On Shortest Single/Multiple Path Computation Problems in Fiber-Wireless (FiWi) Access Networks

Transcription

1 2014 IEEE 15th International Conference on High Performance Sitching and Routing (HPSR) On Shortest Single/Multiple Path Computation Problems in Fiber-Wireless (FiWi) Access Netorks Chenyang Zhou, Anisha Mazumder, Arunabha Sen, Martin Reisslein and Andrea Richa School of Computing, Informatics and Decision Systems Engineering School of Electrical, Computer, and Energy Engineering Arizona State Uniersity {czhou24, anisha.mazumder, asen, reisslein, Abstract Fiber-Wireless (FiWi) netorks hae receied considerable attention in the research community in the last fe years as they offer an attractie ay of integrating optical and ireless technology. As in eery other type of netorks, routing plays a major role in FiWi netorks. Accordingly, a number of routing algorithms for FiWi netorks hae been proposed. Most of the routing algorithms attempt to find the shortest path from the source to the destination. A recent paper proposed a noel path length metric, here the contribution of a link toards path length computation depends not only on that link but also eery other link that constitutes the path from the source to the destination. In this paper e address the problem of computing the shortest path using this path length metric. Moreoer, e consider a ariation of the metric and also proide an algorithm to compute the shortest path using this ariation. As multipath routing proides a number of adantages oer single path routing, e consider disjoint path routing ith the ne path length metric. We sho that hile the single path computation problem can be soled in polynomial time in both the cases, the disjoint path computation problem is NP-complete. We proide optimal solution for the NP-complete problem using integer linear programming and also proide to approximation algorithms ith a performance bound of 4 and 2 respectiely. The experimental ealuation of the approximation algorithms produced a near optimal solution in a fraction of a second. I. INTRODUCTION Path computation problems are arguably one of the most ell studied family of problems in communication netorks. In most of these problems, one or more eight is associated ith a link representing, among other things, the cost, delay or the reliability of that link. The objectie most often is to find a least eighted path (or shortest path ) beteen a specified source-destination node pair. In most of these problems, if a link l is a part of a path P, then the contribution of the link l on the length of the path P depends only on the eight (l) of the link l, and is obliious of the eights of the links traersed before or after traersing the link l on the path P. Hoeer, in a recent paper on optical-ireless FiWi netork [5], the authors hae proposed a path length metric, here the contribution of the link l on the length of the path P depends not only on its on eight (l), but also on the eights of all the links of the path P. As the authors of [5] do not present any algorithm for computing the shortest path beteen the source-destination node pair using this ne metric, e present a polynomial time algorithm for this problem in this paper. This result is interesting because of the nature of ne metric proposed in [5], one key property on hich the shortest path algorithm due to Dijkstra is based, that is, subpath of a shortest path is shortest, is no longer alid. We sho that een ithout this key property, not only it is possible to compute the shortest path in polynomial time using the ne metric, it is also possible to compute the shortest path in polynomial time, ith a ariation of the metric proposed in [5]. The rest of the paper is organized as follos. In section III, e present the path length metric proposed for the FiWi netork in [5] and a ariation of it. In section IV e proide algorithms for computing the shortest path using these to metrics. As multi-path routing offers significant adantage oer single path routing [6], [7], [8], [9], e also consider the problem of computation of a pair of node disjoint paths beteen a source-destination node pair using the metric proposed in [5]. We sho that hile the single path computation problem can be soled in polynomial time in all these cases, the disjoint path computation problem is NP-complete. The contributions of the paper are as follos; Polynomial time algorithm for single path routing (metric 1) in FiWi netorks Polynomial time algorithm for single path routing (metric 2) in FiWi netorks NP-completeness proof of disjoint path routing (metric 1) in FiWi netorks Optimal solution for disjoint path routing (metric 1) in FiWi netorks using Integer Linear Programming One approximation algorithm for disjoint path routing in FiWi netorks ith an approximation bound of 4 and computation complexity O((n + m)log n) One approximation algorithm for disjoint path routing in FiWi netorks ith an approximation bound of 2 and computation complexity O(m(n + m)log n) Experimental ealuation results of the approximation algorithm for disjoint path routing in FiWi netorks II. RELATED WORK Fiber-Wirelss (FiWi) netorks is a hybrid access netork resulting from the conergence of optical access netorks such as Passie Optical Netorks (PONs) and ireless access netorks such as Wireless Mesh Netorks (WMNs) capable of proiding lo cost, high bandidth last mile access /14/$ IEEE 131

2 Because it proides an attractie ay of integrating optical and ireless technology, Fiber-Wireless (FiWi) netorks hae receied considerable attention in the research community in the last fe years [1], [2], [3], [4], [5], [8], [9]. The minimum interference routing algorithm for the FiWi enironment as first proposed in [4]. In this algorithm the path length as measured in terms of the number of hops in the ireless part of the FiWi netork. The rationale for this choice as that the maximum throughput of the ireless part is typically much smaller than the throughput of the optical part, and hence minimization of the ireless hop count should lead to maximizing the throughout of the FiWi netork. Hoeer, the authors of [5] noted that minimization of the ireless hop count does not alays lead to throughput maximization. Accordingly, the path length metric proposed by them in [5] pays considerable importance to the traffic intensity at a generic FiWi netork node. The results presented in this paper are motiated by the path length metric proposed in [5]. III. PROBLEM FORMULATION In the classical path problem, each edge e E of the graph G =(V,E), has a eight (e) associated ith it and if there is a path P from the node 0 to k in the graph G =(V,E) k k then the path length or the distance beteen the nodes 0 and k is gien by (P 0, k )= k Hoeer, in the path length metric proposed in [5] for optical-ireless FiWi netorks [1], [2], [3], the contribution of e i to the path length computation depends not only on the eight i, but also on the eights of the other edges that constitute the path. In the folloing section, e discuss this metric and a ariation of it. We also also formulate the multipath computation problem using this metric. The Optimized FiWi Routing Algorithm (OFRA) proposed in [5] computes the length (or eight) of a path P from 0 to k using the folloing metric ( ) (P 0, k )=min ( u )+max ( u) P u P u P here u represents the traffic intensity at a generic FiWi netork node u, hich may be an optical node in the fiber backhaul or a ireless node in ireless mesh front-end. In order to compute shortest path using this metric, in our formulation, instead of associating a traffic intensity eight ( u ) ith nodes, e associate them ith edges. This can easily be achieed by replacing the node u ith eight u ith to nodes u 1 and u 2, connecting them ith an edge (u 1,u 2 ) and assigning the eight u on this edge. In this scenario, if there is a path P from the node 0 to k in the graph G =(V,E) k k then the path length beteen the nodes 0 and k is gien by + (P 0, k ) = k + max( 1, 2,... k ) k = i + max k i=1 i i=1 In the second metric, the length a path P 0, k : k, beteen the nodes 0 and k is gien by (P 0, k ) = = k i + CNT(P 0, k ) max( 1, 2,... k ) i=1 k i + CNT(P 0, k ) max k i=1 i i=1 here CNT(P 0, k ) is the count of the number of times max ( 1, 2,... k ) appears on the path P 0, k. We study the shortest path computation problems in FiWi netorks using the aboe metrics and proide polynomial time algorithms for solution in subsections IV-A and IV-B. If max = max( 1, 2,... k ), e refer to the corresponding edge (link) as e max. If there are multiple edges haing the eight of max, e arbitrarily choose any one of them as e max. It may be noted that both the metrics hae an interesting property in that in both cases, the contribution of an edge e on the path length computation depends not only on the edge e but also on eery other edge on the path. This is so, because if the edge e happens to be e max, contribution of this edge in computation of + (P 0, k ) and (P 0, k ) ill be 2 (e) and CNT(P 0, k ) (e) respectiely. If e is not e max, then its contribution ill be (e) for both the metrics. As multipath routing proides an opportunity for higher throughput, loer delay, and better load balancing and resilience, its use hae been proposed in fiber netorks [6], ireless netorks [7] and recently in integrated fiber-ireless netorks [8], [9]. Accordingly, e study the problem of computing a pair of edge disjoint paths beteen a sourcedestination node pair s and d, such that the length of the longer path (path length computation using the first metric) is shortest among all edge disjoint path pairs beteen the nodes s and d. In subsection IV-C e proe that this problem is NP-complete, in subsection IV-D, e proide an optimal solution for the problem using integer linear programming, in subsections IV-E and IV-F e proide to approximation algorithms for the problem ith a performance bound of 4 and 2 respectiely, and in subsection IV-F e proide results of experimental ealuation of the approximation algorithms. IV. PATH PROBLEMS IN FIWI NETWORKS In this section, e present (i) to different algorithms for shortest path computation using to different metrics, (ii) NP-completeness proof for the disjoint path problem, (iii) to approximation algorithms for the disjoint path problem, and (i) experimental ealuation results of the approximation algorithms. 132

3 It may be noted that, in both metrics + (P 0, k ) and (P 0, k ), e call an edge e P 0, k crucial, if(e) = max k i=1 (e ), e P 0, k. A. Shortest Path Computation using Metric 1 It may be recalled that the path length metric used in this case is the folloing: + (P 0, k )= k i=1 i + max k i=1 i.if the path length metric as gien as (P 0, k )= k i=1 i, algorithms due to Dijkstra and Bellman-Ford could hae been used to compute the shortest path beteen a source-destination node pair. One important property of the path length metric that is exploited by Dijkstra s algorithm is that subpath of a shortest path is shortest. Hoeer, the ne path length metric k i=1 i +max k i=1 i does not hae this property. We illustrate this ith the example belo. Consider to paths P 1 and P 2 from the node 0 to 3 in the graph G =(V,E), here P 1 : and P 2 : If 1 =0.25, 2 =5, 3 =4.75, 4 = 2, 5 =4, the length of the path P 1, + (P 1 )= max( 1, 2, 3 )= max(0.25, 5, 4.75) = 15 and the length of the path P 2, + (P 2 )= max( 4, 5. 6 ) = max(2, 4, 4.75) = Although P 1 is shortest path in this scenario, the length of its subpath is max (0.25, 5) = 10.25, hich is greater than the length of a subpath of P max (2, 4) = 10, demonstrating that the assertion that subpath of a shortest path is shortest no longer holds in this path length metric. As the assertion subpath of a shortest path is shortest no longer holds in this path length metric, e cannot use the standard shortest path algorithm due to Dijkstra in this case. Hoeer, e sho that e can still compute the shortest path beteen a source-destination node pair in polynomial time by repeated application of the Dijkstra s algorithm. The algorithm is described next. For a gien graph G =(V,E),.l.o.g, e assume V = n and E = m. Define G e as subgraph of G by deleting edges hose eight is greater than (e). Also, as Dijkstra s algorithm does, e need to maintain distance ector. We define dist be distance (length of shortest path) from s to, Π be predecessor of and maxedge be eight of the crucial edge from s to ia the shortest path, ans be optimal solution (length) from s to. Different G e can be treated as different layers of the G. For any path P, e define the function e (P ) as the crucial edge along P. It is easy to obsere that if P d is the optimal path from s to node d then (P d ) = e P (e) and + (P d ) = (P d )+(e (P d )). It may be noted that henceforth, e shorten P s,d to P d, because e consider that the source is fixed hile the destination d is ariable. Lemma 1. (P d ) is minimum in G e (P d ). Proof: It is obious that P d still exists in G e (P d ), since edges on P d are not abandoned. Suppose P d is not shortest, then there must be another path P d s.t. (P d ) < (P d). Noting that the crucial edge on P d, namely e, is no longer than Algorithm 1 Modified Dijkstra s Algorithm 1: Initialize ans = for for all V 2: sort all edges according to (e) in ascending order 3: for i =1to m do 4: Initialize dist =, Π = nil, maxedge =0for all V 5: dist s =0 6: Q = the set of all nodes in graph 7: hile Q is not empty do 8: u = Extract-Min(Q) 9: for each neighbor of u do 10: if e u, E(G ei ) then 11: t = MAX {maxedge u, (e u, )} 12: if dist u + (e u, ) < dist then 13: dist = dist u + (e u, ) 14: maxedge =t 15: Π = u 16: else if dist u +(e u, ) == dist then 17: if maxedge >tthen 18: maxedge = t 19: Π = u 20: end if 21: end if 22: end if 23: end for 24: end hile 25: for each node do 26: ans =min{ans, dist + maxedge } 27: end for 28: end for e (P d ) since they both belong to G e (P d ). Hence + (P d )= (P d )+(e ) <(P d )+(e (P d )) = + (P d ), contradicting P d is optimal. Lemma 2. Modified Dijkstra s Algorithm (MDA) computes shortest path hile keeping the crucial edge as short as possible in eery iteration. Proof: Line 4 to 24 orks similar to the standard Dijkstra s algorithm does. Besides, hen updating distance, MDA also updates the crucial edge to guarantee that it lies on the path and hen there is a tie, MDA ill choose the edge ith the smaller eight. Theorem 1. Modified Dijkstra s Algorithm computes optimal solution for eery node in O(m(n + m)logn) time. Proof: Lemma 1 indicates for any node V, optimal solution can be obtained by enumerating all possible crucial edges e (P ) and computing shortest path on G e (P ). By sorting all edges in nondecreasing order, eery subgraph G e (P ) is considered and it is shon in lemma 2, MDA correctly computes shortest path for eery node in eery G e (P ). Then optimal solution is obtained by examining all shortest path using the () metric plus the corresponding crucial edge. Dijkstra s algorithm runs O((n + m)logn) time hen using 133

4 binary heap, hence MDA runs in O(m(n+m)logn) time hen considering all layers. B. Shortest Path Computation using Metric 2 Gien a path P, let e (P ) be the crucial edge along the P and CNT(P ) be the number of occurrence of such edge. No our objectie becomes to find a path Q, such that (P )= e Q (e)+cnt(q) (e (Q)) is minimum. The layering technique can also be used in this problem. Hoeer, shortest path under a ceratin layer may not become a alid candidate for optimal solution. Here, e introduce a dynamic programming algorithm that can sole the problem optimally in O(n 2 m 2 ) time. Input is a eighted graph G =(V,E), V = n, E = m ith a specified source node s. In this paper, e only consider nonnegatie edge eight. As shon before, e use G e to represent the residue graph by deleting edges longer than e in G. Different from MDA1, in order to consider the number of crucial edges, dist is replaced by an array dist 0, dist 1,...dist n. One can think dist c be the shortest distance from s to by going through exactly c crucial edges and possibly some shorter edges. Similarly, e replace Π by Π c, 0 c n. Each Π c records predecessor of for the path corresponding to dist c. Lastly, ans is used as optimal solution from s to. Lemma 3. If P is the best path from s to, i.e., (P ) is minimum among all s- path, then P is computed in G e (P ) and dist CNT(P) = (P ). Proof: By definition, P exists in G e (P ) and CNT(P ) 1 since any path should go through at least one crucial edge. Noting (P )=(P )+CNT(P ) (e (P )), on one hand if e treat CNT(P ) as a fixed number, then e need to keep (P ) as small as possible. Inspired by idea of bellman-ford algorithm, e can achiee it by enumerating P, i.e., number of edges on P. On the other hand, e need to keep tracking number of crucial edges as ell. Hence, dist c is adopted to maintain such information, superscript c reflects exact number of crucial edges. From line 12 to line 25, dist c is updated either hen it comes from a neighbor ho has already itnessed c crucial edges or it comes from a neighbor ith c 1 crucial edges and the edge beteen is crucial. In either case, node gets a path, say P, ith exact c crucial edges on it and (P ) is minimum. At last, P can be selected by enumerating number of crucial edges and that is hat line 30 to 32 does. Lemma 4. Maxedge Shortest Path Algorithm(MSPA) runs in O(n 2 m) time for each G e. Proof: We can apply similar analysis of bellman-ford algorithm. Hoeer, e need to update dist c array, it takes extra O(n) time for eery node in eery iteration hen enumerating P. Hence, total running time is O(n 2 m). Theorem 2. MSPA computes optimal path for eery V in O(n 2 m 2 ) time. Algorithm 2 Maxedge Shortest Path Algorithm 1: Initialize ans = for for all V 2: sort all edges according to (e) in ascending order, say e 1,e 2,..., e m after sorting 3: for i =1to m do 4: Initialize dist c =, Π c = nil for all V and all 0 c n 5: dist 0 s =0 6: for j =1to n 1 do 7: for k =0to j do 8: for eery node V do 9: if dist k = then 10: continue 11: end if 12: for eery neighbor u of do 13: if (e u, ) >(e i ) then 14: continue 15: else if (e u, )==(e ) then 16: if dist k + (e u, ) < dist k+1 u then 17: dist k+1 u = dist k + (e u, ) 18: Π k+1 u 19: end if 20: else = 21: if dist k + (e u, ) < dist k u then 22: dist k u = dist k + (e u, ) 23: Π k u = 24: end if 25: end if 26: end for 27: end for 28: end for 29: end for 30: for i =1to n 1 do 31: ans =min{ans, dist i + i (e i )} 32: end for 33: end for Proof: By Lemma 3, ifp is obtained hen computing G e P. Then, by considering all possible G e, e could get P in one of these layering. It takes O(m) to generate all G e, by Lemma 4, MSPA runs in V in O(n 2 m 2 ) time. C. Computational Complexity of Disjoint Path Problem In this section, e study edge disjoint path in optical ireless netork. By reduction from ell knon Min-Max 2-Path Problem, i.e., min-max 2 edge disjoint path problem under normal length measurement, e sho it is also NP-complete if e try to minimize the longer path hen + length is applied. Then e gie an ILP formulation to sole this problem optimally. At last, e proide to approximation algorithm, one ith approximation ratio 4, running time 134

5 O((m + n)logn), the other one ith approximation ratio 2 hile running time is O(m(m + n)logn). Min-Max 2 Disjoint Path Problem (MinMax2PP) Instance: An undirected graph G =(V,E) ith a positie eight ( e) associated ith each edge e E, a source node s V, a destination node t V, and a positie number X. Question: Does there exist a pair of edge disjoint paths P 1 and P 2 from s to d in G such that (P 1 ) (P 2 ) X? The MinMax2PP problem is shon to be NP-complete in [10]. With a small modification, e sho NP-completeness still holds if + length measurement is adopted. Min-Max 2 Disjoint Path Problem in Optical Wireless Netorks (MinMax2OWFN) Instance: An undirected graph G =(V,E) ith a positie eight ( e) associated ith each edge e E, a source node s V, a destination node t V, and a positie number X. Question: Does there exist a pair of edge disjoint paths P 1 and P 2 from s to t in G such that + (P 1) + (P 2) X? Theorem 3. The MinMax2OWFN is NP-complete Proof: Eidently, MinMax2OWFN is in NP class, gien to edge joint path P 1 and P 2, e can check if + (P 1) + (P 2) X in polynomial time. We then transfer from MinMax2PP to MinMax2OWFN. Let graph G =(V,E) ith source node s, destination t and an integer X be an instance of MinMax2PP, e construct an instance G of MinMax2OWFN in folloing ay. 1) Create an identical graph G ith same nodes and edges in G. 2) Add one node s 0 to G. 3) Create to parallel edges e 01,e 02 beteen s 0 and s, (e 01 )=(e 02 )=max e G(E) (e) 4) Choose s 0 to the source node in G and t to be the destination. 5) Set X = X +2(e 01 ) It is easy to see, the construction takes polynomial time. No e need to sho a instance of MinMax2OWFN hae to edge disjoint paths from s 0 to t ith length at most X if and only if the corresponding instance hae to edge disjoint paths from s to t ith length at most X. Suppose there are to edge disjoint paths P 1 and P 2 from s 0 to t in G, such that + (P 1) + (P 2) X.Bythe ay e construct G, P 1 and P 2 must go through e 01 and e 02. W.l.o.g. e say e 01 P 1 and e 02 P 2. Since (e 01 )= (e 02 )=max e E(G ){(e)}, therefore e 01 and e 02 are the crucial edge on P 1 and P 2 respectiely. Hence, P 1 e 01 and P 2 e 02 are to edge disjoint path in G, ith length no greater than X 2(e 01 )=X. Conersely, no suppose P 1 and P 2 are to edge joint paths in G satisfying (P 1 ) (P 2 ) X. We follo the same argument aboe, P 1 + e 01 and P 2 + e 02 are to desired paths, ith length not exceeding X +2(e 01 )=X. D. Optimal Solution for the Disjoint Path Problem Here, e gie an ILP formulation for MinMax2OWFN. ILP for MinMax2OWFN min MP s.t. f i,j,1 1 i = s f j,i,1 = 1 i = t (i,j) E (j,i) E 0 otherise f i,j,2 1 i = s f j,i,2 = 1 i = t (i,j) E (j,i) E 0 otherise f i,j,1 + f i,j,2 1 (i, j) E 1 f i,j,1 (i, j) (i, j) E 2 f i,j,2 (i, j) (i, j) E MP 1 + f i,j,1 (i, j) MP 2 + (i,j) E (i,j) E f i,j,2 (i, j) f i,j,1 = {0, 1}, f i,j,2 = {0, 1} (i, j) E The folloing is a brief description of this ILP formulation. The first to equation represent flo constraint as normal shortest path problem does. f i,j,1 =1indicates path P 1 goes through edge (i, j), and 0 otherise. So it is ith f i,j,2 and path P 2. Constraint 3 ensures to edges are disjoint, since f i,j,1 and f i,j,2 cannot both be 1 at the same time. 1, 2 act as the eights of the crucial edges on P 1 and P 2 respectiely. Finally, e define MP to be the maximum of + (P 1 ) and + (P 2 ) and therefore try to minimize it. E. Approximation Algorithm for the Disjoint Path Problem, ith approximation factor 4 Next e propose a 4-approximation algorithm hich runs in O((n + m)logn) time. Gien G =(V,E) ith source s and destination t, the idea of approximation algorithm is to find to disjoint P 1 and P 2 such that (P 1 )+(P 2 ) is minimized. Such P 1 and P 2 can be found either using min cost max flo algorithm or the algorithm due to Suurballe presented in [11]. And e need to sho both + (P 1 ) and + (P 2 ) are at most four times of the optimal solution. Algorithm 3 MinMax2OWFN Approximation Algorithm 1 (MAA1) 1: Run Suurballe s algorithm on G, denote P 1, P 2 be to resulting path. 2: Compute + (P 1 ) and + (P 2 ). 3: Output max{ + (P 1 ), + (P 2 )}. 135

6 Lemma 5. For any path P, + (P ) 2(P ). Proof: By definition, + (P )=(P)+(e (P )). Since (e (P )) (P ), then + (P ) 2 (P ). Lemma 6. If P 1 and P 2 are to edge joint path from s to t such that (P 1 )+(P 2 ) is minimum, then + (P 1 ) and + (P 2 ) are at most four times of the optimal solution. Proof: Say opt is the optimal alue of a MinMax2OWFN instance and Q 1,Q 2 are to s t edge disjoint path in one optimal solution. W.l.o.g, e may suppose + (P 1 ) + (P 2 ) and + (Q 1 ) + (Q 2 ). Let (P 1 )+(P 2 )=p and (Q 1 )+(Q 2 )=q, by assumption, p q. Also, e hae + (P 1 ) = (P 1 )+e (P 1 ) 2p, opt = + (Q 1 ) = (Q 1 ) + e (Q 1 ) > q 2. Hence, + (P 2) opt + (P 1) opt < 2p q/2 4 Theorem 4. MAA1 is a 4-approximation algorithm running in O((n + m)logn) time and 4 is a tight bound. Proof: By Lemma 5 and 6, MAA1 has approximation ratio at most 4.Then e sho MAA1 has approximation at least 4 for certain cases. Consider the folloing graph. Algorithm 4 MinMax2OWFN Approximation Algorithm 2(MAA2) 1: set ans = 2: for eery G e of G do 3: Run Suurballe s algorithm on G e, denote P 1, P 2 be to resulting path. 4: Compute + (P 1 ) and + (P 2 ). 5: ans =min{ans, max{ + (P 1 ), + (P 2 )}}. 6: end for 7: Output ans. (e ). It suffices to sho it by contradiction. Suppose Which follos, (P 1)+(e ) max{ + (Q 1), + (Q 2)} (P 1)+(e ) max{ + (Q 1), + (Q 2)} (P 1 )+(e ) + (Q 1 )+ + (Q 2 ) < 2. Weproe 2, then (P 1 )+(e ) (Q 1 )+(e (Q 1 )) + (Q 2 )+(e (Q 2 )) By definition, e is one of e (Q 1 ), e (Q 2 ). Hence, (P 1 ) >(Q 1 )+(Q 2 ) It is easy to check, P 1 = {s t}, P 2 = {s r t} are to edge disjoint path ith minimum length 2k+2, + (P 1 )= 4k > + (P 2 )=3. Hoeer, let Q 1 = {s u 1 u 2... u k 1 u k r t}, Q 2 = {s r k 1 k t}, then (Q 1 )+(Q 2 )=2k +4 hile + (Q 1 )= + (Q 2 )=k (P 1) + Q 1 = 4k k+3 4 hen k is sufficiently large. Hence, 4 is a tight bound for MAA1. We need O((n + m)logn) time running Suurballe s algorithm and O(n) time computing + (P 1 ) and + (P 2 ). Therefore total running time is O((n + m)logn). It is impossible since (P 1 )+(P 2 ) is minimum in layer G e. Theorem 5. MAA2 is a 2-approximation algorithm running in O(m(n + m)logn) time and 2 is a tight bound. Proof: By Lemma 7, in one of the layer, e guarantee to hae a 2-approximation solution. Since e take minimum outcome among all layers, the final result is no orse than tice of the optimal solution. No e need to sho there exists certain case, such that MAA2 is no good than tice of the optimal solution. Consider the folloing graph F. Approximation Algorithm for the Disjoint Path Problem, ith approximation factor 2 In MAA1, layering technique is not used and e only consider the original graph. Hoeer, by taking all G e of G into account, e can hae a better approximation ratio. Say Q 1, Q 2 are to disjoint paths in one optimal solution. Let e =max{e (Q 1 ),e (Q 2 )} and P 1, P 2 be the resulting paths hen computing layer G e ;.l.o.g, e may assume (P 1 ) >(P 2 ). Also, let ans e =max{ + (P 1 ), + (P 2 )}. Lemma 7. ans e < 2max{ + (Q 1 ), + (Q 2 )}. Proof: Noting that (e (P 1 )) (e ) and (e (P 2 )) (e ) since they both belong to G e. Then ans e (P 1 )+ There is only one layer, and P 1 = {s x 1 x 2... x 2k 1 x 2 k t}, P 2 = {s r t} are to edge disjoint path ith minimum length 2k +3, + (P 1 )= 2k +2> + (P 2 )=3. Again, set Q 1 = {s u 1 u 2... u k 1 u k r t}, Q 2 = {s r k 1 k t}, then (Q 1 )+(Q 2 )=2k +4 hile + (Q 1 )= + (Q 2 )=k (P 1) + Q 1 = 2k+2 k+3 2 hen k is sufficiently large. Hence, 2 is a tight bound for MAA2. Finally, it is easy to see that the running time is O(m(n + m)logn). 136

7 S D Opt Approx Approx Approx Approx node node Sol Sol 1 ratio 1 Sol 2 ratio Fig. 1. The ARPANET graph ith 20 nodes and 32 links TABLE I COMPARISON OF THE APPROXIMATE SOLUTIONS WITH THE OPTIMAL SOLUTION FOR THE ARPANET GRAPH G. Experimental Results for the Disjoint Path Problem In this section, e present the results of simulations for comparing the performance of our approximation algorithms ith the optimal solution hen + () metric is applied. The simulation experiments hae been carried out on the ARPANET topology (as shon in Fig 1 ith nodes and links shon in black) hich has tenty nodes and thirty to links. The eights of the links hae been randomly generated and lie in the range of to and eleen (as shon in red in Fig 1) and e consider the graph to be undirected. The results of the comparison is presented in Table I. We hae compared the lengths of the longer of the to edge disjoint paths computed by the optimal and the approximate solutions for seenteen different source-destination pairs. It may be noted that for almost 65% of the cases, the approximate algorithms obtain the optimal solution. In the remaining cases, the approximate solutions lie ithin a factor of 1.2 of the optimal solution Thus, een though the approximation ratio in the orst case are proen to be 4 and 2, in practical cases, it is ithin 1.2. From these experimental results, e can conclude that the approximation algorithms produce optimal or near optimal solutions in majority of the cases. It may be noted that the to approximation algorithms perform in a similar fashion in the ARPANET graph, hoeer, as proen theoretically, the to approximation algorithms differ in their orst case approximation ratio. V. CONCLUSION In this paper, e study the shortest path problem in FiWi netorks. Based on the path length metrics proposed in [3], [5], e present polynomial time algorithms for the single path scenario. In the disjoint path scenario, e proe that the problem of finding a pair of disjoint paths, here the length of the longer path is shortest, is NP-complete. We proide an ILP solution for the disjoint path problem and propose to approximation algorithms. Both the approximation algorithms hae a constant factor approximation bound. Hoeer, there is a trade-off beteen the quality of the solution (approximation bound) and the execution time. Finally, e sho that both the approximation algorithms obtain near optimal results through simulation using the ARPANET topology. REFERENCES [1] N. Ghazisaidi, M. Maier, and C. M. Assi, Fiber-Wireless (FiWi) Access Netorks: A Surey, IEEE Communications Magazine, ol. 47, no. 2, pp , Feb [2] N. Ghazisaidi, and M. Maier, Fiber-Wireless (FiWi) Access Netorks: Challenges and Opportunities, IEEE Netork, ol. 25, no. 1, pp 36-42, Feb [3] Z. Zheng, J. Wang, X. Wang, ONU placement in fiber-ireless (FiWi) netorks considering peer-to-peer communications, IEEE Globecom, [4] Z. Zheng, J. Wang, X. Wang, A study of netork throughput gain in optical-ireless (FiWi) netorks subject to peer-to-peer commuincations, IEEE ICC, [5] F. Aurzada, M. Leesque, M. Maier, M. Reisslein, FiWi Access Netorks Based on Next-Generation PON and Gigabit-Class WLAN Technologies: A Capacity and Delay Analysis, IEEE/ACM Transactions on Netorking, to appear. [6] A. Sen, B.Hao. B. Shen, L.Zhou and S. Ganguly, On maximum aailable bandidth through disjoint paths, Proc. of IEEE Conf. on High Performance Sitching and Routing, [7] M. Mosko, J.J. Garcia-Luna-Acees, Multipath routing in ireless mesh netorks, Proc. of IEEE Workshop on Wireless Mesh Netorks, [8] J. Wang, K. Wu, S. Li and C. Qiao, Performance Modeling and Analysis of Multi-Path Routing in Integrated Fiber-Wireless (FiWi) Netorks, IEEE Infocom mini conference, [9] S. Li, J. Wang, C. Qiao, Y. Xu, Mitigating Packet Reordering in FiWi Netorks, IEEE/OSA Journal of Optical Communications and Netorking, ol. 3, pp , [10] C. Li, S.T. McCormick and D.Simchi-Lei, Complexity of Finding To Disjoint Paths ith Min- Max Objectie, Discrete Applied Mathematics, ol. 26, pp , [11] J. W. Suurballe, Disjoint paths in a netork, Netorks, ol. 4, pp ,