nevertheless, it is a very important fact that all known practical algorithms for finding the shortest path between two A. Algorithm SPAWG specified vertices are contained in {he class described by Step I (Initialization): algebraic decision trees. Therefore, if we restrict ourselves to this class we do not impose any essential restriction from 1) U 1 := 0; the practical point of view, but it allows us to prove optimal 2) If .i 1 then U, := a l j else u := 00 ( V i ) i running time, as stated below. 3) R; := min {uj : j is in the ith row, j # l}, ( V i ) Theorem: For any network of n nodes with IC wavelengths 4) Cj := min{ui : i is in the jth column, i # I}, ( V j ) Algorithm MCSLP finds an optimal semilightpath in time 5 ) P := {l}; := ( 2 , , N } . T

in column s via zero-weight edges and each vertex in column

t is connected to t again via zero-weight edges. Then, clearly, the shortest path from s to t provides us with a solution.

912

IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 14, NO. 5. JUNE 1996

O [ ( n k ) n k ] . This order of magnitude for the worst case running time is the best possible in the algebraic decision tree model of computation. Proof: See the Appendix. Remark: As mentioned above, the algebraic decision tree model contains all known practical shortest path algorithms. This gives a strong evidence for expecting that in general no practical solution is possible for the lightpath problem with lower order of magnitude for the worst case running time, as a function of n and k . For example, if we run Dijkstras algorithm on the WG, this yields a complexity of O ( N 2 ) . where N = kn.On the other hand, if e.g., k M n,then our method yields O[(N1l2 N 1 / 2 ) N ] O(N). =

V. CONCLUSION An algorithm was presented to optimally solve the problem of fast routing of lightpaths and semilightpaths in wide area fiber optic networks. The obtained lightpath or semilightpath minimizes an overall cost function that accounts both for using the wavelengths on the links and for doing wavelength conversion at nodes when necessary. We have proven that the running time of the algorithm is the best possible with respect to the order of magnitude in the class of algorithms where all known related practical algorithms belong. The simplicity and fast running time of the algorithm make it a good candidate for efficient practical implementation. Given the many theoretical as well as practical efforts for constructing optical WDM networks, this result can be readily implemented in the emerging networks. APPENDIX
Proof of the Theorem: Algorithm SPAWG differs from Dijkstras algorithm (see e.g., [ 5 ] ) in that (1) the search for a new permanent label is restricted to a row or column, and (2) the revision of tentative labels is restricted to the row and column where the new permanently labeled vertex has been found. Since the new permanent label is searched for in the row or column with minimum R or C value and these values represent the minimum U value in the respective subsets, therefore, we find the same vertex for a new permanent label, as if the search were done in the whole graph. The row and column minimum update in Step 3 guarantees that this situation is correctly maintained. The revision of tentative labels, although restricted to the row and column of the new permanent label, also results in the same revision as if it were done in the whole graph, since, by the structure of the WG, the rest of vertices are not connected to the new permanently labeled vertex. Thus, despite the restricted search domains, the algorithm results in the same labels as Dijkstras algorithm. This implies that the shortest path is found correctly. Concerning the complexity of Algorithm SPAWG we can observe that all the searches in Steps 2, 3, and 4 are restricted to O ( n k ) elements. This is repeated N - 1 = n k - 1 times, since in each iteration a new vertex is permanently labeled. The initialization takes O ( N ) = O ( n k ) time. Thus, the overall complexity of Algorithm SPAWG is 0[ (n k)nk]. The

auxiliary steps in Algorithm MCSLP are clearly dominated by this order of magnitude, yielding the same overall complexity for MCSLP. For the proof of optimal running time it is enough to restrict ourselves to the case of k 5 n, for the case k > n works the same way, by interchanging the role of k and 71. For k 5 n we have O ( n k ) = O ( n ) ,therefore, the complexity of our algorithm is O ( k n 2 ) in this case. For simplicity, we also assume that k is an odd number, this does not change the order of magnitude. The idea of the proof of optimal running time is to show that if we have a lightpath algorithm then it can be used to find shortest paths in k independently chosen n-vertex graphs simultaneously. Then applying a known lower bound for the shortest path problem in the algebraic decision tree model, the result will follow. The detailed construction is as follows. Take k arbitrary connected graphs G I , . . . , GI, with weighted edges on the same set of n vertices. Represent each graph by a wavelength plane and the weight of a link with respect to wavelength i is set to the weight of the corresponding edge of Gi. Now select two vertices s and t. Set the cost of wavelength conversion such that it is 0 at node s from wavelength i to i 1 for even values of i and it is 0 at node t from wavelength i to i 1 for odd values of i. All other conversion costs in the network are very high (infinite). Add two new nodes s; t to the network. Connect s to s and t to t. such that on the link (s; s) only wavelength 1 is available, while on the link (t;t ) only wavelength k is available. Now assume we have an algorithm for the lightpath problem. Run this on the above constructed network such that we search for a semilightpath from s to t. The above construction implies that the only optimal solution must have the following structure. Go first from s to s on wavelength 1. Then follow a shortest path from s to t , using only wavelength 1. Then make a wavelength conversion at t to wavelength 2. Then follow a shortest path back to s, using only waveleingth 2. Then make a wavelength conversion at s to wavelength 3. Then follow again a shortest path back to t , using only wavelength 3, and so forth. Finally, we arrive to t on wavelength k , then we can move to t on the link ( t ,t),using wavelength k . The above semilightpath gives a shortest path from s to t in each of the k independently chosen Gi graphs. Assume the lightpath algorithm used C time complexity to find the solution. Let S(n) be best possible complexity of finding shortest paths between a given pair of vertices in a-vertex graphs. Since we have found the paths in k independent graphs, therefore, C 2 k S ( n ) must hold. Now we use the known result ([6], [7]) which says that S ( n ) cannot be smaller than O ( n 2 )for the: algebraic decision tree model of computation (which includes all known practical shortest path algorithms). The lower bound implies that C is at least O(h2).the other hand, as seen above, our algorithm On really achieves this lower bound, which proves the optimal running time.

ACKNOWLEDGMENT The authors are grateful for the valuable comments given by the anonymous reviewers and the editor R. L. Cruz.

CHLAMTAC et ul.: LIGHTPATH (WAVELENGTH) ROUTING IN LARGE WDM NETWORKS

913

REFERENCES
111 I. Chlamtac, A. Ganz, and G. Karmi, Lightpath communications: A novel approach to high bandwidth optical WANS, /EEE Trans. Commun., July 1992. J:F. P. Labourdette and A. S. Acampora, A multichannel multihop local lightwave network, in GLOBECOM87, Tokyo, Japan, Nov. 1987. [31 A. S. Acampora, Partially reconfigurahle multihop lightwave networks, in GLOBECOM90, San Diego, Dec. 1990. . [41 1 Chlamtac, A. Ganz, and G. Karmi, Lightpath routing in the lightnet architecture, in 2nd IFIP WG6.lhVG6.4 International Workshop on Protocols .for High-speed Networks, Stanford, CA, Nov. 1990. E. L. Lawler, Combinatorial Optimization: Networks and Matroids. New York: Holt, Rinehart and Winston, 1976. D. R. Karger, D. Koller, and S. J. Phillips, Finding the hidden path: time hounds for all-pairs shortest paths, in IEEE 32nd Annu. Symp. Foundations Computer Sci., 1991, pp. 560-568. [71 P. M. Spira and A. Pan, On finding and updating shortest paths and spanning trees, in IEEE 14th Annu. Symp. Switching Automata, 1973. [8] S. B. Alexander et al., A precompetitive consortium on wide-hand alloptical networks, J. Lightwave Technol., vol. 1 1 , no. 5/6, pp. 714-735, May/June 1993. [9] C. B. Brackett et al., A scaleable multiwavelength multihop optical network: A proposal for research on all-optical networks, J. Lightwuve Technol., vol. 11, no. S/6, pp. 736-753, May/June 1993. [ 101 K. Bala, T. E. Stern, D. Simchi-Levi, and K. Bala, Routing in linear lightwave network, /EEE/ACM Trans. Networking, vol. 3, no. 4, Aug. 1995.

Andrhs Farag6 received the M.Sc. and Ph.D. degrees from the Technical University of Budapest, Hungary, both in electrical engineering, in 1976 and 1981, respectively. He is an Associate Professor at the Department of Telecommunications and Telematics, Technical University of Budapest. In the Department, he is the Research Director of the High Speed Networks Laboratory based on a joint research project between the Technical University of Budapest and Ericsson. He worked as a Visitinr: Senior Research Fellow at the University of Massachusetts at Amherst in 1491-1994. His main interest is in developing algorithms and applying mathematical methods for the design and optimization of various types of telecommunication networks and services. He has published more than 80 papers in the field. Dr. Farag6 is a member of the IFIP Working Group 6.3 Performance on Communication Systems, member of ACM, and a founding member of the Hungarian Chapter of ACM.

Imrich Chlamtac (M86-SM86-F93) received the Ph.D. degree from the University of Minnesota, Minneapolis. He is the President of BCN, a company dealing with research and development of wireless products and network systems integration in the United States and Europe. He is also a Professor of Electrical Engineering at Boston University, MA, where he leads the Advanced Communications Technologies (ACT) Laboratory. His publications include more than 140 papers in referred journals and conferences, multiple book chapters and books, including the first textbook on LANs, Local Networks: Motivation, Technology and Pe~ormunce (Lexington Books, 1981). He is the founding Editor in Chief of the ACM/Bultzer Wireless Networks Journul, and has served on the editorial hoard of IEEE TRANSACTIONS COMMUNICATIONS numerous other ON and journals. In the past, he chaired several ACM and IEEE workshops and conferences and is the founder of the ACM MohiCom Conference. Dr. Chlamtac bas been an IEEE Distinguished Lecturer and a Fulbright Scholar.

Tao Zhang (S92-M93) received the B S degree in electrical engineenng and the M S degree in computer engineering, hoth from Northern Jiaotong University. Beijing. PROC, in 1984 in 1987. respectively. He received the Ph D. degree in computer engineenng for research in ATM and WDM fiber optic networks from the University of Massachusetts, Amherst, MA in 1993 From 1987 to 1989, he was a faculty member of the Department of Telecommunications, Northern Jiaotong University, Beijing, PROC, where he initiated and lead a research project on integrated datdvoice wireless local area networks. From 1993 to 1995, he was with L.ehman Brothers as a Senior Communications Engineer and later with Citibanlc, N A as an Assistant Vice President, where his work was focuyed on corporate network planning and design as well as building global trading systems He is a Research Scientist at Bellcore, where he works in the Network Architecture and Analysis Research Department. His current research interests include ATM networks, WDM fiber optic networks, network architectures for multimedia applications, and network optimization algorithms

You might also like