Impact of IP Layer Routing Policy on Multi-Layer Design [Invited]

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1 Impact of IP Layer Routing Policy on Multi-Layer Design [Invited] E. Palkopoulou, O. Gerstel, I. Stiakogiannakis, T. Telkamp, V. López, and I. Tomkos Abstract In this work we, first, investigate multi-layer network planning approaches with different levels of integration by extending a commercial planning tool that accurately models the IP network. The different levels of integration are defined with respect to the degree to which information from the optical layer is considered in the optimization of the IP layer. The impact of traffic demand and topology size is evaluated in terms of the obtained cost savings. Case study results indicate cost savings reaching 25% by holistic multi-layer network planning. We then proceed to evaluate the impact of the IP layer routing policy, expressed by the IP link weights, on the cost and latency of a holistic multi-layer network design. As many network operators prefer to assign IP link weights based on simple rules that are unaffected by the network conditions, we focus on two weight setting policies that are agnostic to the network conditions: the Hop-Based Policy and the Distance-Based Policy. We find that different IP layer policies may lead to different optimized IP topologies for the same traffic and optical layer topology. It is shown that the optical network's regeneration requirements affect the optimal policy. Index Terms Multi-layer Planning, IGP Routing Policy, ECMP, Latency M I. INTRODUCTION ulti-layer network planning has been extensively researched as a way to drive down network costs. Many different flavors of network planning fall under the multi-layer category provided that information from different network layers is exploited in the planning process. In addition, new multi-layer planning methodologies are needed to reap the benefits of new multilayer capabilities, For example, cost savings may be achieved by adopting enhanced multi-layer resilience Manuscript received July 1, 2014. E. Palkopoulou and T.Telkamp are with Cisco Systems (epalkopo@cisco.com) O. Gerstel is with Sedona Systems, Israel. I. Stiakogiannakis is with Inria, France V. López is with Telefonica I+D, Global CTO Unit, Madrid Spain I. Tomkos is with the Athens Information Technology Center, Peania, Greece mechanisms such as optical restoration or by optimizing the explicit routing of demands in the IP/MPLS and the optical layers [1] [-6]. However, most multi-layer studies fail to consider requirements posed by real networks. For example, the IP layer is often assumed to route traffic along a single shortest path, while the more common equal-cost multi-path routing (ECMP) is not included. Furthermore, while the impact of failures on the optical layer is considered for dimensioning the IP layer links it is often ignored during the design of the IP topology itself. For example, the IP links are dimensioned to have enough capacity installed in order for there to be no traffic loss under single optical link failures. However, the initial selection of which adjacencies should be set-up on the IP layer (affecting routing and grooming) is performed considering only the fail-free case. This leads to an inaccurate assessment of the impact of a certain IP layer design over an optical topology affecting, for example, the validity of studies of router bypass. This paper aims at bringing out important characteristics of the IP layer that impact such multi-layer planning. The increasing pressure to improve the cost efficiency of networks has led to the development of advanced traffic engineering (TE) and network planning techniques. MPLS TE enables the explicit routing and arbitrary splitting of traffic, which are both desirable characteristics for optimization purposes. However, scalability and robustness may become issues in MPLS TE [7]. TE can also be achieved with pure IP networks. In such IP-based TE networks, the link weights of the interior gateway protocol (IGP) are set in order to meet a given objective [8]. IP TE provides better scalability but less control over the selection of the individual paths. This in turn requires for the IGP link weights (which are added up to define the shortest path in the network) to be carefully tweaked in order to match the new conditions. However, altering the IGP weights of IP links is highly undesirable by network operators. As a result, many network operators prefer to assign link weights based on simple rules that are unaffected by the network conditions. Thus, in this work we focus on two IGP weight setting policies that are agnostic to the network conditions: the Hop-Based Policy and the Distance-Based Policy. We extend a widely used commercial network planning tool [9] to implement different multi-layer planning approaches in IP-over-DWDM networks. In order to provide

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2 Input: IP routing policy Optical topology IP topology Traffic demand Failure states Network equipment characteristics Network Planning and Optimization Tool Approach I: Divide-and-Conquer Approach II: Designing for Failures Approach III: Holistic Multi-Layer Design Output: Optimized network design Required network equipment Total network cost (a) (b) Fig. 1. (a) Network planning and optimization tool (b) Network planning approaches with different levels of multi-layer integration. a fair comparison, all planning approaches assume the same protection/restoration techniques and provide the same level of resilience. The relative cost benefits of introducing different levels of integration in multi-layer are quantified under different traffic demand volume and network topology size assumptions. We demonstrate that jointly considering optical link failures in the design of the IP topology leads to significant cost savings. Finally, we show that counter to intuition including knowledge of the optical layer costs as part of the IP topology design phase does not necessarily yield additional cost reductions. Furthermore, we examine the effect of different IGP weight setting policies when performing integrated multilayer planning. We find that different IGP policies may lead to different optimized IP topologies for the same traffic and optical layer topology. We show that the Hop-Based Policy is more cost-efficient than the Distance-Based Policy, for networks that do not require significant regeneration. However, these cost savings come at the expense of increased end-to-end latency. For networks with high regeneration requirements the Distance-Based Policy is more cost efficient and yields lower latency. Finally, we evaluate the impact of ECMP on the two examined IGP policies. This work extends work presented in [10]. This paper is organized as follows. In Section II we present network planning approaches with a different level of multi-layer integration. Furthermore, different IGP routing policies are discussed: the Hop-Based Policy and the Distance-Based Policy. In Section III case study results are presented quantifying the impact of: (i) integrated multilayer network planning and (ii) IP layer routing policies. II. MULTI-LAYER NETWORK PLANNING AND OPTIMIZATION In this section we first discuss network planning approaches with a different level of multi-layer integration. Then different IP layer routing policies are presented. A. Network Planning Approaches Fig. 1(a) presents a network planning and optimization tool, which can assume network planning approaches with different levels of multi-layer integration. The traffic demand, the optical topology, the network equipment characteristics, and the failure states under which survivability should be provided are given as input to the tool. The output of the tool includes the IP topology, the required network equipment in different network layers (e.g. IP ports, transponders, regenerators), as well as the total network cost. The optimization objective differs depending on the assumed network planning approach. Note that all planning approaches result in networks dimensioned to provide the same level of survivability (e.g. 100% survivability under single optical link failures). We adopt today s resiliency approach, in which optical link failures are always handled by the IP layer. The optical circuits affected by optical link failures are not rerouted to

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3 Optimized Topology for Hop-Based IP Routing Policy Optical Topology Network Planning and Optimization Tool Approach III: Holistic Multi-Layer Design Optimized Topology for Distance-Based IP Routing Policy Fig. 2. Different IP layer routing policies may lead to different optimized IP layer topologies. another path in the optical layer (i.e., optical restoration is not assumed). This means that the IP links that were carried over affected optical circuits will be operationally down until the fiber failure is repaired. For this reason the IP topology has to be provisioned with enough capacity for there to be no traffic loss under optical link failures. The constraints imposed by the network planning tool are discussed in the following. It is required that the IP topology remains connected under the considered failure scenarios (i.e., single optical link failures). Furthermore, it is required that the IP links provide enough capacity to support the traffic under failure scenarios. This requires examining different operational states for the network, which increases the complexity of the problem. The worst-case traffic on each IP link under optical link failures is calculated (i.e., the maximum traffic over all failure scenarios). This information is used to provision enough capacity on the IP links. In the following we present three different network planning approaches, which are shown in Fig.1 (b). Approach I represents a typical planning method, which is used by network operators today. It consists of three steps. In the first step, the IP topology is designed and dimensioned for normal operation (i.e., assuming that there are no failures in the network). The objective of the design process is the minimization of the IP layer costs. Note that during the IP topology design process the optical layer is not considered at all. As a second step, the IP links from the IP topology provided by the first step are routed on the shortest path in the optical layer. Thus, a mapping between IP links and optical links is obtained. This enables us to capture the impact of optical link failures on IP links by performing failure analysis. During failure analysis, the considered failure states are sequentially examined and the worst case traffic on each IP link is recorded. As a result, the IP links are dimensioned according to the expected worst case traffic. Finally, in the third step the optical circuits are dimensioned. This entails the determination of the required number of transponders and regenerators per optical circuit - according to the information provided by the second step. Approach II comprises two steps. In the first step, the IP topology is designed and dimensioned considering all failure states (i.e., failure analysis is conducted as part of the IP topology design phase). The optimization objective is set to the minimization of the IP layer costs. Note that in contrast to Approach III - the optical layer costs are not considered in the IP topology design. In the second step the optical circuits are dimensioned enabling the calculation of the optical layer costs. Approach III adopts a more holistic multi-layer methodology by jointly considering failure analysis and the optical layer costs in the network design. As shown in Fig.1(b), Approach III consists of a single step, in which the optimization objective is set to the minimization of the IP layer and the optical costs. B. Exploring IGP Metrics When performing integrated multi-layer planning, the deployed IP routing policy may impact the obtained results. We focus on two IGP weight setting policies that are agnostic to the network conditions: the Hop-Based Policy and the Distance-Based Policy. This is because many network operators prefer to assign link weights based on simple rules that are unaffected by the network conditions. Hop-Based Policy: In this policy the IGP metric of all IP links is set to a constant value. For example, all IGP metrics are set to 10. Distance-Based Policy: In this policy the IGP metric of an IP link is set to be proportional to the physical distance on the optical circuit that it corresponds to. For example, if an IP link is carried over an optical circuit with a length of 432 km, then its IGP metric is set to 432. We apply the holistic multi-layer methodology (Approach III) to a Spanish national backbone network [2]. As shown in Fig.2, we find that different IGP routing policies may

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4 Table I: Total costs of Approach I over total costs of Approach II Traffic Demand Multiplier Topology Stretch Factor x1 x3 x6 x1 125% 122% 110% x2 119% 117% 114% x3 109% 106% 105% x4 118% 117% 115% Table II: Total costs of Approach II over total costs of Approach III Traffic Demand Multiplier Topology Stretch Factor x1 x3 x6 x1 100% 101% 110% x2 100% 100% 101% x3 103% 106% 110% x4 101% 100% 101% (a) Topology stretch factor: x1 (b) Topology stretch factor: x6 Fig. 3. The normalized total network cost of the different planning approaches is presented for the initial traffic demand volume (x1). lead to different optimized IP layer topologies. Note that the locations of the IP core routers are provided as input. However, the established IP adjacencies are subject to optimization. III. CASE STUDIES In this section we examine the impact of integrated multilayer network planning and IP layer routing policies. The Telefonica backbone network and the corresponding traffic matrix based on the aggregation of the traffic of Spanish regional networks are used as a basis [2]. In order to perform a sensitivity analysis of the impact of the optical cost, we vary the geographical area that the network topology spans. Therefore we examine scaled versions of the initial topology. This is achieved by uniformly scaling the length of the optical links by the so-called topology stretch factor. For example, a topology stretch factor equal to three corresponds to a network topology with three times longer optical links. In a similar manner, we examine the impact of traffic magnitude by uniformly scaling the traffic matrix by a given traffic demand multiplier. The cost for IP ports, transponders, and regenerators is considered as these constitute the most important contributors to the overall costs. We assume next generation networking deploying 16QAM modulation, which enables a capacity of 200 Gb/s per wavelength and assume IP ports that feed each such wavelength with 2x100GE links. The maximum transparent reach of the transponders is assumed to be 800 km (we do not consider an accurate impairment model). Regenerators are assumed to incur a cost overhead of 80% compared to the transponders. Transponders are assumed to cost half of the IP port costs this roughly corresponds to recent price points. This respectively corresponds to relative cost values for transponders, regenerators, and IP ports of 1, 1.8, and 2. The failure set under which full survivability is provided includes all single optical link failures as these are the main failures typical networks are planned for. A. Impact of Integrated Multi-Layer Planning We evaluate the performance of the network planning approaches presented in Section II.A. The IP routing policy is set to the Distance-Based policy in all cases. Table I

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5 (a) Spanish national backbone network (b) Scaled Spanish national backbone network by a factor of 6 Fig. 4. Relative performance of Hop-Based over Distance-Based IGP policies in terms of latency and cost (a) (b) Fig. 5. Spanish national backbone network: (a) The IP link utilization of different IGP policies; (b) The total network cost with set to be ECMP inactive (i.e., the maximum number of ECMP paths is set to 1) over the total cost with ECMP set to be active (i.e., no restrictions are imposed on the maximum number of ECMP paths) for different IGP policies. presents the total costs of Approach I over the total costs of Approach II for different values of the traffic demand multiplier and the topology stretch factor. We find that savings reaching 25% can be achieved by Approach II compared to Approach I. Including failure analysis in the IP topology design has the effect of indirectly introducing more traffic in the design process. When IP link failures occur, the remaining operational IP links have to carry additional traffic load that was previously carried by the failed links. This additional traffic load that occurs in different failure states provides an incentive to establish more bypassing IP links. On the other hand, the more physical links an IP link traverses on the optical layer the more physical link failures require the traffic that it is carrying to be re-routed on other links. Hence, the incentive is provided to establish IP links that are routed on shorter optical paths in terms of hops. Overall, Approach II leads to more IP adjacencies to be established, which tend to have longer optical paths than Approach I. As the topology stretch factor increases, the costs for regeneration increase. As a result, the relative benefits of Approach II over I slightly decrease as the topology stretch factor increases. Table II presents the total costs of Approach II over the total costs of Approach III. We find that the relative benefits of Approach III over Approach II are insignificant for the initial network topology. However, as the topology stretch factor increases, these benefits reach up to 10%. This can be explained by the following. First observe that the number of the transponders is directly correlated to the number of the IP ports. As a result, by minimizing the cost for IP ports the minimization of the cost for transponders is achieved as a valuable by-product. Thus, only the cost for regenerators is not accounted for in the IP topology design of Approach II. Furthermore, as previously explained, an incentive is provided for Approach II to establish IP links that are routed on shorter optical paths in terms of hops. Note that in the examined topology there is a direct correlation between the hop count and the physical length. Thus, Approach II indirectly favors paths with lower regeneration requirements. As a result, Approach III does not always reap additional benefits compared to Approach II. In Fig.3 the total network cost of the different planning approaches is presented for the initial traffic demand. The cost is normalized with respect to Approach III. Fig.3(a) and Fig.3(b) correspond to a topology stretch factor of x1 and x6 respectively. It is evident how the relative cost contribution of regenerators is more significant in geographically larger networks (what is important is the relationship between the transparent reach of the transponders and the size of the

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6 network). As the relative cost contribution of the regenerators increases, so do the benefits of adopting Approach III over Approach II. Finally, it is noted that the variations of the relative performance of the different approaches with respect to the traffic demand show no significant trend in both Tables I and II. B. Impact of IP Layer Routing Policy In the following we focus on the impact of IP layer policies on network cost and latency in integrated multilayer design. To this end we adopt network planning Approach III. The optimization objective is set to the minimization of the total network costs - which include IP ports, transponders, and regenerators. Case study results on the performance of different IGP policies are discussed. Fig. 4 presents the relative performance of the Hop-Based Policy over the Distance-Based Policy in terms of latency and cost as a function of the traffic demand volume. Both average and maximum values are presented with respect to latency (averaging and maximization are applied to the different paths that all demands are routed over). Let us first focus on Fig. 4(a), which corresponds to the presented Spanish national backbone network, and examine the case of traffic demand multiplier 1. We find that costs savings of 16% can be achieved by adopting the Hop-Based Policy as opposed to the Distance-Based Policy. However, these savings come at the expense of 11% higher average end-toend latency and 17% higher maximum latency compared to the Distance-Based Policy. For networks that span a relatively small geographical area, we observe that the Hop- Based Policy incurs less costs than the Distance-Based Policy (on average 11% cost savings). However, it leads to higher latency. While the penalty in terms of increased latency may be expected as the Distance-Based Policy forces demands to be routed over the path with the shortest physical distance the gain in terms of cost savings is not as straightforward. In the following we analyze this behavior. First of all, the Hop-Based Policy captures quite well the current network cost model for networks with low regeneration requirements - as the main contributor to overall network costs are IP ports and transponders that are deployed at the end-points of each IP link. So a demand that is routed on the path with the minimum number of network hops actually consumes the minimum portion of such costly resources. However, this is only part of the picture. Fig. 5(a) presents the IP link utilization of the different IGP policies as a function of the traffic demand volume. We find that the maximum IP link utilization in both policies is approximately 90% - which is the upper limit of the IP link utilization given as input to the multi-layer planning and optimization tool. What is interesting to observe is that the average IP link utilization is significantly lower in the Distance-Based Policy. Thus, better load balancing is achieved for the Hop-Based Policy. We find that this is greatly affected by ECMP. The ECMP mechanism allows for a demand to split among multiple sub-routes with equal IGP link weights. Traffic is evenly split to the next hop routers on these paths again influencing the IP-optical design. Fig. 5(b) shows the impact of ECMP on different IGP policies as a function of the traffic demand volume. In order to quantify this impact we examine the ratio of the total network cost with ECMP set to be inactive over the total cost with ECMP set to be active. Note that when ECMP is set to be inactive, the maximum number of allowed ECMP paths is set to 1. When ECMP is set to be active, no restrictions are imposed on the maximum number of ECMP paths. We find that on average 14% cost savings can be achieved when no restrictions are imposed on the maximum number of allowed ECMP paths for the Hop-Based Policy. On the contrary, no such savings are achieved for the Distance-Based Policy (on average a 2% penalty in terms of cost is incurred). This is affected by the nature of the equal cost paths in the Distance-Based Policy. When two paths have the same metric (corresponding to paths with the same physical length) but a different number of hops, there will be no incentive for the path with fewer hops to be used. Thus, the traffic is split over paths with actually unequal network costs. Furthermore, the probability of paths with equal costs to arise in the Hop-Based Policy is greater. Let us now focus on Fig. 4(b), which corresponds to the Spanish national backbone network scaled by a factor of 6. We find that the Distance-Based Policy is more cost efficient (up to 8% savings) for geographically large networks with high regeneration needs without sacrificing on latency. In this case the Distance-Based Policy captures better the network cost model. IV. CONCLUSION We investigate the benefits from multi-layer network planning techniques with different levels of integration by extending a commercial network planning tool. We find that including failure analysis of the optical links in the design of the IP layer topology yields cost savings reaching 25%. The benefits from including the optical layer costs in IP topology design depend on the size of the network (i.e., the geographical area that is spans), but counter to intuition are much less significant. Furthermore, we evaluate the impact of Hop-Based and Distance-Based IGP routing policies in terms of the network cost and latency. Case study results on a Spanish backbone network indicate that the Hop-Based Policy incurs less costs than the Distance-Based Policy (up to 16% cost savings) but leads to penalties in terms of end-to-end latency (up to 12% with respect to average latency). For networks with significant regeneration requirements the Distance-Based policy yields both lower costs and latency. We find that ECMP, which splits demands among multiple sub-routes with equal IGP metrics, allows significant cost savings to be achieved only for the Hop-Based policy. ACKNOWLEDGMENT This work is partially supported by the CISCO University Research Program and the EC FP7 project n 317999 (IDEALIST).

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 7 REFERENCES [1] O. Gerstel, C. Filsfils, T. Telkamp, M. Gunkel, M. Horneffer, V. Lopez and A. Mayoral: Multi-Layer Capacity Planning for IP- Optical Networks, in IEEE Communications Magazine Feature Topic "Advances in Network Planning", January 2014, Vol. 52, Issue. 1, pp. 44-51. [2] P. Iovanna, A. Germoni, and V. López, Integration of Planning and Control Plane in Packet Optical Multilayer Network, In Proc. Optical Fiber Communication Conference and Exposition and the National Fiber Optic Engineers Conference (OFC/NFOEC), NW1F.2, Mar 2013. [3] M. Pickavet, P. Demeester, D. Colle, D. Staessens, B. Puype, L. Depré, and I. Lievens., Recovery in Multilayer Optical Networks, Journal of Lightwave Technology, vol.24, no.1, pp.122,134, Jan. 2006 [4] M Ruiz, O Pedrola, L Velasco, D Careglio, J Fernández- Palacios, and G Junyent, "Survivable IP/MPLS-over-WSON multilayer network optimization," Journal of Optical Communications and Networking 3 (8), 629-640, August 2011. [5] O. González de Dios, F.J. Jiménez Chico, J.P. Fernández Palacios, S. Oliveira Guerra, and M. A. Callejo Rodríguez, "CAPEX Savings by a Scalable IP Offloading Approach," in Proc. Optical Fiber Conference (OFC), OTuR5, Mar 2011. [6] H. Zhu, H. Zang, K. Zhu, and B. Mukherjee, "A novel, generic graph model for traffic grooming in heterogeneous WDM mesh networks," IEEE/ACM Transactions in Networking, 11:285/299, 2003. [7] N. Wang, K. Ho, G. Pavlou, and N. Howarth, An overview of routing optimization for internet traffic engineering, Communications Surveys & Tutorials, IEEE, vol.10, no.1, pp.36,56, First Quarter 2008. [8] B. Fortz and M. Thorup, Robust Optimization of OSPF/IS-IS Weights, In Proc. International Network Optimization Conference (INOC), pages 225-230, October 2003. [9] Planning and Designing Networks with the MATE Portfolio, Cariden white paper, http://www.cariden.com/download/datasets/822/planning-and- Designing-Networks-with-the-MATE-Portfolio.pdf [10] E. Palkopoulou, O. Gerstel, I. Stiakogiannakis, T. Telkamp, V. López, and I. Tomkos, Impact of IP Layer Routing Policy on Multi-Layer Design, in Proc. Optical Fiber Conference (OFC), W1K.4, Mar 2014.