OSPF Weight Setting Optimization for Single Link Failures

International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 DOI : 10.512 1/ijcnc.2011. 3111 168 OSPF W EIGHT S ETTING O PTIMIZ ATION F OR S INGLE L IN K F AILURES Mohammed H. S qalli, Sadiq M. S ait, and Sy ed Asadullah Computer Engineeri ng Department King Fahd Un iversity of Petroleum & Mine rals Dhahran 31261, Sa udi Arabia {sqalli, sadiq,sas ad}@kfupm. edu.sa A BSTRACT In operatio nal networks, node s a re c onnecte d via m ultiple links fo r l oad sharing and redun dancy. This is done to make sure that a failure of a link does not disconnect or isolate some parts of the network. However, link failures have an effect on routing, as the routers find alternate paths for t he traffic originally flowin g thr ough th e link w hich has faile d. Thi s effect is severe in case of f ailure of a critic al link i n the network, such as backbone l inks or the l inks carrying highe r traffic loads. When routing i s done using the Open Shortest Path Firs t ( OSPF) routing protoc ol, the ori ginal weight selection for the normal state topology m ay not be as ef ficient fo r t he failure state. In this paper, we inve stigate the s ingle link fa ilure issue with an obje ctive to find a weight s etting which resul ts in efficient rout ing i n normal and failure states. We en gineer Tabu Search I terative heuristic usi ng tw o different impleme ntation strategies to solve t he O SPF weight se tting proble m f or link failu re scenarios. We evalu ate t hese heuristics and show thro ugh exp erimental r esults t hat both he uristics e fficiently h andle weight setting f or t he failure state. A co mparison of both strategies is al so presente d. K EYWORDS Routing, O pen Shortest Path First ( OSPF), OSPF W eight Setti ng Problem, Iterative Heuri stics, Link Failure, T abu Search. 1. I NTRODUC TION OSPF is an i ntra-domain routing protocol that uses link weights t o make rout in g decisions and compute the shortest paths. Dif ferent we ight assignment strategies have been discussed in the literature [11] including the Unit OSPF, Inverse Capacity OSPF, Random OSPF etc. A better selection of the OSPF link weights can lead to efficient network utili zation [1, 2]. Iterative heuristics have been extensively used [10, 3, 4] and implemente d using different strategies to achieve this goal. Ericsson et al. [4] proposed a Genetic Algorit hm and used the set of t est problems considered in [11]. A hybrid GA was also proposed by them [5] which makes use of the dynamic shortest path algorithm to recompute shortest p aths after the mod ification of link we ights. Sridharan et al. [6] developed another heuristic f or a slightly different version of t he problem, in which the fl ow is split among a subset of the outgoing links on the shortest paths to the destination IP address. In this wor k, we have used a Tabu Search (T S) al gorithm [ 7] to solve t he OSPFWS problem. Tabu search is an iterative heuristic that h as b een applied for solving a range of combinatorial optimization problems in different fields. The detailed d escription and related r eferences fo r tabu search can be found in [7]. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 169 However, all strategies work on the ass umption t hat the topology is fixed and there are no failures in the network. A n etwork ma y expe rience a link fai lure resu lting in a change in topology due to the l oss of a link, when the n etwork st ate changes (Failure State) due to l ink failure, t he routing paths are also not t he same as in th e original state (Nor mal State). The optimized weights for the ori ginal t opology and d emand may no l onger be good enou gh f or the new t opology wit h the failed link. The abs ence of the failed l ink causes the traff ic which was originally flowing t hrough this link t o flow through other ava ilable links. The fact t hat the network was not optimize d for these flows c an result in an inefficient mapping of traffi c on to available links. This may also cause c ongestion i n some parts of th e network, especia lly in the case of higher demands. One solution to this issue is to apply a new set of OSPF weights to links which optimize the new topology ( Failure State). However, it is cum bersome to change the weights on each link in the entire topology and also n ot very practical in ca se o f larger networks. One would suppose that once the set of OSPF weights have been fi xe d, the operator would not want to c hange t hese weights in order to adapt for such changes in the state of the network. He nce, it is required to adapt the original heuristic t o optimize link weights taking i nto consideration single l ink failure scenarios. In other words it is required to find a set of weights that work for both the normal and failed state of the network without cons iderable degradation in performanc e in both states. Link fail ure scenarios requir e dealing with two states of a network. The first state where all links are functional i s denoted as Nor mal state and the other state where a lin k has failed i s denoted as Failure state. In t his p aper, which is an extension of Sqalli et al. [8], two different strategies are devised and implemented to address this i ss ue. The first strategy viz. LinkFailure- FT is similar to the approach a dopted by Fortz and T horoup [9] with some modifications. Another new strategy viz. LinkFailur e- SS is proposed, where the weights are first optimized for the Failure state. Keeping these weights fixed, all combinations of weights are tried for the added li nk to find the best cost for the Normal state. Both strategies are discu ssed further in this paper. Similar problem has been attempted by Fortz and Thoroup [9]. In their approach, a set of links was considered as critical, and in eac h it eration one of these link s was failed based on the maximum uti lization among critical links. The co st of nor mal topolo gy and the resulting failed topology wa s a verage d and the se arch w as driven t o find a solution which min imize s the average cost. One of ou r impl eme ntations in this work is similar to thi s approach bu t with the modification th at the link failed is always t he one connec ted between no des carrying th e highest demand. This simulates the worst case scenario. The rest of the paper is organize d a s follows; The OSPFWS probl em statement a nd the c ost functions proposed in the literature are presented in Sectio n 2. The t wo Link Failure algorithms are di sc ussed and analyzed in Section 3. This i s followed by the experimental results inclu ding the compa rison of both a lgorithms under Normal and Fail ure state in Sec tion 4. Finall y, we conclude in Section 5. 2. P ROBLE M S TATEMENT AND COST FUNC TION The OSP F w eight s etting probl em can be stated as fo llows: Given a directed network of nodes and arcs ) , ( = A N G , a demand matrix D , and capacity a C for each arc A a ∈ , determine a positive integer weig ht ] [1, max a w w ∈ for each ar c A a ∈ such that the objective function or cost functi on Φ i s minimized. Wh en rou ting is done using OSPF the assigned link weig hts completely determine t he shor test paths, and hence the traffic flows. Base d on these tr affic International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 170 flows the p artial loads on each arc f or a given destination are computed. This i s done for all destination nodes. The aggregate d partial l oads for all destinations on a particular arc give the total load a l on that arc. The cost of sending traffic t hrough t his arc is given by ) ( a a l Φ . The cost value depends on t he utilization of the arc and is given by the lin ear function proposed b y Fortz and Thoroup.            ≤ ≤ ≤ ≤ ≤ ≤ Φ infinity c l for c l for c l for c l for c l for c l for l a a a a a a a < / 11/10 5000 11/10, < / 1 500 1, < / 9/10 70 9/10, < / 2/3 10 2/3, < / 1/3 3 1/3, < / 0 1 = ) ( ' (1) The Fortz cost function is give n in equation 2. ) ( = a a A a l Φ Φ ∑ ∈ (2) The objective is to minimize Φ , subject to these constraints: , = ) , ( ) , ( A a f l t s a NXN t s a ∈ ∑ ∈ (3) 0 ) , ( ≥ t s a f (4) In constraint 3, for traffic between source destination pair (s,t), ) , ( t s a f indicates t he amount of traffic flow that goes ove r ar c a . The detail ed steps showing the f ormulation of thi s cost function can be found in t he li terature [11, 10]. 3. L INK F AILURE Handling li nk failure scenarios requires dealing with two sta tes of a network. In the Normal state, the topology is said to have 1 + n links. There exist a set of weights W which optimize the cost for t his topology . The cost function for this i s denoted by OH n 1) ( + Φ , where OH stands for Original Heuristic and 1 + n indicates a topology wi th 1 + n links. In the case of failure, these weights fo r the new topology will re sult i n another cost and is denoted by OH n 1) 1] ([ − + Φ . Here, 1 1] [ − + n indicates failure of link and topolog y change from 1 + n to n links. The abov e functions are representa tive of the costs when the No rmal state top ology was optimized using the original heuristic. 3.1. L inkFailure – FT In L inkFailure – FT strategy, to find optimum we ights representing both the n ormal and the failed states, the i dea is no t t o minimize the cost of each state individually but to minimize the combined or average cost of both s tates. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 171 For a given solutio n or set of weights W for the Normal stat e, the cost is denoted by Φ (n+1) and for the Failure state w ith the same set of weights minus t he failed link (W-a) the cost i s ` Φ (n). The objective is to find the set of weights w hich minimi zes the new cos t function: Avg Φ = 2 1 / ( Φ (n+1) + ` Φ (n)) (5) Starting with a random ini tial solution for t he No rma l state ( Norm T ) the s ame set of weights, except the weight of th e failed li nk, are transferred to the failed state ( Fail T ) and bo th topologies find the shortest paths and the cost of the initial solution. Tabu Search is started on Norm T by making random moves , and every time the same move is again transferred to Fail T . Both topologies find th e shortest path and the corresponding cost after a move. Th e cost of the new solution for Norm T is denoted as Avg n 1) ( + Φ and the new cost of Fail T i s denoted as Avg n ) ( Φ . The c ost of the current solution Avg Φ is the average of Avg n 1) ( + Φ and Avg n ) ( Φ . Avg Φ = 2 1 / ( Avg n 1) ( + Φ + Avg n ) ( Φ ) (6) Here, Avg n 1) ( + Φ and Avg n ) ( Φ indicate the co st of Norm T and Fail T respectively while optimizing the average co st function. We c ontinue Tabu Searc h and compute the average cost for ea ch iterati on until the termination criteria i s met. The s et of weights whi ch gives the least value of Avg Φ is the best solution obtained by the new heuristi c. Fig ure 1 shows t he structure of the LinkFailure – FT algorithm. 3.2. Perfor mance Evaluation of LinkFailure – FT The performance of t his strategy can be e valuate d by comparing the cost obtained fo r Norm T and Fail T using thi s heuri stic with that of the original. The difference b etween t he co sts of the original and the new heuristic would indicate a ga in or loss i n the solution quality. For Norm T , this difference would be: Norm δ = OH n 1) ( + Φ - Avg n 1) ( + Φ (7) Optimizing weights using the original heuristic is expected to give a bet ter cost than opt imizing for average cost. Hence, the value of Norm δ is expected to be negative, indicating a loss in solution quality in the Norma l state. A smaller Norm δ value or a value close t o zero would indicate that the heuristic is pe rforming well in the Normal state. In the case of the Failure sta te the cost difference would be indicated a s: Fail δ = OH n 1) 1] ([ − + Φ - Avg n ) ( Φ (8) The purpose of optimizing t he weights for li nk failure is to achieve a b etter cost in case of a Failure state than would have been achieved with the original heuristic. Hence, Fail δ must be a International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 172 positive value indicating an i mproveme nt in the solution quality. A larger Fail δ value would indicate t hat the new heuristic is pe rforming w ell in the ca se of a F ailure st ate. Hence, a combination of smaller Norm δ value and larger Fail δ value wo uld b e an ideal case indicating minimal lo ss i n the case of the Normal state and significant impro vemen t i n the case o f t he Failure state. Algorithm: Link Failure-FT 0 S : initial solution. S : solution. b S : best solution. a : failed arc. i W : Weight of arc i. Norm T : Normal state. Fail T : Failure state. Begin Norm T : 1. Generate 0 S ; 2. Transfer { 0 S - a} to Fail T ; 3. Compute Avg n 1) ( + Φ ; Fail T : 4. Compute Avg n ) ( Φ ; do Norm T : 5a. Move (i, i W ); 6a. Compute Avg n 1) ( + Φ ; Fail T : 5b. Move (i, i W ); 6b. Compute Avg n ) ( Φ ; 7. Avg Φ = 2 1 / ( Avg n 1) ( + Φ + Avg n ) ( Φ ; While (Termination crite ria is not met) 8. b S = S for min( Avg Φ ); End Figure 1 : Structur e of the L inkFailure – FT algo rithm. 3.3. L inkFailure – SS In the previous strategy, we have tried to o pt imize weights for t he ave rage cost of Norm T and Fail T . In this section, we propose another strat egy whi ch optimizes weights for Fail T and fi nds the best solution f or Norm T b y keeping the weights obtained from Fail T unchanged and trying all possible weights f or t he one additional lin k. The test cases and b enchmark topologies used were the same as for the previous s trategy. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 173 We start with a random i nitial solution for Fail T and find the shortest pa ths and corresponding cost for thi s solution. Tabu Search is started on Fail T by making random moves and after each move, t he shortest paths and corresponding cost are computed. The co st of the new solutio n for Fail T is denoted as OH n ) ( Φ which indicates that th e cost is for th e topology with n li nks optimized using the or iginal heuristic. Once the termination criterion is met, we obtain the bes t solution for Fail T and compute its best cost. The final n weights are t ransferred to Norm T . The w eight on the additional (n+1 th ) link is assigned values from 1 to 20. For eac h w eight i W , the cost of the th i solution is computed. The twenty costs obtained are compared to find the best solution for Norm T . Thi s is denoted by 20 1) ] ([ + Φ OH n which indicates that the cost is for topology with 1 + n links where n links are optimized using the or iginal heuristic and o ne additional link is o ptimized by fi nding t he best solution from th e twenty possible combinations. Figure 2 shows the structure of the LinkFailure – SS algorithm Algorithm: Link Failure-SS 0 S : initial solution. S : solution. b S : best solution. a : failed arc. i W : Weight of arc i. Norm T : Normal state. Fail T : Failure state. Begin Fail T : 1. Generate 0 S ; 2. Compute OH n ) ( 0 Φ ; do 3. Move(i, i W ); 4. Compute OH n ) ( Φ ; While (Termination crite ria is not met) 5. b S = S for min( OH n ) ( Φ ); 6. Transfer b S to Norm T ; Norm T : 7. Compute a n 1) ( + Φ ; for a={1, 2, ..., 20} 8. 20 1) ] ([ + Φ OH n = min( a n 1) ( + Φ ); 9. b S = S for 20 1) ] ([ + Φ OH n ; End Figure 2:Structure of the L inkFailure-SS algorithm. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 174 3.4. Perfor mance Evaluation of LinkFailure – SS Similar to the FT approach, the performance of this strategy c an be evaluated by comparing the cost obtained for Norm T and Fail T using the SS h euristic wi th t hat of the or iginal h euristic (OH). The difference between costs of the original and the new heuristic would indicate a gain or loss in the solution quality. For Norm T this difference would be: Norm δ = OH n 1) ( + Φ - 20 1) ] ([ + Φ OH n (9) In the case of t he Failure st ate , the original heuristic wi ll end up with a cost OH n 1) 1] ([ − + Φ and the SS heuristic with a co st of OH n ) ( Φ . Hence, the c ost difference would be indicated as: Fail δ = OH n 1) 1] ([ − + Φ - OH n ) ( Φ (10) In the SS approach, the w eights are optimized for Fail T and are expected to achieve a bett er cost in t he c ase of a link f ailure than would have been ac hieved with th e original heuristic. Hence, Fail δ must be a p ositi ve value indicating an improve ment in the solution quality. 3.5. FT versus SS In the case of LinkFailure-FT , we simultaneously optimize t wo states of a network viz. Norm T and Fail T , whereas in L inkFailure-SS we only optimize Fail T and then try the best possible weight for th e one additional link to optimize Norm T . Hence, the SS approach has a faster convergence when compared t o FT; which is a major factor when de aling with larger networks and higher demands . As discussed earlier, SS is op timized for the Fail ure st ate and hence should n ot only give b etter solution when compared to OH but also should p erform better than FT in the Failure state. In the FT approach, the wei ghts are selected to optimize t he average cost and not the best cost for individual states. Any heuristic, to be accepta ble, must not degrade the performance of the network in the Normal state. In ot her words it should result in a solution q uality as close to the Original Heuristic (OH) as possible . 4. R ESULTS In this section, we present the experimental re sults fo r t he t wo h euristics mentioned i n the previous section. The benchmarks used for the evaluation of the o riginal heuristic for no failure case [12, 11] were also used for the link failure case. Due to the change in topology ( different number of links) i n the two states, the original test case would repre sen t only o ne of the states and a modified test case w ould r eprese nt the other state. Representing the failed st ate with a modified test case would r equire deletion of the corresponding link entries from the files representing the graph and capacity of li nks . This could also result in a disconnection of the graph. T o avoid this, we represented t he Fa i lure state wit h the original test cases. To represent the Normal state, we add an additional link between t wo nodes 1 n and 2 n . The nod es selected were th e ones with t he hi ghest demand between them in the demand matrix. Failing this p articula r link which i s directly connecte d between the t wo nodes h aving th e h ighest d emand betw een them would cause t he worst effect on th e network. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 175 Hence, if our heuristic is able to optimize weights for the worst case scenario th en it is expected to be robust. The notations used to denote Cost in the Normal and Failure state are shown below: N FT N SS N OH avg N 1) ( + Φ 20 1) ] ([ + Φ N oh N 1) ( + Φ F FT F SS F OH avg N ) ( Φ oh N ) ( Φ oh N 1) 1] ([ − + Φ 4.1. FT versus OH Experimental results f or the t wo strategies implem ented for the single link failure scenario are presented in this section. The i ndividual performance of each strategy can be evaluated by comparing its results in Norma l and Failure States to the Origina l Heuristic. Table 1 shows th e Co st v alues obtained using FT Strategy and OH for five different d emands using th e test case h100N360a. From the table, it can be seen t hat i n t he Normal state the Cost of FT is marginally higher than OH, which can be seen in the Norm δ c olumn wh ich shows the Cost difference for the two st rategies in the Normal S tate. Ne gative values indicate a loss. As expected, there is some loss in the Normal St ate. In the Fail ure State, for all demands except Demand-9, the FT Cost is l ess than th e OH Cost a s indicated by a positive value in the column Fail δ . Hence, t here is some gain in t he Failure State. The o verall gain or loss is indicated in the column δ . The value of Norm δ is more than the value of Fail δ for higher demands D11, D12 which i mplies that the margin of lo ss in Normal state is more than the gain in the Failure State f or this case at higher demands. Results a lso show an overall gain for the two demand s D8 and D10. Table 1: Cost Comparison F T versus OH in Normal and Fa ilure State for h100N360a Network. D N FT N OH N δ F FT F OH F δ δ D8 1.313 1.320 0.006 1.336 2.743 1.406 1.413 D9 1.482 1.448 -0.033 1.538 1.494 -0.0 44 -0.077 D10 2.096 1.985 -0.111 2.315 5 .711 3.396 3.285 D11 4.498 4.369 -0.129 6.017 6 .057 0.040 -0.089 D12 17.973 14.076 -3.897 24.398 25.487 1.089 -2.809 International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 176 4.2. SS v ersus OH Table 2 shows similar comparison for the SS Strategy. Even in this case, the values of Norm δ for SS are marginally higher than those of OH, and the values of Fail δ for SS are well below those of OH f or all five demands shown in the table. This shows that there is a slight lo ss in the Normal State and a significant i mprovement in the Failure State. There is an o verall gain as indicated by a positi ve values in the l ast column δ . Hence, there is an i mprovement in performance due to the use of SS strateg y compared to OH. Table 2: Cost Comparison SS v ersus OH in Normal a nd Failure State for h100N360a Network. D N SS N OH N δ F SS F OH F δ δ D8 1.329 1.320 - 0.009 1.343 2.743 1.399 1.390 D9 1.480 1.448 -0.032 1.487 1.494 0.00 7 -0.025 D10 1.986 1.985 -0.001 2.010 5.711 3.70 1 3.700 D11 4.389 4.369 -0.019 5.330 6.057 0.72 7 0.708 D12 14.316 14.076 -0.240 18.158 25.487 7.329 7.089 4.3. FT versus SS We have seen that both strategies are pe rforming better than the Original Heuristic in the Failure state while OH has slightly b etter r esults f or the Normal state. We now compare t he SS and FT results t o show which of the two h euristics performs bett er. The comparison is sho wn i n Table 3. In t he Normal state, for the demands D8 - D10 both strategies h ave almost the same cost values with marginal di fferences in favour of SS. For the hi ghest demand D12, SS cl early performs better t han FT. Overall, for the Normal State, it ca n be sa id that SS performs better t han FT for this test case. For the Fail ure State, SS clearly outp erforms FT for all demands. Th is is expected as the strategy is sp ec ifically designed t o optimize weights for the Failure State or in ot her words to minimize the Failure State Co st. Hence, S S is always expected t o produce better results for a Failure State. The overall compariso n shows superiorit y of SS over FT for this test case. Comparison of all three strategies for this test case is presented below. Table 3: Cost Comparison F T versus SS in Normal and Failure State for h100N360a Network. Dema nd N FT N SS F FT F SS D8 1.31326 1.32905 1.33621 1.34312 D9 1.48152 1.48005 1.53819 1.48682 D10 2.09604 1.98619 2.31527 2.01001 D11 4.49806 4.38878 6.01734 5.33037 D12 17.9732 14.3157 24.3984 18.1582 4.4. OH versus FT versus SS Figure 3 shows the graph with the Cost compa rison of all the thr ee heuristics in the Normal s tate and in Figure 4 for the Fa ilur e State for the h100N360a Network. In Figure 3, it can be se en that OH has the best Cost in the Normal state which is very clo sely matched by SS. FT comparatively has the worst Cost in the Normal state. In the Failure state SS International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 177 outperforms both FT and OH as s een in Figure 4. Hence, SS has prov ed to be ha ving a marginal loss (negligible in the case of lower demands) i n the Normal st ate and a sig nificant gain in the case of Failure, which is the idea l requirement for these types of problems. Experiments were conducted for five more t est cases and are presented i n the following fi gures for both the normal and failure states. All t hese figures provide a comparison for all three algorithms, i.e., FT, SS, an d OH. A su mmary of t he results obtained is p resented at the end of this section. Figure 3 : Cost Comparison FT, SS and OH in the Norma l state for h100N360a Network. Figure 4: Cost Comparis on FT, SS and OH in the Failure state for h100N360a Ne twork. Figure 5 shows the graph with the Cost compa rison of all th e three heuristics in the Normal s tate and in Figure 6 for the Fa ilur e State for the r50N228a Netw ork. In Figure 5 , it can be seen that both SS and FT show comparable results in the Nor mal state. In the Failure state, SS outperforms both FT and OH as seen in Fig ure 6. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 178 Figure 5: Comparis on of FT, SS, and OH in the Normal state for r50N228a Network. Figure 6: Comparison of FT, SS, an d OH in the Failure s tate for r50 N228a Network. Figure 7 shows the graph with the Cost compa rison of all the thr ee heuristics in the Normal s tate and in Figure 8 for the Fa ilur e State for the r100N503a Ne twork. In Figure 7 , it can be seen that both SS and FT show comparable results in the Nor mal state . In the Failure state, SS outperforms both FT and OH as seen in F igure 8. Figure 7: Comparison of FT, SS, an d OH in the Normal state for r100 N503a Ne twork. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 179 Fig ure 8: Comparison of FT, SS, and OH in the Failure state for r100N503a Network. Figure 9 shows the graph with the Cost compa rison of all the thr ee heuristics in the Normal s tate and in Figure 10 for the Failure State for the w5 0N169a Network. In Figure 9, i t can be seen that all strategies perfor m equally well in the Normal state for all demands. In th e Failure state, similarly all strategies p erform equally well for al l demands as seen in Figure 10. This indicates that a link failure does not h ave significant eff ec t on network performance for this test case . Figure 9: Comparis on of FT, SS, and OH in the Normal state for w50N169a Network. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 180 Figure 1 0: Comparison of FT, SS, and OH in the Failure state for w50N169a Network. Figure 1 1 shows the graph with t he Cos t comparison of all the three heuristics in the Normal state and in Figure 12 for the Failure State for the w100N476a Ne twork. In Figure 11, it can be seen that all str ateg ies p erform equally w ell in the Normal state for all demands. In th e Failure state, similarly all strategies p erform equally well for al l demands as seen in Figure 12. This indicates that a link failure does not h ave significant eff ec t on network performance for this test case . Fig ure 11: Comparison of FT, SS, and OH in the Normal state for w100N476a Network. Figure 12: Compa rison of FT, SS, and OH in the Failure state for w100N476a Network. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 181 Figure 1 3 shows the graph with t he Cos t comparison of all the three heuristics in the Normal state and in Figure 14 for the Failure State for the h50N148a Netw ork. In Figure 1 3, it can be seen that OH has the best Cost in the Normal state which is very clo sely matched by SS. FT comparatively has the worst Cost in the Normal state. In the Failure state SS outperforms both FT and OH a s seen in Figure 14. Figure 1 3: Comparison of FT, SS, and OH in the Normal state for h50N148a Network. Figure 1 4: Comparison of FT, SS, and OH in the Failure state for h50N148a Network. 4.5. Summary of Results In all t he test cases, SS ac hieves the be st results for the Failure state ( Fail δ ) and also for the overall improvemen t ( δ ). SS i s fol lowed by FT in t he Failure state, which performs b etter than OH. I n th e Normal state, SS p erforms slightly b etter than FT for t he two test cases h50 N148a and h100N360a and has comparable results f or the two cases r100N503a and r50N228a. For the two Waxman graphs, w50 N169a and w100N476a, all strategies perform equally well in Normal and Failure state f or all demands. This i ndicates t hat a l ink failure does not have signif icant effect on netwo rk performa nce for these two cases. Finally, it can also be obs erved that for lower demands (Demand-8, Demand- 9), th e results are almost the same for all th e six test cases. This indicates t hat, if the load on the network is low, there is mini mum eff ec t of the link failure International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 182 on t he network performance and the or iginal heuristic itself i s efficient enough t o h andle single link failures. 5. C ONCLUSIONS The single li nk failure i s sue in OSPF routi ng was addressed in this work to find a weight setting for t he links wh ich r esults in efficient routing in no rma l and fail ure states. Two new heuristics based on Tabu Search were proposed i n this paper, namely LinkFailure-SS and LinkFailure-FT. Both heuristics were evaluated and they both produced better results when compare d to the original heuristic i n the Failure state. In additi on, the SS approach is found to gi ve better results than the FT approach in bot h normal and fail ure states. Therefore, it can be con cluded that the SS approach is an efficient way to tac kle single lin k failure issues. It was also shown th rough experimental results that at l ower demands and tr affic loads the effect of li nk failure on network performance is less and th e original heuristic can also handle single link failures if the traff ic load on the network is low. A CKNOWLEDGEMENTS Acknowledgement goes to KFUPM for supporting t his research work. This m aterial is based i n part o n work supported by a K FUPM project under G rant No. SAB-2006-10. The authors wi sh to thank Bernard Fortz and Mikkel Thorup fo r sharing the test problems. R EFERENCES [1] Be rnard Fortz, J. Rexford & Mikkel T horup (2002) “Traffic eng ineering with traditio nal IP routing pr otocols”, IE EE Communicatio ns Magazine , 1 18-124. [2] Be rnard Fortz & Mikkel Thorup (2000) “ Internet Traffic Engineering b y Optimizing OS PF Weights”, IE EE INFOCOM . [3] S adiq M. Sait, Mohammed H. Sqalli & Mohammed Aijaz M ohiuddin ( 2006) “Engineering Evolutiona ry A lgorithm t o S olve Multi-objective OSPF Weight Setting Problem ”, Australian Conference on Artificial Inte lligence , 950-955. [4] M Resende Ericss on & P P ardalos (2002) “A genetic al gorithm for the we ight setti ng pr oblem in OSPF r outing”, Com binatorial Optimi sation Conference . [5] L.S. Buriol, M.G.C. Resende, C.C. Ribeiro & M. Thorup (200 5) “A Hybrid Genetic Al gorithm for the Weight Set ting Pr oblem i n OS PF/IS-IS Rout ing”, AT &T La bs Re search Te chnical Report, TD-5 NTN5G. [6] A . Sridharan, R. Gué rin & C. Diot (200 5) “Achie ving near-opt imal traffi c engineerin g solutions for current OS PF/IS- IS networks”, Spri nt ATL Tec hnical Report TR02-ATL-0220 37, Sprint Labs. [7] S adiq M. Sait & H abib Youss ef (1999) Iter ative Computer Algorithms and the ir Application t o Engineering, IEEE C omputer Societ y Pres s. [8] M ohammed H. Sqa lli, Sadiq M. Sait & Syed A sadullah ( 2010) “O ptimizing O SP F Routing for Link Failure Scenar ios”, The Fifth Inter national Works hop on A dvanced Comput ation for Engineering Applicatio ns (ACEA'2010), Taif University , Taif, Sau di Arabia. [9] Be rnard Fortz & Mikkel Thorup ( 2003) “Robust optimizati on of OSP F/IS-IS wei ghts”, Proceedings of the Intern ational Netw ork Optimiz ation Conference , 225-230. [10] Mohamme d H . Sqalli, Sadiq M. Sait & Mohamme d Aijaz Mohiuddin (20 06) “A n Enhanced Estimator to Multi-objecti ve OSP F Weight Settin g Problem” Netw ork Opera tions and Managemen t Symposium, NOMS . [11] Bernard Fortz & Mikkel Thorup (2 000) “Increasing Inte rnet Capacity U sing L ocal S earch”, Technical Report IS-MG. International Journal o f C omputer Netw orks & C ommunications (IJCNC) Vol.3, N o.1, Januar y 2011 183 [12] W. Zegura (19 96) GT-I TM: Geor g ia Tech in ternetwor k t opology model s (sof tware), http://www.cc .gatech.e du/faq/Ellen.Ze gura/gt-itm/gt-itm.t ar.gz . Authors Mohamme d H. S qalli recei ved a degree of “Ingenieur d’Etat” in Comp uter Science fr om Ecole Mohammadia d’Ingenieurs, Rabat, Morocco in 1992. He earned a Maste r’s degree in C omputer Science in 1996 and a Ph.D. degree in Engineerin g - S ystems D esign i n 2002, both from the Universit y of Ne w Hampshire , Durha m, NH, U SA. He i s a recipient of a Fulbright Sch o larship f or the period of 19 94-1998. Mohammed H. S qalli is currentl y an a ssistant pr ofessor in the C omputer Engineeri ng De partment at KFUPM. He i s als o an IEEE member . His re search intere sts inclu de: Netw o rk Sec urity, Cloud Computing, Networ k Design and M anagement , Tra ffic Engineering, and Itera tive H euristics. He has over 2 5 publicati ons in relate d areas. Sadiq M. Sait obtai ned a Bachelor 's degree in Electronics from Bangalore University i n 1981, and Mas ter's and P hD degrees i n Electrica l Engineeri ng from King Fahd Universit y of Petroleu m & Minerals (KFUP M), Dhahran, Saudi Ar abia i n 1983 & 1987 respectivel y. Sadiq M. Sait is the co-a uthor of t he book V LSI PHYSIC AL D ESIGN AUTOMATION: Theor y & Pra ctice, published by M cGraw-Hill Boo k Co., Europe, (and als o co-published b y I EEE Pre ss), January 1995, a nd ITERA TIVE COMPU TER ALGORITH MS with APP LICATIONS i n EN GINEE RING (S olving Combinat orial Optimization Problems): publishe d b y IEEE Comp uter Society Pr ess, California, USA, 1999. He was the Chairman of Computer E ngineering Dep artment, KFUP M from J anuary 2001 - Decembe r 2004. Presentl y he is the Director of Information Technol ogy Center (IT C) at KFUP M, since J anuary 2005. S y ed Asadulla h received a Bachel o r of Technolog y ( B. Tec h) degree in El ectronics an d Communica tions Engi neering from Jawa herlal Nehru Techn ological Universit y , Hyderaba d, India in 2 000. He als o obtained his Master’s degree in Co mputer Netw orks from KFUPM, Saudi Arabia i n 2008.