Bandwidth Protection in MPLS Networks Using p-cycle Structure

Transcription

1 Bandwidth Protection in MPLS Networks Using p-cycle Structure Jianghui Kang, Martin J. Reed Department of Electronic System Engineering University of Essex Colchester, CO4 3SQ, United Kingdom {jkangf, Abstruct- This paper addresses the routing of backup tunnels with bandwidth guarantee in MPLS networks. Backup tunnels implemented by the packet layer in MPLS are seen as a useful alternative (or addition) to the schemes offered by a SONETBDH layer in particular for providing protection for real time packet services such as voice over IP. Many schemes have been proposed, working with MPLS, to provide the switching architecture for backup tunnels. However, the routing of the tunnels in these schemes is not clearly defined and hence the motivation for this work. This paper investigates the use of the p-cycle scheme to route the bandwidthguaranteed backup tunnels, a scheme that was originally devised for optical transport networks and IPlMPLS networks as well. We demonstrate that the problem (as for all but trivial p-cycle schemes) is too large to solve optimally and propose a relaxation method that pre-selects candidate cycles that are likely to provide a good solution. Furthermore, we show that working load distributions can significantly affect the performance of a p-cycle scheme and compare two routing algorithms to show how good distribution of working load can help increase the efficiency of the system. Index Terms-MPLS, facility backup, p-cycle, capacity planning, heuristic I. INTRODUCTION Fast backup protection in the order of 50 ms has long been a feature of circuit oriented networks such as SONETlSDH optical transport networks and is widely used and specified for carrying communications traffic such as voice. Packet networks working at a higher layer, in particular Internet protocol networks, have often resorted to slower restoration schemes that provide low-cost solutions to failure but with much slower time frames (usually measured in minutes). However, there is increasing interest in using packet based systems to provide real-time service transport working with quality of service techniques, for example voice over IP. For these systems to provide features comparable with a circuit oriented transport network, fast backup in the order of tens of milliseconds (i.e. protection or fast restoration) is required [l]. One solution to this is This work is supported by the EPSRC under the project GR/R05413/01 Audio over next generation Internet protocol networks to use MPLS or Multi Protocol Label Switching as the switching mechanism as it provides an efficient way to overlay connectionless IP traffic over a connection oriented layer. In practice, MPLS networks may be enabled over SONET using Packet-over-SONET technology. While, the SONETlSDH layer can provide reliable and fast protection in the occurrence of faults, at a cost, recent work has suggested that protection in the MPLS layer may be more cost effective. In addition, some envisage that the various layers existing today will be replaced by a simplified integrated IP over optical architecture using MPLS as an enabling technology. Consequently, the functions provided by the disappearing middle layers must be implemented in these two layers. This makes the investigation of the protectiodrestoration function in MPLS layer very important and routing of backup paths within the MPLS layer is the subject of this paper. In [l], a facility backup scheme for MPLS is proposed that uses one backup label switched path (LSP) tunnel to protect a set of LSPs. The former LSP is called a labelstacked bypass tunnel, which can protect all the LSPs passing its two end nodes (local repair points) by label stacking. When a fault occurs, all affected packets will be pushed in an additional label representing the bypass tunnel and redirected to the bypass tunnel at the local repair points. At the penultimate node, the additional label will be popped off the stack and the packets will be switched along the original LSPs. As the local repair points and detection points are very near to the fault, this scheme has fast rerouting speed. However, the design and management of these bypass tunnels are still not very clear although constrained shortest path first (CSPF) has been proposed to compute the tunnels. The p-cycle scheme has been proposed for protection in WDM or SONET networking and for the IP layer as well [2]. Here, we adapt the pcycle concept to the MPLS layer where it controls bypass tunnels engineered using label stacking under the assumption that all the LSPs in the network have certain bandwidth requirements and a backup tunnel must provide guaranteed bandwidth to the LSPs protected by it. Although the facility backup using /03/$ IEEE 356

2 p-cycle scheme can provide both node and link protection, in this paper we will concentrate on link protection. The scheme proposed in this paper only provides protection to the traffic in the occurrence of a link failure as the fast part of the whole restoration process. It should cooperate with an interior gateway routing protocol that should continue to act according to its own protocol to be the slower part of the whole restoration process [2]. This two-part scheme should be able to achieve both the reduction of interference to the traffic and good capacity efficiency. This paper first describes MPLS and the pcycle scheme. Our original studies are presented in Section I11 and IV. In Section 111, the paper states the constraint formulation used in this work to route the bandwidth-guaranteed backup tunnels, followed by descriptions of how the solution search space is reduced and the results of investigation into routing and subsequent load distribution. Section IV investigates the work applied to a realistic network and the scalability of the relaxation method is also discussed. Note that the MPLS-layer pcycles have been first proposed in [2] and our work is thereby motivated to pursue aspects of that approach further. 11. OVERVIEW OF MPLS AND p-cycle SCHEME MPLS integrates the advantages of both layer 2 and layer 3 and has better support for QoS and traffic engineering than traditional IP networking. In MPLS networks, packets will be assigned labels at ingress points. Then these packets with a same label will be forwarded along the same labelswitching path (LSP). These LSPs can be thought of as traffic trunks that carry aggregated traffic according to forwarding equivalence classes (FEC). Service providers can use explicit routing to create bandwidth guaranteed LSPs and use them to provision bandwidth guaranteed services to their customers. In order to provide certain guarantee to these customers, the service providers need to provide certain protection mechanisms to ensure these bandwidth guaranteed LSPs are quickly protected upon failures. Here we deploy the pcycle structure to manage label-stacked bypass tunnels to provide bandwidth guaranteed, quick protection to LSPs. The pcycle scheme was originally proposed for fault restoration in transport networks (WDM or SONET) [3]. The essence of this scheme is that pre-configuring the spare capacity of a network into certain patterns can achieve better restoration performance than others and in [41,[6] it is proven that pre-configured cycles have the best performance among trees, linear segments, closed cycles and the arbitrary mixtures of all these patterns. Furthermore, in [3] it is found that pcycles scheme can achieve 100% restorability with no or little increase of the spare capacity in meshbased restorable networks while remaining the BLSR-like restoration speed. In fact in [4] it is proven that pcycles have the same lower bound on the ratio of spare to working capacity as a span-restorable mesh network, which is =&, where 2 is the average span degree of the nodes in the network. It also shows that larger pcycles tend to provide higher capacity efficiency, from which we derive a heuristic algorithm for pre-selecting the candidate cycles in Section 111. Then in [2] it is demonstrated that this scheme can also be used in IPIMPLS networks to provide restoration in IP layer, still achieving full restoration almost without more spare capacity than optimised span restorable mesh networks. In order to make the difference between the spare capacity design problems in transport networks and MPLS networks with bandwidth guarantee clear, a MIP formulation (1)-(4) for spare capacity design in transport networks is presented below, which has been developed in [2] for the deployment of p-cycle in a SONET or WDM span restoration context. We can compare it with the formulation (5)-(9), which is adapted from a formulation for the spare capacity design problem of packet switching networks in [2]. First a set of all elementary cycles is generated from the network graph. Each elementary cycle can have certain number of copies of p-cycles that have one unit of capacity, e.g., an add-drop or cross-connection signal unit. The MIP will generate an optimal pcycie plan by determining the number of pcycle copies of each elementary cycle. Where the set of all the elementary cycles of the network graph is denoted as P, E is the set of all network spans, S = IEJ is the number of network spans. sj and wj are number of spare and working links on span j, respectively. ni is the number of unit capacity copies of elementary cycle i, pi,j is the number of spare links required on span j to build a copy of cycle i. xi,j is the number of paths that a single copy of p-cycle i provides for restoration of span j, cj is the cost or length of span j. The coefficients pi,j and zi,j can be determined in advance for each cycle in P. Then the formulation will be solved to determine the value of each ni, which will provide an optimal pcycle spare capacity design for the given network. p-cycles can be used to provide bandwidth-guaranteed 357

3 fast restoration in packet based MPLS networks as proposed in [2] in the same logical way as in SONET and WDM networks since we can manage the bandwidth of a packet MPLS network explicitly by using explicit routing and setting up bandwidth-guaranteed LSPs. But, unlike SONET or WDM networks, there is no distinct concept of unit spare capacity. This means that we cannot set up a certain number of copies of unit capacity p-cycle to provide restoration for the links with exactly the same capacity. Therefore we cannot directly use the above formulation designed for WDM or SONET networks to solve the pcycle optimisation problem in MPLS context. A method to adapt the circuit-switched p-cycle solution to design problems in IP networks is proposed in [2]. This considers the effects of restoration as an over-subscription on a link when affected traffic is moved onto new paths after a failure occurs, which follows naturally from the statistical multiplexing properties of IP networks and that sources will back-off in presence of over-subscription with reduced throughput but not total failure. Consequently, the objective function, which is to be minimised, measures the maximum over-subscription ratio on any link during a restoration event. As we need to provide bandwidthguaranteed protection here, this formulation also cannot be used directly and will be adapted in next section PROVISION OF BANDWIDTH PROTECTION IN MPLS NETWORKS USING p-cycles A. Spare Bandwidth Design Problem In order to provide bandwidth guaranteed restoration for the LSPs with certain bandwidth requirement, the formulation in [2] is adapted as following: NC min 2 ci si i= 1 ai,jlbj Vi=1,2,..., Ns,j=1,2,..., Nc i=l NC j=1 NC si 2 x,&,j,k * wj. aj,k k=l Vi = 1,2,..., Ns,j = 1,2,..., Ns,i # j Where Nc is the number of elementary cycles of the network graph, NS is the number of links of the network, bj is 1 if cycle j is selected in the design and 0 otherwise, ai,j is 1 if cycle j is used to protect link i and 0 otherwise, Pi,j,k is the link load ratio on link i when cycle k is used to protect link j. Pi,j,k is 1 if link j is on cycle k and 0.5 if link j is straddling cycle IC, it will be 0 otherwise. The different values of &,k for on-cycle links and straddling links imply the different protection schemes for these two kinds of links. Here we assume that an on-cycle link is protected by only one bypass tunnel formed by the remaining part of the cycle while a straddling link is able to utilize both parts of the cycle as the bypass tunnels and the traffic of the failed link can be evenly diverted into these two tunnels. Si and Wi are the total amounts of spare capacity reserved for restoration and working traffic on link i, respectively. ci is the cost of link i. (6) ensures that link i uses cycle j only when cycle j is selected in the design. (7) requires cycle j to be selected in the design only if it is used to protect any link. (8) ensures that each link will be protected by only one cycle and (9) ensures link i can provide enough guaranteed bandwidth even in the worst situation. The objective function is to minimise the weighted backup capacity for protection. Here we consider only single link failure, as it is assumed that the probability of multi-link failures is small with fast restoration in operation. Furthermore, we assume that the LSPs and pcycles are all bi-directional and have equal priority. As discussed above, all the LSPs passing a link will be protected as a single entity in the bypass tunnel formed by a pcycle, which leads to a relatively coarse protection granularity. Finer protection granularity is likely to improve the capacity efficiency further and provides more flexible protection service, e.g., more protection priority classes and different LSPs passing the same link can be protected by different cycles. This certainly will make the problem much more difficult to solve. However, as our main aim is to investigate the capacity efficiency of the pcycle scheme in the management of bandwidth-guaranteed bypass tunnels, this paper concentrates on the spare capacity design problem under the assumption of single priority, link-based protection. As the pcycle optimisation design is an NP-hard combinatorial optimisation problem [2], some relaxation or offline pre-selection of cycles is necessary in order to make its application practical. In this paper we consider the preselection of elementary cycles to reduce the size of the problem. In Section III.B, a pre-selection method to find an effective subset of cycles from all elementary cycles is proposed. All the work presented here was conducted by solving the above formulation (5)-(9) using linear program solver GLPK3.1, given the network topology, the traffic demand matrix and the allowed size or size range of the candidate elementary cycles. The output is the selected optimal set of cycles that will be used as pcycles and the corresponding links protected by them. It is worth noting that although all the work in this study was conducted 358

4 - I I I I I 1.6 Topology 1- Topology Topology Topology ROIJWX D"lrrpp E r m i Size of Cycles Figure 1 - BCR of Different Size of Cycles of 7 Topologies. ROWER 0 Figure 2 - Effect of Cycle Size with Unbalanced Load. by solving MIP, the proposed scheme is not restricted to the MIP method, especially for large networks. Other algorithms, e.g., genetic algorithm (GA) may be more suitable than the MIP method when the network is large. This will be considered in detail in Section IV. B. Reduction in search space through cycle selection We performed our first set of experiments on seven randomly generated networks with a uniform traffic demand and the results obtained from the experiments are shown in Figure 1. The randomly generated networks have an average size of 9 nodes and the average node degree varies between 3.8 and 4.2. The objective of this set of experiments is to investigate the relationship between the BCR (Backup Capacity Ratio), which is the ratio of backup capacity for restoration and working capacity of the network, and the size (in hops) of candidate cycles. For this set of experiments, the link capacities were set to infinity and we assumed there was unit bandwidth LSP between each pair of nodes. Then we used the formulation mentioned above to obtain the optimal pcycle design within certain hop limit, i.e., the cycles with certain hop number. Here the range of the size of cycles is from 3 to 9 hops except for network 3, where the largest cycle has 8 hops. As shown in Figure 1, in most cases the BCR will decrease with the increase of the size of the candidate cycles. This means that large cycles have higher efficiency in these cases, which has been substantiated theoretically in [4]. However, in two of the seven networks there are contrary results, the BCR increases when larger cycles are deployed. After careful analysis we found that it was because of the following reasons. The first reason is that too much working traffic passes the same link in these two networks because of the routing algorithm (pure shortest path routing) that we deployed in the experiments. This means the working load is not distributed evenly. So if we deploy larger cycles to protect the heavily loaded link, more backup capacity will have to be reserved than using smaller cycles in this case. Figure 2 depicts this. Let us assume link Router A, Router C is heavily loaded. If we use cycle A, B, D, C, A to protect link A, C, we only need reserve the same amount of capacities on three links. But if we select cycle A, B, E, F, G, D, C, A to protect link A, C, we will have to reserve the same amount of capacities on six other links. And as the link A, C is much more loaded that any other link, the large amount of backup bandwidth reserved for it on the cycle can not be shared fully by other links. This will degrade the capacity efficiency of pcycle scheme greatly. There will be overloading in some parts of the network while under-utilization in other parts of the network. This can be avoided in MPLS networks by traffic engineering and by deploying more efficient routing algorithms that can balance the working traffic distribution. The second reason is that the number of candidate large cycles is small for these two networks. The limited number of candidate cycles will lead to less flexibility thus decreasing the efficiency of the solution. These two problems will be considered carefully later in this section. Following the arguments above, we conjecture that larger cycles, as a general trend (but not always), have better capacity efficiency performance overall than smaller ones. Consequently, in the ideal case, the pcycle design problem can be simplified by reducing the number of candidate cycles (searching space) to just that of large cycles while not decreasing the overall capacity efficiency. Having established that large pcycles are optimum in the ideal case of uniform traffic distribution across all links, but, that in the non-ideal case of non-uniform traffic distribution this may not be true, we investigate the effect of the routing algorithm (and the subsequent traffic distribution) on the capacity efficiency. Consider a fully connected network with N nodes as shown in Figure 3. In order to make the analysis simpler, without loss of generality assume that one of the on-cycle links has working traffic of z while all of the other 359

5 BCR of Different a CL Figure 4 - BCR of Different Load Ratio a. Figure 3 - Relationship between BCR and Working Traffic Distribution. N- 1 on-cycle links carry working traffic of y where x > y. Furthermore, assume that all the - N straddlingcycle links have working capacity of 2y. In this case, the protected working capacity C can be derived as N(N - 1) c = x + y(n - 1) + 2y[ -NI (10) while the reserved backup capacity B will be B = z(n - 1) + y (11) Thus after simplifying (10) and (ll), the BCR can be derived as BCR B x(n - 1) + y = - (12) c z+y(n2-2n-l) Let a = f then N-l+a BCR = 1 + a(n2-2n - 1) (13) We plot the computed result for BCR in Figure 4 for N = 9,15,21 with a from 0.1 to 1. The figure shows that for a fixed N, the BCR will decrease with an increase of a and for a fixed cy the BCR will decrease with increased N. This implies that the capacity efficiency of the pcycle will decrease when the difference between x and y becomes bigger and will increase when the size of cycle becomes larger. This result confirms that the general trend is for large cycles to have improved capacity efficiency but that nonuniform traffic distribution limits this. In order to diminish the efficiency degradation caused by reducing the number of candidate cycles (searching space) to just that of the largest cycles in the cases where the traffic distribution is non-uniform, some smallest cycles should also be deployed to mop up the working traffic of heavily loaded links in localized points. Based on above analysis, the heuristic for the pre-selection of elementary cycles can be described as following. Let M be the number of sizes of the elementary cycles in the network, the cycles will have one of the following sizes: SI, 5 2,..., SM, where SI < S2 <... < SM. The set of cycles with size S is represented by Cs. So there are M sets of cycles: Csl, Cs,,..., Cs,. According to above analysis, we can reduce the searching space of the problem to the set of the largest cycles plus the set of smallest cycles, i.e., Csl U CsM. While this heuristic chooses the largest available cycles in networks, in practice large cycles may be a poor design due to practical implementation. For example, a large cycle is more susceptible to multiple failures, may introduce unacceptable delay (on the unacceptable number of nodes in the bypass path) and/or may result in signal degradation in optical transmission system. Clearly, for a practice implementation, the largest allowable cycle size may be set from the network policy. In this paper we ignore this. The simple heuristic presented here provides part of a solution to an unbalanced network load. An unbalanced traffic distribution is highly likely in a practical network as the traffic demand across network ingresdegress could differ and links could be deployed with differing capacities. However, the routing algorithm used in the network can also affect traffic distribution. Consequently the effect of routing together with the improvement of using the described cycle selection heuristic will be described together in the next section. C. The effects of routing algorithm on eflciency Here we investigate the effect of the simple cycle selection heuristic and the effect of the routing on the capacity efficiency measured using BCR. Two different routing algorithms were investigated: pure shortest path routing algorithm and a dynamic weight shortest path routing algorithm. The former is here simply termed shortest path (SP) algorithm. In this case, when a demand pair appears the shortest path between the source and destination nodes will 3 60

6 Network 1 TABLE I THE EFFECT OF ROUTING ALGORITHM ON BCR sc LC SLC SP I DWSP SP I DWSP SP I DWSP I I I I I I I I I I I I I SC = Smallest Cycles, LC = Largest Cycles SLC = Smallest + Largest Cycles be computed using Dijkstra algorithm, where the weights of links are fixed values. The second routing algorithm is termed here dynamic weight shortest path (DWSP), where the weights of the links are dynamically determined according to the working load of the links. In [5] a new LSP computing algorithm is proposed to keep the interference to demand routing of later demand pairs minimum. Thus it can achieve better load balancing than the shortest path algorithm. But this routing algorithm requires intensively solving Max-flow problem. So in 171 a new algorithm is proposed based on the one in [5] to reduce computation complexity and bound the lengths of the solutions as well. Here in our experiment, we adapted the algorithm in [7] to make it suitable for routing the demand pairs in an incapacitated network with complete working traffic information. This routing algorithm can achieve better working load balancing to improve the capacity efficiency of p-cycle further. In this experiment, four randomly generated network topologies with average node degree of 4 were deployed with randomly-generated traffic demand of uniform distribution. In addition, in order to test our conclusion about the size of pcycle, the experiment used different pre-selected candidate cycle sets, i.e., candidate cycle sets consisting of only the smallest elementary cycles Cs, (SC), only the largest cycles Cs, (LC) and the smallest plus the largest cycles Cs, U Cs, (SLC) respectively. Table I shows the results of the experiment regarding the above conclusions. From Table I we can see that in all the networks, DWSP algorithm has better performance (lower BCR) than SP algorithm regarding capacity efficiency for different sizes of cycles. For example, in Topology 1, the BCR can be reduced from to in the SLC case (our proposed algorithm), which can lead to a significant cost reduction. This is entirely consistent with our earlier conjecture. The table also provides the capacity efficiency of candidate cycle sets of SC and LC and SLC. It shows that large cycles indeed have better performance than their smaller counterparts and largest plus smallest cycles can achieve the best performance. For example, with same routing algorithm DWSP, the BCR of Topology 4 is decreased from to Figure 5 - CN Core Network of Project Cost when SC is replaced by LC and is reduced further to when LC is replaced by SLC. IV. REALISTIC NETWORK EXAMPLE In this part of this paper we will deploy the conclusions obtained from above experiments and analysis to test the performance of this scheme with a realistic network topology and traffic demand. The network topology that we used here is proposed by CN Core Network of project Cost 266 as shown in Figure 5. The traffic demand is the average Internet traffic of year 2002 on this network and given in a form of Source, Destination, and Bandwidth. The first thing to do is to route these working demands on the network as working flows with certain bandwidth requirements. As discussed above, in order to implement p cycle scheme more efficiently, we need a routing algorithm with good load balancing ability. So in our experiment we adopted the idea in [9] to control the weights of links. In [9] in order to optimise the weight setting for an AT&T World backbone with projected demands, the authors use a local search heuristic to solve the optimisation problem and get the weight settings for links that are very near to the values of optimal general routing method. The method has excellent performance in improving the network utilization and reducing congestion and link overloading by deciding the weights of links according to load and giving more penalties to higher loaded links. We modified the parameters of this method according to our own demand and network capacity. First let us see how our method can reduce the complexity of the optimisation problem. Table I1 is the distribution of cycles of all sizes existing in this network topology. Using the notation given in Section III.B, Table I1 gives M = 12 (there are 12 sizes of cycles), S1 = 4 (the minimum cycle size), 5 12 = 15 (the maximum cycle size). Without 361

7 TABLE I1 THE DISTRIBUTION OF ELEMENTARY CYCLES IN FIGURE 5 (TOTAL NUMBER OF CYCLES IS 118) appropriate measures, the solution space of the optimization problem is the combination of all 12 sets of cycles, which have 118 cycles altogether and thus all of them will make Z118 possible combinations. This will make the application computationally intensive. But, if our proposed method of reducing the solution space is used, we need only consider the cycles of set Csl U Cslz, which has only 8 cycles in this case. Clearly, the cycle selection heuristic proposed in Section 1II.B reduces the searching space to a size that is implementable and following our methodology the smaller candidate cycles will only be included if they form an advantage, hence, it will not degrade performance compared to a simple large cycle selection scheme. The capacity efficiency is depicted by the ratio of spare capacity reserved for backup and the working load, which is about 93% in this case. Further work being carried out at the moment is to investigate the optimisation of the routing algorithm such that the capacity efficiency can be further increased. For the design problem of even larger networks of several tens of nodes, there is more to consider. Firstly, as mentioned in Section III.B, the term the largest cycles have to be replaced by the largest allowable cycles to avoid utilizing cycles of too many hops. From our analysis it is obvious that in this case the proposed method can still achieve relatively high capacity efficiency. Secondly, as having been found in our experiments, for some large networks, even after relaxation using our pre-selection heuristic, the search space may still be too large to be solved by an MIP solver. In this case, other effective algorithms can be deployed to solve the design problem and the pre-selection heuristic can still provide an effectively reduced search space to speed up the solving of the problem. Because of the complexity of the pre-selection algorithm is determined only by the complexity of the algorithm that is used to enumerate all the elementary cycles in the network and polynomial time cycle-enumeration algorithms are available, the proposed method is scalable. bandwidth-guaranteed bypass tunnels in packet MPLS networks. In order to reduce the complexity of the problem, we investigated the reduction of the solution space using a cycle pre-selection scheme that produces good results. The paper investigated the application of this scheme to a realistic network scenario that has demonstrated that our scheme can achieve very good performance without intensive computation. The relation of capacity efficiency and load distribution has been further investigated and it was found that load balancing (a feature of MPLS traffic engineering) can further improve the efficiency of this scheme. Finally, we conclude that using adapted pcycle scheme with the proposed relaxation method provides an excellent scheme for the routing of bandwidth-guaranteed bypass tunnels in MPLS networks. REFERENCES Ping Pan et al., [Online], Fast reroute extensions to RSVP-TE for LSP tunnels, IETF Intemet Draft, Available: 15 August 2002 [date accessed]. D. Stamatelakis and W.D. Grover, IP layer restoration and network planning based on virtual protection cycles, IEEE Journal on Selected Areas in Communications, vol. 18, pp , October W.D. Grover and D. Stamatelakis, Cycle-oriented distributed preconfiguration: Ring-lie speed with mesh-lie capacity for self-planning network restoration, in Proceedings of ICC 98, pp , Atlanta, Georgia, June 07-11, D. Stamatelakis and W.D. Grover, Theoretical underpinnings for the efficiency of restorable networks using pre-configured cycles, IEEE Transactions on Communications, vol. 48, pp , August M. Kodialam and T. Lakshman, Minimum interference routing with applications to MPLS traffic engineering, Proceedings of IEEE IN- FOCOM 2000, vol. 2, pp , Tel Aviv, Israel, March 26-30, D. Stamatelakis and W.D. Grover, Network restorability design using pre-configured trees, cycles and mixtures of pattem types, TRLabs Canada, Technical Report TR , Issue 1.0, October 20, G. Baneqee and D. Sidhu, Path computation for traffic engineering in MPLS networks, in Proceeding of First Intemtional Conference on Nefworking (ICN ZOOI), Part 11, pp , Colmar, France, July 09-13, G. Hjalmtysson, P. Sebos, G. Smith, and J. Yates, Simple IP restoration for IP/GbWlOGbE optical networks, OFC 2000, Baltimore, MD, March B. Fortz, M. Thorup, Intemet traffic engineering by optimising OSPF weights, Proceedings of IEEE INFOCOM 2000, vol. 2, pp , Tel Aviv, Israel, March 26-30, V. CONCLUSION In this paper, using p-cycles to manage and control the bandwidth-guaranteed backup bypass tunnels in MPLS networks has been investigated. The study is based on the p-cycle scheme proposed originally for transport networks and IPMPLS networks but here is adapted for use with 362