Network Protection Design Models, a Heuristic, and a Study for Concurrent Single-Link per Layer Failures in Three-Layer Networks

Transcription

1 Network Protection Design Models, a Heuristic, and a Study for Concurrent Single-Link per Layer Failures in Three-Layer Networks Iyad Katib a,, Deep Medhi b a King Abdulaziz University, Jeddah, Saudi Arabia b University of Missouri Kansas City, USA Abstract Multilayer network design has received significant attention in current literature. However, the explicit modeling of IP/MPLS over OTN over DWDM in which the OTN layer s technological constraints are specifically considered has not been investigated before. In this paper, we present an optimization design model for protecting an IP/MPLS over OTN over DWDM three-layer network. While considering the technological constraints of each layer, we provide a protection mechanism at each layer that guarantees the multilayer network survivability when three links fail simultaneously where each layer suffers a single failure. We present a heuristic approach to reduce the complexity of the problem and present a study based on varying several network parameters to understand their impacts on the protection capacity and the overall network cost. In addition, we present and solve three variations of our original model where we exclude each layer protection in each one of them to compare the cost performance of all models. We observe that generally the DWDM layer protection is the most expensive capacity component. The IP/MPLS layer protection becomes more expensive only when the IP/MPLS unit cost is high. Keywords: DWDM, IP/MPLS, Multilayer Network, Network Optimization, OTN, Protection, Survivable Design. 1. Introduction Multilayer network survivability has been a crucial research topic in recent years as network traffic is unstoppably rising. A survivable network, in general, is a network that provides some ability to recover ongoing traffic disrupted by a network failure. Large ISPs need to ensure their networks can meet customer satisfaction and expectations. In addition, in today s world where businesses rely heavily on computer networks, network failures can severely affect their revenues. Thus, network survivability has always been a vital factor in designing current and future communication networks. In two-layer networks such as IP-over-WDM, a single recovery mechanism could be provided at either layer. In this design, a critical question arises: where do we provide the protection mechanism? The benefits of an upper layer protection are: (1) in case of failure (either at the upper or lower layer), the network could be fully recovered, (2) since the upper layer often carries Corresponding Author addresses: Iyad.Katib@gmail.com (Iyad Katib), DMedhi@umkc.edu (Deep Medhi) differentiated services with different QoS requirements, it is generally easier to offer differentiated survivability at the upper layer. Nonetheless, recovery at the upper layer has some disadvantages: (1) recovery time at the upper layer is usually higher than recovery time at the lower layer due to the nature of the IP, (2) in case of failure at the lower layer, there could be a huge amount of upper layer traffic affected by the failure in which case, a great amount of recovery process at the upper layer is required. On the other hand, recovery at the lower layer has some advantages. It is faster than recovery at the upper layer and it requires considerably fewer actions due to the coarser granularity of the lower layer. The drawback, however, is that some failures (e.g. an IP router failure) cannot be handled by the lower layer. The above discussion elucidates the need for a recovery mechanism to be deployed at each layer of the network to recover from various network failures. Despite the extensive research devoted to multilayer networks, the explicit modeling of IP/MPLS-over- OTN-over DWDM in which the OTN (optical transport network) layer is specifically considered has not been addressed before except in our previous work [1, 2, 3, 4]. We can visualize the OTN layer as being the virtual Preprint submitted to COMPUTER COMMUNICATIONS October 7, 2012

2 optical layer on top of the physical DWDM layer. This architecture, that bridges integration and interaction between the IP and optical layers [5], has been identified as promising. In this paper, we consider a survivability design specifically for a three-layer IP/MPLS-over-OTN-over- DWDM network where only the normal flow of each layer is 100% protected against any single link failure. This leads to a survivable network in case of three concurrent link failures where each layer suffers a single failure. In this architecture, the label switched routers (LSRs) in the IP/MPLS layer are physically connected to optical transport networks that are slated on top of optical cross-connects (OXCs) that are interconnected by a DWDM fiber transmission medium at the physical level. In this setting, we present the network capacity (Normal-protection) design model and a study based on various network parameters. The contributions of this work are as follows: The explicit consideration of each layer in the network and its modularity requirement, including the OTN layer that results in a three-layer model. The consideration of the OTN layer sublayer technological constraints. Note that although each OTN signal quantum (see Section 3 for U k ) can form its own virtual network (sublayer), we can consider them together without considering each sublayer separately because of the way the costing is defined in the objective function; this reduces the number of constraints considerably. The design that is based on the separation of capacity components at each layer to avoid double or triple protection of the upper layer capacity. A study and analysis of results obtained by solving four different network protection design models based on various network parameters to understand their impacts on three-layer networks. 2. Related Work For multilayer network survivability, the problem is how to design a survivable multilayer network with two goals in mind: (1) to maximize the network protection and, (2) to reduce the cost of the network resources. Several survivability mechanisms have been discussed in literature for two-layer networks [6, 7, 8, 9]. The most traditional approach is the redundant protection. In this case, the spare capacity of the upper layer is twice protected; once in the upper layer, and once in the 2 lower layer. Clearly, this leads to a poor utilization of the expensive network resource. A cost reduction can be achieved in this design if the protection (spare) capacity of the upper layer is left unprotected in the lower layer. Fumagalli and Valcarenghi, in [10], review the most common restoration and protection mechanisms available at the IP and WDM layers that can be implemented concurrently in the IP over WDM architecture. Sahasrabuddhe et al., in [11], address the problem of in which layer to provide the fault-management technique (either the IP or WDM layer). Kubilinskas and Pióro, in [12], also present two design problems providing protection in either the WDM layer or the IP layer. Zhang and Durresi, in [13], investigate the necessity, methods, and advantages to coordinate multilayer survivability in the IP over WDM networks. Lei et al. in [14], and Qin et al. in [15], investigate and study joint multilayer survivability in IP/WDM networks. In [16], Bigos et al. compare single layer and multilayer survivability in MPLS over optical transport networks. In [17, 18], the impact of GMPLS on multi-layer survivability has been addressed. We note, however, that in all previous works, the OTN layer that imposes unique technological constraints is not explicitly considered. We observe that usually when an OTN is mentioned, a reconfigurable optical backbone is meant. That is, the core routers are connected through electro-optical cross-connects (OXCs) with no consideration for OTN as a distinct layer. Design models based on this assumption are missing the sub-layer technological constraints within the lightpath. Thus, our capacity design model is more realistic and precise because it goes into the middle layer s (OTN) technological constraints and models its associated sub-component cost, allowing us to investigate its impact on the network. In [4, 19], we have laid the foundation behind the three layer networks including how the three layers interact with each other in operation and provided detailed motivation to aggressively consider this three layer architecture. In [4], we addressed the integrated design problem without survivability. In this paper, we extend our previous work [3] on protection design of threelayer networks. Mainly, we emphasize the benefit of employing a protection mechanism in each layer leading to a fully survivable network in case three links - possibly more - of different switching capabilities fail simultaneously. We also present three different variations of our original comprehensive protection design model. These models differ based on which layers to employ protection. While our main model is overencompassing of all the model variations by providing protection at every layer, in the three sub-models

3 Table 1: OTN Signals, Data Rates and Multiplexing. U k Signal Bit-Rate (Gbps) Max. U k s in a wavelength U U U U U we exclude a protection component from one layer at a time. That is, the first variation has no DWDM layer protection, the second one has no OTN layer protection, and the last variation has no IP/MPLS layer protection. Then, we solve all models and compare their cost performance to investigate each layer protection from a cost standpoint. We also use another real-world existing topology to analyze the results obtained by solving the models. 3. OTN Overview To support network management and supervision functionalities, the OTN system is structured in layered networks consisting of several sublayers [20]. Each sublayer is responsible for specific services and is activated at its termination points. For this paper, we are interested in the Optical Data Unit (ODU) sublayer that provides (1) tandem connection monitoring, (2) end-to-end path supervision, (3) adaptation of client data that can be of diverse formats such as SDH, Ethernet, IP/MPLS, T- MPLS, and so on. The ODU sublayer currently defines five bit-rate client signals, i.e., 1.25, 2.5, 10, 40, and 100 Gbps that are referred to as ODUk (k = 0, 1, 2, 3, 4), respectively. In the rest of the paper, U k denotes ODUk for k = 0, 1, 2, 3, 4. Then, for the multiplexing process, we can write: 2U 0 = U 1, 4U 1 = U 2, 4U 2 = U 3, and 2U 3 = U 4 (see Table 1 rates and how these fit into a wavelength). OTN also defines the ODUk time division multiplexing sublayer. It supports the multiplexing and transporting of several lower bit-rate signals into a higher bit-rate signal and maintains an end-to-end trail for the lower bit-rate signals. This typically occurs when a client signal does not occupy an entire wavelength. The multiplexing of ODUk signals is easy to visualize from the the bit-rates shown in Table 1. There are two additional specifications: ODU2e and ODUflex. For the purpose of capacity planning modeling, ODU2e can be treated as ODU2, which is not considered separately. ODUflex is any rate over ODU1, which from a modelling purpose can be treated as a real variable with a lower bound of 1 Gbps. Since in our 3 model, any ODU modular variables can be relaxed to be real variables, ODUflex is not considered separately. 4. Protection Mechanisms Resource protection can be performed in different layers of a multilayer network. In our architecture, the IP/MPLS layer is protected in the underlying OTN layer that is protected by the DWDM layer. In this case, a failure in a lower layer cannot be seen by the upper layer. For instance, the IP/MPLS layer does not see the failure of the OTN link. Several protection and restoration mechanisms have been introduced in literature [10]. The choice of which method to implement in a network depends on the requirements of the ISP and whether a method is technologically meaningful. In this section, we present our selection of the protection mechanism used per layer of the multilayer network and explain why we selected each one of them. MPLS tunnels can be set up to carry demand volumes for different traffic demand types that require different QoS. This indicates that the MPLS layer can provide transport services through the use of tunnels. In our design model, we assume that each IP demand d can be carried over a single end-to-end primary tunnel. In this case, one of the suitable protection mechanisms from the service provider standpoint is the hot-standby path protection. In this method, a demand is carried over the primary path only, while the protection path is reserved for future use in case that the primary path fails. This is a 1:1 protection technique in which the primary and protection paths are link-disjoint. Note that the protection capacity for one path is not shared with the protection capacity used for other paths that fail in other failure situations. In addition, each failed flow is restored on one single protection path. Since each OTN link carries U k signals, we provide a protection for each U k by using a link restoration on a single path. In this mechanism, the entire capacity of the failed U k is restored on a single path between the end nodes of the failed OTN link. Note that in link restoration the protection capacity is shared by protection paths. However, primary and protection paths are link-disjoint. For the DWDM layer, we provide protection at the aggregate signal level. A common method of protection, at this lambda layer, is protection by using fixed backup paths. In this method, a copy of the data signal is transmitted respectively on a primary and a protection path that are link and node-disjoint. Based on the signal quality, the receiver can make a decision to accept which copy of signal. This is a 1+1 protection technique.

4 Figure 1: IP/MPLS over OTN over DWDM Network 5. An Integrated Capacity (Normal and Protection) Optimization Model Figure 1 shows how we approach the problem. First, we have an IP/MPLS layer normal capacity and its protection capacity. Both must be realized by the OTN layer. However, the OTN layer will only protect its normal capacity that is needed to realize the normal IP/MPLS capacity to avoid protecting the IP/MPLS layer capacity twice; one in the IP/MPLS layer and one in the OTN layer. Then, all OTN layer capacities will be realized by the DWDM layer. Again, only the normal capacity of the DWDM layer is protected to avoid protecting the OTN layer capacity twice; one in the OTN layer and one in the DWDM layer. Note that by separating the capacity components and protecting the normal capacity of each layer, the multilayer network can fully survive three link failures at minimum in case a single link fails in each layer simultaneously. Tables 2 and 3 list the notations used in our formulation Constraints Since protection will be provided to the normal capacity of each layer, we have separated the capacity components at each layer. In our formulation, there are two general sets of constraints. The first is the set of capacity feasibility constraints that assures all flows routed on a particular link do not exceed the capacity of the link. The second is the set of demand constraints that specifies how the capacity of each upper layer link is realized by means of flow allocated to its candidate paths from the routing list in the lower layer. Note that in our model capacity variables are expressed in terms of the number of modules of certain sizes. Next, we start introducing the design constraints. 4 Table 2: List of Notations (Given entities) Indices: d = 1, 2,..., D demands between source-destination pairs of the IP/MPLS layer. p = 1, 2,..., P d candidate pair of (primary, protection) paths (P dp, R dp ) for realizing demand d. e = 1, 2,..., E links of the IP/MPLS layer. q = 1, 2,..., Q e candidate paths of OTN layer for realizing capacity of link e. g, l = 1, 2,..., G links of the OTN layer. r = 1, 2,..., R g candidate restoration paths for link g. z = 1, 2,..., Z g candidate pair of (primary, protection) paths (Z g, A g ) of DWDM layer for realizing capacity of link g. v = 1, 2,..., V g candidate paths of DWDM layer for realizing capacity of link g. f = 1, 2,..., F links of the DWDM layer. k = 0, 1, 2, 3, 4. modular interfaces of OTN link g. Constants: h d : Volume of demand d. δ edp : =1 if link e belongs to the primary path P dp realizing demand d; 0, otherwise. µ edp : =1 if link e belongs to the protection path R dp protecting path P dp of demand d; 0, otherwise. γ geq : =1 if link g belongs to path q realizing capacity of link e; 0, otherwise. ϑ f gz : =1 if link f belongs to primary path Z g realizing capacity of link g; 0, otherwise. θ f gz : =1 if link f belongs to the protection path A g protecting path Z g of link g; 0, otherwise. π f gv : =1 if link f belongs to the path v realizing capacity of link g; 0, otherwise. lgkr : =1 if link l belongs to path r restoring OTN interface k on link g; 0, otherwise. M: Module size for IP/MPLS layer. U k : Module size for OTN layer link capacities k = 0, 1, 2, 3, 4. N: Module size for DWDM layer link capacities. η e : Cost of one capacity unit of module M of IP/MPLS layer link e. β gk : Cost of one capacity unit of type U k of OTN layer link g. ξ f : Cost of one capacity unit of module N of WDM layer link f. In this architecture, we assume that an IP demand d can be carried over a single pair of primary and protection paths ( pp-path-pair ) out of the set of candidate pairs of (primary, protection) paths that are numbered 1, 2,..., P d. Path pairs are obtained after computing a number of paths off-line using k-shortest paths based on Dijkstra s algorithm and fed to the model. We define x dp as a binary decision variable for selection of a pppath-pair for demand d. This can be expressed as in constraints (1). Constraints (2) are the capacity feasibility constraints of the normal flows routed on link e where M is the allowable granularity of each MPLS tunnel. Here, δ edp determines if link e belongs to the primary path P dp carrying the normal flow of demand d. Protection in the IP/MPLS layer is achieved using a hot-standby path for each primary path. Constraints (3) are the capacity feasibility constraints of the protection

5 Table 3: List of Notations (Variables) Variables: x dp : IP/MPLS flow allocated to path pair p (P dp, R dp ) of demand d (non-negative, binary). m eq : OTN flow allocated to path q realizing normal capacity of link e (non-negative integral). m eq: OTN flow allocated to path q realizing protection capacity of link e (non-negative integral). s gkz : DWDM flow allocated to path pair z (Z g, A g ) realizing normal capacity of link g of interface k (non-negative integral). s gkv : DWDM flow allocated to path v realizing protection capacity of link g of interface k (non-negative integral). s gkv : DWDM flow allocated to path v realizing OTN capacity of link g of interface k that realizes protection capacity of the IP/MPLS layer (non-negative integral). c gkr : flow restoring normal capacity of interface k of link g on restoration path r. u gkr : binary flow variable associated with c gkr. y e : Number of modules M to be installed on link e for normal capacity of the IP/MPLS layer (non-negative integral). y : Number of modules M to be installed for protection capacity e of link e. w gk : Number of modules U k to be installed on link g for normal capacity in the OTN layer (non-negative integral). w gk : Number of modules U k to be installed for protection capacity of link g (non-negative integral). w gk : Number of modules U k to be installed on link g in the OTN layer for realizing IP/MPLS layer protection capacity (nonnegative integral). b f : Number of modules N to be installed on link f in the DWDM layer (non-negative integral). b f : Protection capacity on link f in the DWDM layer (nonnegative integral). b f : Number of modules N to be installed on link f in the DWDM layer for realizing OTN layer protection capacity (nonnegative integral). b f : Number of modules N to be installed on link f in the DWDM layer for realizing OTN capacity that realizes IP/MPLS layer protection capacity (non-negative integral). flows on link e. Here, µ edp determines if link e belongs to the protection path R dp that protects the primary path P dp. Pd p=1 x dp = 1 d = 1, 2,..., D (1) Dd=1 h Pd d p=1 δ edpx dp My e e = 1, 2,..., E (2) Dd=1 h Pd d p=1 µ edpx dp My e e = 1, 2,..., E (3) Constraints (4) are the demand constraints that specify how the normal capacity of each IP/MPLS layer link e is realized by means of flow m eq and is allocated to its candidate paths from the routing list in the OTN layer. Similarly, Constraints (5) are the demand constraints of the protection capacity of the IP/MPLS layer. Qe q=1 m eq = y e e = 1, 2,..., E (4) 5 Qe q=1 m eq = y e e = 1, 2,..., E (5) The OTN layer s normal capacity feasibility constraints are expressed in (6). These constraints assure that all normal flows routed on each OTN layer link g do not exceed their capacity that is allocated in modules of sizes U k that represent the five modular interfaces of OTN. Likewise, constraints (7) are the OTN layer protection capacity feasibility constraints. M E e=1 Qe q=1 γ geqm eq 4 k=0 U k w gk g = 1, 2,..., G (6) M E e=1 Qe q=1 γ geqm eq 4 k=0 U k w gk g = 1, 2,..., G (7) Protection in the OTN layer is achieved using a link restoration on a single path. Constraints (8) (11) assure that only normal capacity of each link g can be restored using only the protection capacity of the remaining links l(l g) on a single restoration path r. Note that we avoid double protection of the IP/MPLS protection capacity by protecting only the capacity w gk required for m eq flow of the IP/MPLS normal capacity y e. Note that constraints (8) to (10) force that c gkr = u gkr w gk, but the right-hand side cannot be used directly in the formulation because it is a term containing a multiplication of two variables. Also, constraints (11) assure that normal capacity of each OTN interface k can be restored using only the protection capacity of the remaining links l(l g). Rg r=1 c gkr = w gk g = 1, 2,..., G k = 0, 1, 2, 3, 4 (8) Rg r=1 u gkr = 1 g = 1, 2,..., G k = 0, 1, 2, 3, 4 (9) c gkr U k u gkr g = 1, 2,..., G k = 0, 1, 2, 3, 4, r = 1, 2,..., R g (10) Rg r=1 lgkrc gkr w lk k = 0, 1, 2, 3, 4, l = 1, 2,..., G, g = 1, 2,..., G l g (11) Constraints (12) and (13) are the OTN over DWDM demand constraints for the normal, and protection capacity, respectively. They specify how the capacity of each OTN layer interface k of link g is realized by means of flow allocated to its candidate paths from the routing list in the DWDM layer. Note that we separated the normal capacity w gk from protection capacity w gk to avoid protecting the OTN signals twice, once in the OTN layer and once in the DWDM layer. Constraints (14) are the OTN over DWDM demand constraints for the OTN capacity required to realize the IP/MPLS protection capacity.

6 Zg z=1 s gkz = w gk g = 1, 2,..., G k = 0, 1, 2, 3, 4 (12) Vg v=1 s gkv = w gk g = 1, 2,..., G k = 0, 1, 2, 3, 4 (13) Vg v=1 s gkv = w gk g = 1, 2,..., G k = 0, 1, 2, 3, 4 (14) Protection in the DWDM layer is achieved using fixed back-up paths. Constraints (15) to (18) are the DWDM layer capacity feasibility constraints. They assure that the capacity of each physical link f is not exceeded by the flow using this link. Note that N is the module size of the DWDM layer link capacity that is equal to the wavelength capacity, and b f would be the normal number of wavelengths to be installed on link f. At this layer, we have four capacity components: b f for the normal DWDM layer capacity, b f for the protection capacity, b f for the capacity required to realize the OTN protection capacity, and b f for the capacity required to realize the OTN capacity that realizes the IP/MPLS protection capacity. Figure 1 shows all capacity components at each layer. G 4k=0 g=1 U k Zg z ϑ f gz s gkz Nb f f = 1, 2,..., F G 4k=0 Zg g=1 U k z=1 θ f gzs gkz Nb f f = 1, 2,..., F (15) G 4k=0 Vg g=1 U k v=1 π f gvs gkv Nb f (16) f = 1, 2,..., F (17) G 4k=0 Vg g=1 U k v=1 π f gvs gkv Nb f 5.2. Objective and Cost Model f = 1, 2,..., F (18) The goal in our design model is to minimize the total network planning cost of the normal and protection capacity. The objective is given by: F = E η e (y e + y ) + G e e=1 g=1 k=0 4 β gk (w gk + w gk + w gk ) (19) + F ξ f (b f + b f + b f + b f ) f =1 This objective function captures the total cost of network resources over all three layers generically, where η e, β gk, and ξ f are the weights across the three metrics associated with the three layers. The three layer cost structure is shown in Figure 2. An advantage of our cost structure model is that this allows us to consider a number of different cost combinations that are helpful in understanding inter-layer interactions. 6 Figure 2: Cost Structure of The Three-Layer Network For the IP/MPLS layer, η e is the unit cost of link e; this is defined as the sum of the interface cost for the upper layer η U e ends of the connection between the IP/MPLS layer node and the OTN layer node, i.e., η e = 2η U e, where 2 is to count for both ends. At the OTN layer, β gk is the unit cost of link g, and is equal to the cost of the interface of the U k signal on link g, β U g, plus the cost of multiplexing OTN signals β k g, i.e., β gk = 2β U g + 2β k g. For the DWDM layer, ξ f is the cost of link f, and is equal to the interface cost for line-cards connected to the transport end of a physical node to another physical node ξ I f, the optical transponders cost ξt f, the OXC ports ξ o f, plus a physical link distance cost f, i.e., ξ f = 2(ξ I f + ξ t f + ξo f ) + f. The capacity (Normal and Protection) optimization problem (P) for the IP/MPLS-over-OTN-over-DWDM multilayer is to minimize the cost F given by (19) subject to the set of constraints (1) (18), with variables as defined in Table A Three-Phase Solution Approach Problem (P) has a large number of discrete variables and constraints. The number of variables is P D + 2(E(1 + Q)) + 8G(1 + R) + 12GZ + 4F, where P denotes the average number of paths for each demand d, and the number of constraints is D + 4(E +GR + F +G 2 ) + 22G. Furthermore, the problem is NP-hard, since simpler forms of network design problems, such as the singlepath flow allocation or modular link design, are shown to be NP-hard [9]. It is extremely difficult to solve problem (P) using an ILP solver such as CPLEX even for a small size network. We note, however, that if we decompose the problem into three subproblems, then we can solve the problem for moderate size networks tak-

7 Table 4: Summary of Cost Values for Each Layer. Cost Notation Unit Cost Values IP-cost (η e ) 5, 10, 20, 40 U k -cost (β gk ) 2/4/8/16/32, 2/5/13/20/50, 2/6/18/54/162 W-cost (ξ f ) 140 Figure 3: Phases of the Solution Approach ing a phased approach. Therefore, we solve Model (P) in three phases as follows: Phase 1: Solve the following design problem: Minimize E η e (y e + y ) + G e e=1 g=1 k=0 4 β gk (w gk + w gk ) (20) subject to the set of constraints (1) (7). Then, w gk will be a constant in the phase 2. Phase 2: Solve the following design problem: Minimize G g=1 k=0 4 β gk w gk + F ξ f b f (21) subject to the set of constraints (8) (11), (13), and (17). Phase 3: Solve the following design problem: Minimize F f =1 f =1 ξ f (b f + b f + b f ) (22) subject to the set of constraints (12), (14) (16), and (18). Note that w gk and w gk are constant to this phase made by solving phase 1. Figure 3 shows the phases of the solution. Note that even by breaking the original problem into three subproblems, each one of the them is still NP-hard on its own. We have managed to reduce the magnitude of its complexity but it is still difficult to solve for large networks. 7. Study Environment In the formulation of problem (P), η e is defined as the cost of one unit of module M of the IP/MPLS layer link e. In our study, this is also referred to as the IP-cost. Likewise, β gk is the cost of one capacity unit of module type U k of the OTN layer link g. We refer to this cost as the U k -cost for (k = 0, 1, 2, 3, 4). At the DWDM layer, ξ f is the cost of one capacity unit of module N of the 7 DWDM layer link f. This will be referred to as the W- cost. According to [16], one of the cost ratios of future network elements is 8, 0.5, and 1, representing costs of a DWDM transponder, an IP/optical interface card, and a photonic OXC port, respectively. Based on our cost model in Section 5.2, the IP/MPLS layer cost becomes 2 (0.5) = 1, and the DWDM layer cost considering only the transponders and OXC port is 2 (8 + 1) = 18. Then, we add other costs to the DWDM layer to include the interface cost for line-cards connected to the transport end of a physical node to another physical node plus a physical distance cost; we assume this is a fixed cost of 10. This means when the IP/MPLS layer cost is 1, the DWDM cost 28. We transform this value so that when the IP-cost is 5, the W-cost is 140. We fixed the W-cost at 140 throughout our study and adjusted the other units costs to understand the impact due to the cost ratio change at different layers. Specifically, for the IP-cost, we vary the cost starting from IPcost= 5 and double the cost to IP-cost= 10, 20, and 40 to study the impact of different IP-cost scenarios while the W-cost is fixed. For the OTN layer parameter values, we have three possible cost scenarios of U k (0 k 3): UK-cr1: 2 U k = U k+1 UK-cr2: 3 U k > U k+1 UK-cr3: 3 U k = U k+1 To represent them, we consider the following U k cost (k = 0, 1,..., 4), 2/4/8/16/32, 2/5/13/20/50, and 2/6/18/54/162, for UK-cr1, UK-cr2, UK-cr3, respectively. Note that the actual values of U k s are not as important as the relationships between them. Note that we avoid unrealistic U k cost relationships such as when U k = U k+1 or when 4U k = U k+1. The former indicates an equal cost of two different OTN units, and the latter follows one of the signal multiplexing rules we explained in Section 3. We summarize each layer s cost values in Table 4. The experiments we conducted for this study with various parameter values allowed us to examine the impact of each layer cost and IP/MPLS modularity on

8 other layers and ultimately the overall network cost. We wish to answer a number of questions. For instance, how do the IP-cost and the size of M influence the required protection capacity at each layer and the overall network cost? How does the cost of each U k scenario affect the final types and numbers of U k s needed to satisfy a given set of demands? In this three-layer work setting, we consider the 14- node NSFNET in addition to a real-world existing 23- node metro topology as DWDM equipments are being employed into this medium-haul metro topology [21]. In our three-layer case, the NSFNET is considered as 14 nodes in each layer that results in 42 total nodes, and the number of physical fiber links F is 21. We assume that the virtual topologies of the IP/MPLS and OTN layers follow the connectivity of the physical layer, hence, E = G = 21 resulting in 63 total links. The total number of demands is 91 bidirectional demands assuming a demand between every LSR pair where the average demand is 5 Gbps. Therefore, we consider three values of M: 2.5, 5, and 10 Gbps to represent three cases: below average, equal average, and above average demand in the network. Similarly, the metro topology is considered as 23 nodes in each layer that results in 69 total nodes, and the number of physical fiber links F is 30. Again, we assume that the virtual topologies of the IP/MPLS and OTN layers follow the connectivity of the physical layer, hence, E = G = 30 resulting in 90 total links. The total number of demands is 253 bidirectional demands assuming a demand between every LSR pair where the average demand is 5 Gbps. Demands between the LSRs in the network are generated according to the demand model in [22]. For each demand, five primary paths and five protection paths are available at each layer. Figure 4 and 5 show the multilayer design of NSFNET and metro network when IPcost=5, M=2.5 Gbps, and the Uk-cost is UK-cr1 using our phased design approach. Here, a black link indicates that all capacity components of the layer are present on the link. If not a black link at the OTN layer, we use a blue dashed link to indicate the normal capacity, a green dashed link for the protection capacity, and a red dashed link for the capacity that realizes the IP/MPLS protection capacity. At the DWDM layer, we use the same colors to relate this layer s capacity with the OTN capacity components except that the orange dashed link is used for the DWDM layer protection capacity. All results are close-to-optimal derived by solving the three phases of problem (P) as described in Section 6 using the CPLEX 12.2 optimization package where we limit the number of nodes to be visited in the branch and cut tree to 500,000 by declaring set mip limits 8 Figure 4: 14-node per Layer NSFNET Design Figure 5: 23-node per Layer Metro Topology Design nodes The average optimality gap of the solutions is 1.83%. Table 5: Notation and Abbreviation Mapping. Notation Abbreviation Discreption y e N-IP Normal IP capacity y e P-IP Protection IP capacity w gk N-OTN Normal OTN capacity w gk P-OTN Protection OTN capacity w gk P-IP-OTN OTN capacity of P-IP b f N-W Normal fiber capacity b f P-W Protection fiber capacity b f P-OTN-W Fiber capacity of P-OTN b f P-IP-OTN-W Fiber capacity of P-IP-OTN 8. Illustrative Numerical Results 8.1. Total Network Cost To make it easier to follow the results, we present Table 5 that maps each capacity notation in the formulation to an abbreviation. Figure 6 and Figure 7 show the

9 Figure 6: Total NSFNET Cost Figure 8: Capacity of Different Components when IP=5 and UK-cr1 in NSFNET Topology Figure 7: Total metro Network Cost total network cost for each considered scenario in considered topologies. We observe that when the IP-cost is fixed, M=2.5 is more expensive than M=5 which is also more expensive than M=10. This is due to the following reason. Under fixed IP-cost, the larger the M the more demands it can satisfy without increasing the cost. We also observe the cost increase as we increase the IPcost from 5 to 40. By observing the effects of M and its cost, we note that both are essential to be considered jointly in multilayer networks. For instance, the cost of the case when IP=20 and M=2.5 is higher than when IP=40 and M=10. This is simply comparing the cost per Gbps (8 vs. 4). However, not only is the cost per Gbps important to consider, but also the size of M as this parameter will impact the lower layers. Obviously, when M and the IP-cost are fixed, the case of UK-cr3 is the most expensive followed by UK-cr2 and UK-cr1. This is because the gap between the U k -cost is the largest in UK-cr3. Moreover, the cost gap between UK-cr3 and UK-cr2 is larger than between UK-cr2 and UK-cr1. We will explain the reason in Section Capacity of Different Components Figure 8 and Figure 9 show the capacity required for each component when the IP-cost=5 and UK-cr1 in NSFNET and metro topology. We note that the protection capacity of each layer is larger than its normal capacity. Furthermore, the gap between the normal capacity and its protection capacity is increased as we go down in the network s layers. That is, the gap between N-IP and P-IP is less than the gap between the N-OTN and P-OTN, which is also less than the gap between N- W and P-W. This is primarily due to the larger granularity of the lower layers and the longer the protection paths. The gap is the largest in the DWDM layer where each wavelength bit rate N=100 Gbps (compared to M=2.5, 5, or 10 Gbps for the IP/MPLS layer) and the protection paths are usually longer than the primary paths. We note the same trends for different IP-cost and UK-cost scenarios in both topologies. We note that on average, 24% over the N-IP is required for N-OTN, and 33% over the N-OTN is required for N-W. In other words, the required N-OTN is N-IP + 24% of N-IP, and the required N-W is N-OTN + 33% of N-OTN Protection Capacity In this section, we present the protection capacity required for each individual layer. Figure 10 and Figure 11 depict the required P-IP for each scenario in NSFNET and metro topologies. We observe that the cases when M=10 requires more protection capacity than when M=2.5 and 5. M=2.5 is the best case to minimize P-IP. An interesting observation is that the IP-cost is not a significant factor when the M=2.5. Increasing the IP-cost when M=2.5 does not significantly change the required capacity. This is unlike the case when the impact of the IP-cost is noticeable in M=10 scenarios.

10 Figure 9: Capacity of Different Components when IP=5 and UK-cr1 in metro Topology Figure 10: Protection Capacity of the IP/MPLS Layer in NSFNET Topology Figure 12 and Figure 13 depict the required P-OTN for each scenario in both topologies. In this capacity component, the smallest required protection capacity is usually achieved under UK-cr3 and the largest is under UK-cr2. Moreover, the case when M=2.5 is often the case that minimizes P-OTN. This suggests that when M is below the average demand, it is often the best case to minimize the required protection capacity at the OTN layer. Figure 14 and Figure 15 show the required P-W under different scenarios. We note that UK-cr3 is the best case to minimize this capacity component along with M=2.5. The largest capacity is required under UK-cr1 and M=10. We also observe that the protection capacity on average is 22% more than the normal capacity at the IP/MPLS layer, 87% of the normal capacity at the OTN layer, and triple the normal capacity of the DWDM layer. We note that the protection capacity does not exceed the normal capacity at the OTN layer since the method of protection is a link restoration on a single path allows resource sharing. On the other hand, paths of the DWDM layer are link and node disjoint leading to much longer protection paths which explains the significant overprovisioned capacity in this layer. Figure 11: Protection Capacity of the IP/MPLS Layer in metro Topology Figure 12: Protection Capacity of the OTN Layer in NSFNET Topology 8.4. No. of OTN Signals for P-OTN Figure 16 and Figure 17 present the OTN layer U k signals required in the P-OTN component in both topologies. We observe that in UK-cr3 scenarios, U 4 is not used. This is because the cost of 2 U 3 is less than the cost of a U 4. And since 2 U 3 take only 80 Gbps of the wavelength, the rest is mostly filled with U 0. We observe that U 4 is used in UK-cr2 but with fewer numbers than U 3. It is only in UK-cr1 that U 4 becomes higher than U 3 due to the small gap between their cost. We 10 also observe the low numbers of U 1 and U 2 in all cases, suggesting their limited benefits in these scenarios. That is, the IP/MPLS demands will largely be served by U 0 and U 3 under UK-cr3 or a mix of U 4, U 3, with very low numbers of U 0, U 1, and U 2, under UK-cr1 and UK-cr2. The same general trends can be observed when varying the IP-cost or for different OTN capacity components. However, the numbers of U k in P-OTN are higher than those in N-OTN which explains the larger capacity required for P-OTN as shown in Figure 8 and Figure 9.

11 Figure 13: Protection Capacity of the OTN Layer in metro Topology Figure 16: No. of U k of P-OTN OTN Layer Components When IP=5 in NSFNET Topology Figure 14: Protection Capacity of the DWDM Layer in NSFNET Topology Figure 17: No. of U k of P-OTN OTN Layer Components When IP=5 in metro Topology Figure 15: Protection Capacity of the DWDM Layer in metro Topology We stated in Section 8.1 that the cost gap between UKcr3 and UK-cr2 is larger than between UK-cr2 and UKcr1. This is because of the costs of the U k used in each U k -cost scenario. For instance, if we want to consume the full capacity of an OTN link (100 Gbps) under UKcr1, then the optimal solution would be to use 1 U 4, which costs 32. Under UK-cr2, the optimal solution would also be 1 U 4 which costs 50. In UK-cr3, U4 is not the optimal solution since 2U 3 +16U 0 = 140 < 162, the cost of a U 4. We note that the gap between UK-cr1 11 and UK-cr2 in this example is (50 32 = 18, or about a 56% increase). On the other hand, the gap between UK-cr2 and UK-cr3 is ( = 90, or about a 180% increase) Cost vs. Protection Capacity We mentioned in Section 8.1 that the case when M=10 achieves the lowest overall network cost performance when the IP-cost is fixed. This may seem contradictory to the observations pointed out in Section 8.3 that more capacity is often required for protection when M=10 leads to more cost for protection. However, it is important to note that the overall network cost is determined by all capacity components not just the protection

12 components. More important, the cost per Gbps plays a significant role in the IP/MPLS layer. To illustrate, when the IP-cost=40, the costs per Gbps are 16 and 4 for the cases M= 2.5 and 10, respectively. To carry a 10 Gbps demand on a link, 4 Ms are required when M=2.5 is clearly more expensive than the 1 M required when M=10. From a cost perspective, the higher the size of M, the better. However, if we look from a capacity standpoint, a higher size of M implies more capacity needed at each layer Model (P) Special Cases Considering Figure 1, we observe a considerable amount of redundant capacity over the basic normal capacity components: N-IP, N-OTN, and N-W. This is because in our design model each layer is a 100% protected against a single link failure. Furthermore, the multilayer network is guaranteed to survive three link failures when each layer encounters a link failure at the same time. It may also survive more than a single failure per layer providing that a failed link of a primary path does not appear in the protection path of another failed primary path. However, in cases that such overallocated capacity for maximum protection is not required at all layers, we could develop special cases models that are derived from Model (P). For instance, if protection is not critical at the DWDM layer, we can eliminate P-W from the model. That is, modify Model (P) and derive Model-(W) that do not provide any protection to the normal DWDM layer capacity. Moreover, we can develop Model-(O) in which there is no OTN protection component for the OTN normal flow. A third model variation is Model-(I) where we exclude the IP protection component (P-IP) and its lower layer capacity components. The new models are as follows: Model-(W): (no DWDM layer protection) minimize (23) subject to constraints (1)-(15), (17) and (18). Model-(O): (no OTN layer protection) minimize (24) subject to constraints (1)-(7), (12), (14)-(16) and (18). Model-(I): (no IP layer protection) minimize (25) subject to constraints (1)-(2), (4), (6), (8)-(13), and (15)- (17). F = E η e (y e + y ) + G e e=1 g=1 k=0 4 β gk (w gk + w gk + w gk ) (23) + F ξ f (b f + b f + b f ) f =1 12 F = E η e (y e + y ) + G e e=1 F = E η e (y e ) + G e=1 g=1 k=0 4 β gk (w gk + w gk ) (24) + F ξ f (b f + b f + b f ) f =1 g=1 k=0 4 β gk (w gk + w gk ) (25) + F ξ f (b f + b f + b f ) f =1 Similar to our approach with Model (P) in section 6, we have solved the new models: Model (W), Model (O), and Model (I) using the same heuristic approach. Figure 18, Figure 19, and Figure 20 show the total cost achieved by these models under different UK-cr in metro topology. We observe the following points: Model (P) is our original comprehensive model that provides protection at every layer encompassing all model variations. Its cost serves as an upper bound since it is the most expensive model. We observe that the most expensive protection capacity component is usually the DWDM layer protection. This is why Model-(W) is the cheapest since it has no DWDM protection. There is an exception to the previous observation and that is when the cost per Gb is high at the IP/MPLS layer. That is, when IP=40 and M=2.5 the cost per Gb is 16, compared with cost per Gb equal 0.5 when IP=5 and M=10. In these cases, the IP layer protection become the most expensive capacity component. This explains why in such cases Model-(I) becomes the cheapest, even cheaper than Model-(W) when IP=40 and M=2.5 in the three figures. Increasing the UK-cost makes the previous incident, IP layer protection being most expensive, comes also with lower cost per Gb, e.g. when cost per Gb equals 8 (IP=20, M=2.5) in UK-cr3. IP-cost and M has no significant impact on cost achieved by Model-(I). This is dissimilar to the cost performance of the other models. This is because varying the IP/MPLS unit cost has no tangible impact on Model-(I) since this model does not include IP-P component.

13 Figure 18: Total Cost of Different Models under UK-cr1 in metro Topology Figure 20: Total Cost of Different Models under UK-cr3 in metro Topology Figure 19: Total Cost of Different Models under UK-cr2 in metro Topology Model-(O) is the most expensive model following the comprehensive Model-(P). This is because Model-(O) includes the two most expensive protection components: the DWDM and IP layer protection. The cost performance of all models is increasing as the the UK-cost values increases from UK-cr1 to UK-cr2 to UK-cr3. 9. Summary and Future Work In this paper, we presented a network protection design model for a IP/MPLS over OTN over DWDM three-layer network. In this architecture, we explicitly considered the OTN layer as a distinct layer with its own technological constraints. The survivability design provides protection only for the normal capacity of each 13 layer to reduce the protection resources while maximizing the protection. The design model ensures the survivability of the multilayer network for a minimum of three concurrent link failures in which each layer suffers a single failure. We present a heuristic solution to the problem by solving it in three phases to reduce the complexity of the problem in order to solve it for moderate size networks. We next presented a study based on various network parameters to understand their effects especially on the protection capacity and overall costs in two different topologies. We find that even though the case when M=10 (above the average demand) would be the best case to reduce the overall network cost, it is often the one that requires more protection capacity. The case when M=2.5 (below the average demand) would achieve the lowest amount of capacity needed for protection at each layer. This indicates that the optimal ratio of M to IP demand size is 1:2 would achieve the lowest amount of protection capacity for each layer. On the other hand, the ratio of 2:1 is the best to reduce the overall network cost. We observed that the protection capacity of each layer is larger than its normal capacity, noticeably at the lowest layer (DWDM layer), due to the longer protection paths and the larger granularity of this layer. We also noted the limited usage of the OTN layer U 1 and U 2 signals. Mainly, the IP/MPLS demands will be accommodated by U 0 s and U 3 s under UK-cr3 or a mix of U 4, U 3, with very low numbers of U 0 s, U 1 s, or U 2 s, under UK-cr1 and UK-cr2. We also present and solve three variations of our original model where we eliminate a protection component from each layer and compare their cost performance. We observe that the DWDM layer protection is often

14 the most expensive component. However, the IP/MPLS layer becomes more expensive when the IP/MPLS unit cost is high. For future work, we plan to expand our study by developing a heuristic algorithm to solve the problem for large size networks and provide an extensive analysis. We also plan to consider other protection mechanisms and compare their performance under the three-layer architecture. It may be noted that the design problem here is based on using a static traffic matrix. For dynamically changing load, a multi-hour approach can be taken, which uses traffic matrices at different time of the day. In our case, this will increase the number of variables significantly depending on the number of traffic snapshots used. Another possible direction is considering the demand uncertainty in the design problem. Acknowledgement This work was supported in part by National Science Foundation Grant # CNS References [1] I. Katib, D. Medhi, A study on layer correlation effects through a multilayer network optimization problem, in: Proc. of 23rd International Teletraffic Congress (ITC 2011), 2011, pp [2] I. Katib, D. Medhi, Optimizing node capacity in multilayer networks, IEEE Communications Letters 15 (5) (2011) [3] I. Katib, D. Medhi, A network protection design model and a study of three-layer networks with IP/MPLS, OTN, and DWDM, in: Proc. of 8th International Workshop on the Design of Reliable Communication Networks (DRCN 2011), 2011, pp [4] I. Katib, D. Medhi, IP/MPLS-over-OTN-over-DWDM multilayer networks: An integrated three-layer capacity optimization model, a heuristic, and a study, IEEE Transactions on Network and Service Management(Accepted). [5] M. Bhatta, Four challenges in backbone network, Huawei Communicate issue 44 (2008) [6] P. Demeester, M. Gryseels, A. Autenrieth, C. Brianza, L. Castagna, G. Signorelli, R. Clemenfe, M. Ravera, A. Jajszczyk, D. Janukowicz, K. Van Doorselaere, Y. Harada, Resilience in multilayer networks, IEEE Communications Magazine 37 (8) (1999) [7] S. De Maesschalck, D. Colle, A. Groebbens, C. Develder, A. Lievens, P. Lagasse, M. Pickavet, P. Demeester, F. Saluta, M. Quagliatti, Intelligent optical networking for multilayer survivability, IEEE Communications Magazine 40 (1) (2002) [8] D. Medhi, A unified approach to network survivability for teletraffic networks: Models, algorithms and analysis, IEEE Trans. on Communications 42 (1994) [9] M. Pióro, D. Medhi, Routing, Flow, and Capacity Design in Communication and Computer Networks, Morgan Kaufmann Publishers, [10] A. Fumagalli, L. Valcarenghi, IP restoration vs. WDM protection: is there an optimal choice?, IEEE Networks 14 (6) (2000) [11] L. Sahasrabuddhe, S. Ramamurthy, B. Mukherjee, Fault management in IP-over-WDM networks: WDM protection versus IP restoration, IEEE Journal on Selected Areas in Communications 20 (1) (2002) [12] E. Kubilinskas, M. Pióro, Two design problems for the IP/MLPS over WDM networks, in: 5th International Workshop on Design and Reliable Communication Networks (DRCN 2005), 2005, pp [13] H. Zhang, A. Durresi, Differentiated multi-layer survivability in IP/WDM networks, in: Proc. of IEEE/IFIP Network Operations and Management Symposium (NOMS 2002), 2002, pp [14] L. Lei, A. Liu, Y. Ji, A joint resilience scheme with interlayer backup resource sharing in IP over WDM networks, IEEE Communications Magazine 42 (1) (2004) [15] Y. Qin, L. Mason, K. Jia, Study on a joint multiple layer restoration scheme for IP over WDM networks, IEEE Networks 17 (2) (2003) [16] W. Bigos, B. Cousin, S. Gosselin, M. Le Foll, H. Nakajima, Survivable MPLS over optical transport networks: Cost and resource usage analysis, IEEE Journal on Selected Areas in Communications 25(5) (2007) [17] P. Pongpaibool, H. Kim, Impacts of GMPLS on topology design and protection planning of survivable IP-over-optical networks, in: Proc. of IEEE GlobeCom Workshops 2004, 2004, pp [18] J. Zhao, S. Wang, Z. Tang, Integrated multi-layer network survivability based on GMPLS to improve fault recovery time, in: Proc. of International Conference on Computer Engineering and Technology (ICCET 2009), Vol. 2, 2009, pp [19] I. Katib, IP/MPLS over OTN over DWDM multilayer networks: Optimization models, algorithms, and analyses, Ph.D. dissertation, University of Missouri Kansas City, May [20] ITU-T Recommendation G.709/Y.1331, Interfaces for the optical transport network (OTN), Geneva, December 2009, itu-t. [21] J. Serrat, A. Galis, Deploying and Managing IP over WDM Networks, ARTECH HOUSE, INC., [22] B. Fortz, M. Thorup, Internet traffic engineering by optimizing OSPF weights, Proc. IEEE INFOCOM (2000)