This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts

for publication in the IEEE ICC 2010 proceedings

Carrier Grade Ethernet versus SDH in Optical Networks: Planning Methods and CAPEX Comparisons
Domenico Siracusa, Guido Maier
Department of Electronics and Information, Politecnico di Milano, Via Ponzio 34-35, 20121 Milan, Italy Email: {siracusa,maier}@elet.polimi.it
Abstract The new Carrier Ethernet services offered by providers require not only a huge amount of bandwidth, but also a more exible access to bandwidth that the traditional transport networks, based on circuit-switching, can hardly provide. Carriers have nowadays the option of migrating to new layer-2 frame-switched technologies developed ad hoc to support Carrier Ethernet. An important question is whether this migration is really cost effective. This paper deals with the problem of designing a transport network able to support Carrier Ethernet. We compare two network models, one based on layer-2 switching and Ethernet interfaces, the other based on legacy circuit switching and SDH interfaces. Minimization of investments for network interfaces is carried out for the two models, given the same statictrafc matrix to be supported. Both an ILP formulation and a heuristic method are proposed to solve the design problem. The results presented for comparison are obtained by applying the optimization procedures to a case-study network. Index Terms Optical networks, network planning, Carrier Grade Ethernet

I. I NTRODUCTION In recent years providers are demanding for a new generation of IP networks able to face the challenges posed by new services and increasing bandwidth demand. This main driving force is fostering the current evolution of the transport-network carriers. The traditional transport infrastructure, almost-entirely based on large-capacity circuit-switched trunks, does not seem appropriate any more to deliver packetbased trafc. SONET/SDH interfaces, ubiquitously deployed until few years ago, are today matched by the Ethernetfamily interfaces. Thus, the most interesting option available to carrier operators is the migration from a SONET/SDH-based infrastructure to the apparently less-expensive and packetswitching oriented Ethernet counterpart. The Metro Ethernet Forum (MEF) has dened the services and requirements of the Carrier Ethernet (CE) [1]. In order to provide services with their QoS attributes, an Ethernet Virtual Connection (EVC) [1], a sort of tunnel between two or more endpoints of a network, has to be established. The delivery of the end-to-end Ethernet service must be transparent to the (potentially multiple) transport technologies adopted in the
The work presented in this paper was carried out with the support of the BONE-project (Building the Future Optical Network in Europe), a Network of Excellence funded by the European Commission through the 7th ICTFramework Programme.

networks crossed by the EVC. This leaves open the choice of the technology to the carrier operators. The currently most common transport solutions to support CE service are: Ethernet-over-SONET/SDH (EoS) and IP/MPLS tunneling. Both solutions make use of welldeveloped and established switching equipment respectively operating at layer-1 (circuit-switching digital cross connects) and at layer-3 (label/packet switching routers). The activity of standardization of the MEF has spurred major standardization organizations, such as IEEE, ITU and IETF, to develop standards for brand-new network technologies supporting CE through layer-2 switching. In particular, IEEE recently standardized Provider Backbone Bridging - Trafc Engineering (PBB-TE) [2], while ITU-T developed TransportMPLS (T-MPLS) [3]. After concerning about the full compatibility of T-MPLS with already established IP/MPLS standards, a liaison between ITU-T and IETF is leading to the denition of MPLS-Transport Prole (MPLS-TP). In the next years, we will see which one of the two options, whether to continue with legacy equipment or to migrate to the new layer-2 switching concept, will prevail upon the carriers. Surely, the nal choice will be made on the basis of the cost and the performance of the various solutions. The main objective of this work is to make a comparison between the CE transport solution based on Packet-over-SDH and the upcoming layer-2-switching CE technologies. This preliminary paper is dedicated to off-line network planning and optimization on the basis of a static-trafc matrix. In particular, we will focus on the problem of minimizing the number of deployed node-interfaces (SDH and Ethernet in the two cases, respectively). We provide a mathematical formulation of the network design problem and compare the results given by different optimization algorithms (ILP and heuristics) in a case-study network. Our intent is to propose new CE networkdesign methods and principles. Beside that, we would like to numerically measure (by simulation) how much the bandwidth efciency and multipath capability combine to affect the choice between Ethernet or SDH technologies. From the standpoint of this paper, for the assumptions later reported, the PBB-TE and T-MPLS technologies are treated as similar. Therefore, we will refer to a more general layer-2 packet-switched (or better frame-switched) CE implementation instead of explicitly referring to any of such technologies.

978-1-4244-6404-3/10/$26.00 2010 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

This article is structured as follows: after a short description of prior works (Section II), we discuss in Section III the issue of planning a CE network in the two SDH and layer-2switched implementations, pointing out some assumptions. In Section IV, we introduce the design and optimization problem providing a mathematical ILP formulation. In Section V, we present a simulated-annealing heuristic method. In Section VI, we show the comparison between the different transport technologies. Section VII draws some conclusions. II. P REVIOUS WORKS The main objective of this work is to make a comparison between transport technologies used to support CE. The number of papers currently in literature comparing either the different CE implementations or legacy transport to them is rather limited. Different state of the art papers have been presented, from [4] presented in 2005 which examines CE technologies, to the papers based on the CE solutions, such as [5] and [6] or [7]. In June 2006, a technical and economical analysis of PBB transport technology was presented [8]. A similar analysis was presented by [9]. A paper concerning the design of a CE network appeared in IEEE Journal of Lightwave Technology in January 2008 [10]. The authors studied the problem of designing reliable and cost-efcient high-rate carrier-grade Ethernet in a multiline-rate optical network under signal transmission-range constraints. The same authors presented other two articles: [11] and [12]. The rst one offers an optimal solution of the transmission-range problem by adopting ILP. In the second work, the authors propose an algorithm that provides better performance than the shortest-path one. In May 2008 [13] proposes a control plane for Carrier Grade Ethernet networks, taking into account the T-MPLS transport technology. The paper [14] describes how to use SDH infrastructure to deliver CE services from the providers perspective. In September 2008, IEEE Communication Magazine dedicated a special issue to CE technology: in [15], the work done by [7] is extended; in [16], the attention is focused on the comparison between PBB-TE and SDH technologies from a qualitative point of view. During 2009 several articles regarding Carrier Ethernet technologies have been published. Among them [17] which makes a comparison between T-MPLS/MPLS-TP and PBB-TE from the application point of view and [18] which investigates the topology of IP layers that minimizes the total transport network cost upon a realistic reference network. III. N ETWORK MODELS AND DESIGN - PROBLEM
DEFINITION

Let us better dene the network scenarios we are considering in this comparative work. From now on, we will refer to layer-2 packet-switched transport technologies as Packet and to SDH transport technologies as SDH. There are several reasons for a possible migration of the carrier network infrastructure from the SDH to the Packet technology (Section II). SDH interfaces are regarded as being

more expensive. Moreover, with Packet technology the rate of ows assigned to each EVC varies with continuity, while SDH sets-up circuits with xed bandwidth granularity. We do not consider here other potential advantages of the Packet solutions, such as statistical multiplexing, or other QoS related issues such as delay, delay variation, etc. Such aspects will be analyzed in future extensions of the paper. Moreover, this paper considers only point-to-point EVCs (E-Line service), leaving multicast connections (E-Lan and E-Tree service) to future studies. Since standardization of the control plane is not yet complete for all the CE technologies, in the Packet case we have to make some simplifying assumptions on the behavior of the control of routing operations. In particular, we assume that all the packets (frames) of an EVC are forwarded as a single ow between two endpoints on an unique route through the network, with no bifurcations or merging points. This assumption complies with the examples provided also in standards and in [15]. In the SDH case, thanks to the Virtual Concatenation (VCAT) technology, the trafc of an EVC can be transported on different Virtual Container (VC) paths which can follow different routes through the network. In order to provide a meaningful comparison, we consider only 10 Gbit/s Ethernet (transmission bit rate of 10312.5 Mbit/s and capacity for PDU payload of 10000 Mbit/s) and STM-64 (transmission bit rate of 9953.28 Mbit/s and capacity for PDU payload 9584.64 Mbit/s) interfaces, respectively for Packet and SDH. In the Packet case, any value of bandwidth can be reserved for each EVC on each link of the network. To avoid limitations in the transmission distances (link lengths), we suppose that the Ethernet signals generated by the interfaces are transparently carried by Optical Transport Network (OTN) [19]. These interfaces allow to extend the reach of transmission of Ethernet systems without requiring any intermediate encapsulation protocol or any change to frame formats or line encoding (64B/66B). In particular, the so called 10GBASE-R solution includes the Ethernet frames directly into a slightly over-clocked version of standard Optical Data Unit (ODU) container: the ODU-2e container. In the SDH case, the EVCs have to be accommodated into the SDH paths. We assume that a quantum of bandwidth for SDH corresponds to a VC-4 container (payload 149.760 Mbit/s), which thus represents the minimum granularity for bandwidth reservation in the circuit-switched case. According to a popular and effective EoS implementation, we further assume that Ethernet frames are encapsulated into the VC-4 containers by the Generic Framing Procedure (GFP) (GFP-F mapping). As GFP is a very efcient protocol, we will neglect the limited overhead introduced by frame mapping. Given all the above assumptions, we can now dene the problem we want to deal with. The network topology (nodes and links) is given, as well as a set of EVC requests. The links of the network are optical and contain a variable number of channels, obtained by multiplying the number of bers by the number of wavelengths on each ber. As nodes are

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

assumed to be all opaque (no transparent passthrough allowed, wavelength conversion and regeneration always available at each node), the number of wavelengths and bers exactly used is irrelevant. The only relevant condition is that the number of optical channels active on a link in each direction has to be equal to the number of deployed interfaces. Notice that actual transmission systems are always symmetrical in the two propagation directions across a link. Hence, we will have to impose that the number of interfaces deployed at the two edges of the same link must be the same. Each EVC request species a source and a destination among the nodes of the network and a bandwidth to be reserved for the EVC. EVCs are all unicast and unidirectional and the trafc matrix is assumed static (permanent connections). The design occurs ofine: nodes have to be equipped with a suitable number of interfaces, so that enough capacity exists to accommodate all the EVCs. The function to be minimized is the total number of interfaces to be deployed. The output consists of the number and location of the interfaces as well as the routing of all the EVCs. In both Packet and SDH cases, we assume that trafc injected into each EVC is controlled at the source by performing a continuous shaping or policing. The maximum information rate is strictly enforced, possibly allowing for short temporary violations to accommodate small-size bursts of data. The enforced maximum information rate for each EVC corresponds (after adding frame headers) to the bandwidth reserved in the network for that EVC. This assumption on trafc is particularly important for the following reason. One may wonder if our procedure that minimizes the number of interfaces is not detrimental for QoS, as it tends to overload the remaining interfaces. This side effect is avoided if we assume that statistical multiplexing is not exploited in designing the network in the Packet case, consistently with what previosly stated. Indeed, this condition is ensured if: 1) each ow loading an EVC is forced to guarantee a maximum burst-size; 2) an independent portion of buffer (or an independent buffer) is reserved in each transit node of the network to each EVC, so that the size of such a buffer is signicatively larger than the corresponding maximum burst-size. Under the above hypothesis, a uid-ow model [20], [21] can be adopted to analyze the network. According to this model, the packet loss rate on each EVC (or equivalently the delay) becomes independent of the total load of the links: each EVC can be loaded up to the capacity reserved in the network for it, without causing instability. Obviously the advantages of statistical multiplexing are lost, but this is not a problem. At the same time, the model allows us to better focus our Packet vs. SDH comparison just on bandwidth granularity and interface-related CAPEX. It should be nally noted that the design procedure presented in this paper supports EVCs differentiation into two classes of service: protected and unprotected EVCs. Protected EVCs requires the reservation of the EVC capacity along a

primary path as well as a secondary one, with the two paths link-disjoint (only dedicated path protection is considered). IV. I NTEGER L INEAR P ROGRAMMING FORMULATION The formulation we have adopted allows us to apply it to the two Packet and SDH cases with just a few modications. Here follows the description of data, variables, objective function and constraints:

Data: - G = (V,E): network topology, in which V and E are the sets of nodes and (unidirectional) links, respectively (V=|V|; E=|E|). - bh : connection request number h of b Mbit/s between s,d source node s and destination node d. - Hs,d : number of requests between s and d, H s,d = |Hs,d |. - S: set of source nodes. - D: set of destination nodes. - K: set of deployable interfaces per each link. - B: bandwidth of an interface (bandwidth available for payload). - M : maximum number of interfaces for each link. Variables: s,d,h s,d,h - Xi,j,k : Xi,j,k = 1 (i = j) if connection request number h from s to d is routed across link (i, j) over interface s,d,h k, otherwise Xi,j,k = 0. - Ii,j,k : Ii,j,k = 1 (i = j) if the interface k on the link (i, j) has been deployed, otherwise Ii,j,k = 0 Objective Function: route all connection requests deploying the minimum number of bidirectional interfaces. Minimize :
(i,j)E,kK

Ii,j,k

(1)

Constraints: - Solenoidality constraints, for all s S, d D, and h Hs,d :

s,d,h Xj,i,k = jV,kK jV,lK s,d,h Xi,j,k = 1 jV,kK s,d,h Xi,j,k = 1 iV,kK s,d,h Xi,j,l

i = (s, d)

(2)

iS

(3)

jD

(4)

- Capacity constraints:
s,d,h bh Xi,j,k Ii,j,k B i,j hHs,d

(i, j) V, k K (5)

- Bidirectionality of interfaces constraint: Ii,j,k = Ij,i,k (i, j) E, k K (6)

Regarding the distinction between Packet and SDH network, we introduce the following rules on the data set:

SDH network: - B=64 C where C is VC4 capacity. - Each connection request has the same bandwidth as the payload of a VC4, that is C. - The total number of connection requests is equal to: H s,d = bh /C s,d

Packet network:

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

- B=10 Gbit/s.

Concerning the protected class of service, we use a dedicated path protection algorithm and we introduce a new variable s,d,h s,d,h Yi,j,k that is identical to Xi,j,k but it refers to the protection path. Then the capacity constraint becomes:
s,d,h s,d,h bh (Xi,j,k + Yi,j,k ) Ii,j,k B i,j hHs,d

(i, j) V, k K (7)

The description of Packet and SDH data sets does not differ from the previous one. Some new constraints must be added to the unprotected class of service formulation. These constraints have to be complied for all h Hs,d :
s,d,h Yj,i,k = jV,kK jV,kK s,d,h Yi,j,k

i = (s, d)

(8)

s,d,h Yi,j,k = 1 jV,kK

iS

(9)

s,d,h Yi,j,k = 1 iV,kK

jD

(10)

In the end we consider the link-disjointness constraint:

s,d,h s,d,h (Xi,j,k + Yi,j,k ) 1 kK

(i, j) E, h Hs,d

(11)

The computational complexity of the ILP in terms of the number of variables for the unprotected case is equal to EK[V (V 1)H +1]. In the worst case (fully meshed network) it goes as O[V 4 KH]. For the protected case the number of variables is EK[2V (V 1)H + 1] and, in the worst case, the complexity goes as O[2V 4 KH]. The proof of NP-completeness for the aforementioned problem () can be given by restriction [22], i.e. by proving that contains a known NP-complete problem as a special case. In particular we can restrict the instances of our problem to a network composed by only two nodes and a link. Under these conditions our problem corresponds one-to-one to the bin packing problem, which is NP-hard in the strong sense [22]. The equivalence (not formally proven here for brevity) is due to the following properties: the objective function (1), in which only interfaces are considered as indexes, is the same of the bin packing problem; constraints (2), (3) and (4) can be disregarded; capacity constraint (5) has a clear correspondance to the bin packing capacity constraint; equation (6) only sets a scaling factor which does not affect the decision problem. V. H EURISTIC OPTIMIZATION BY S IMULATED A NNEALING As NP-hard problems typically require long computational time, we need a heuristic approach to solve realistic-size design problems. We chose to develop a meta-heuristic approach based on Simulated Annealing (SA) [23]. We applied it to our case as follows. After a rst routing phase in which all requests are satised, an optimization phase follows, with the purpose of minimizing resource utilization. The number of deployed Ethernet and SDH interfaces in the Packet and SDH case is the cost function we are trying to minimize. The optimization phase uses the Algorithm 1.

Algorithm 1 Simulated Annealing 1) Set the initial temperature (rerouting probability); 2) Save the current state (routing of all connections, deployed interfaces as optimal and current number of deployed interfaces as minimum cost); 3) Randomly choose a link of the network; 4) Try to reroute all the connection requests passing through the link: all requests are successfully rerouted: the released resources (interfaces) can be canceled and a conrmation of the positive result is returned; one or more requests cant be rerouted: the previous state is restored and a conrmation of the negative result is returned; 5) Discriminate on the basis of the outcome of the rerouting routine: positive result: - if the number of deployed interfaces decreases compared to the minimum cost, store the new state as optimal and keep it for the next SA routine; go to 6; - if the number of deployed interfaces increases compared to the minimum cost but it is under a threshold, keep the new state for the next SA routine; go to 6; - if the number of deployed interfaces increases compared to the minimum cost and it is over a threshold, restore the optimal state; go to 6; negative result: generate a random number between 0 and 1, if this number is lower than the rerouting probability, add a channel (two interfaces) to all links, with the exception of the link in which the rerouting routine has been attempted; 6) After a xed number of steps, decrease the temperature (rerouting probability); 7) Repeat the steps from 3 to 6 until the temperature (rerouting probability) is below a xed threshold.

VI. L AYER -2 SWITCHING VERSUS SDH We now discuss the comparison between layer-2 switching and SDH in terms of number of deployed interfaces (CAPEX), which is intended to be the nal goal of the paper. On this account it should be noted that more than providing a response on which is the most cost effective solution, we intend to elaborate a methodology of comparison to quantify the difference between the two technologies in terms of number of deployed interfaces, trying to identify network and trafc conditions that make one solution more suitable than the other. Given the assumptions described so far on the dened network models and scenarios, two main phenomena affect the number of deployed interfaces: (a) granularity; (b) multipath. SDH suffers a granularity impairment compared to layer-2 switching because: i) the bandwidth in SDH can be assigned

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

to connections in chunks of VC-4 capacity; ii) the bit rate of an SDH interface (9953.28 Mbit/s) is slightly lower than the rate of a 10GE interface (10312.5 Mbit/s). It is easy to realize that the global effect of the granularity impairment decreases as the overall volume of trafc and as the average bandwidth required per trafc ow get large. Viceversa, granularity effect is strengthened when the average path-length (in number of hops) increases. On the other hand, the layer-2 switching solution suffers a multipath impairment compared to SDH. This is due to the capability of SDH, given by VCAT, of slicing a single ow on multiple paths across the network, creating a load-sharing effect that is expected to improve performance. However, it should be considered that the multipath effect can be actually appreciated in case two or more paths with disjoint interfaces exist and the amount of available bandwidth on the minimum cost path is less than the requested one; this means that connection requests are forced to pass through different disjoint paths. Thus, low nodal degree and small number of available resources are expected to reduce the impact of the multipath impairment, especially when the requested bandwidth per connection is low. The neat result of the combination of granularity and multipath is not obvious. Let us consider the simple case in which the requested bandwidth is equal for each connection and let us focus a single network node n. The number of outgoing interfaces needed by n in the Packet and SDH case be OIP and OIS , respectively. We have OIP = r/ BP /b OIS = ( b/CV C4 r CV C4 )/BS (12) (13)

Normalized number of output interfaces per node

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5

PACKET SDH

Normalized Requested Bandwidth (

Fig. 1. Number of output interfaces per node needed to setup connections with equal requested bandwidth in a point-to-point link.

where r is the number of connection requests of bandwidth b arriving at the node n; CV C4 , BP , BS , are the bandwidths of a VC4 container, a Packet and an SDH interface, respectively. It should be noted that r accounts for connection requests for which the considered node is either the source node or a transit node. In gure 1 we show the two curves representing OIP and OIS of n in the simple theoretical case of a network of two nodes n and m and one link ln,m . In such case n is source of all the r connection requests (n can not be transit node) and thus r can be assigned arbitrarily (r = 15 in the gure). The two curves are represented as functions of the normalized bandwidth = b/BP . The vertical axes is also normalized to the maximum number of output interfaces, which is given by SDH when is equal to 1 (since BP > BS ). It can be observed that for low values of the two technologies have similar trends; sometimes SDH needs more interfaces, because of the granularity impairment. Around intermediate values of this trend changes, and for 0.5 the SDH technology is noticeably cheaper than Packet, because of the the multipath impairment prevailing. For rational values of , Packet always performs better than SDH because it lls exactly the available payload on the interface and the multipath effect becomes useless, leaving the granularity impairment to

prevail. We have veried that the same behavior holds for non degenerative networks, this suggests that estimating a CAPEX comparison between SDH and Packet by the simple eqs. 12 and 13 is possible. Moreover, {r, b, CV C4 , BP , BS } are the main parameters an operator should consider to decide whether to invest in SDH or Packet interface, independently of the network topology and complexity. We nally show the comparison between the SA heuristic and the exact ILP approach in the optimization of the number of interfaces for static trafc and a realistic network. The comparison has been carried out for the two SDH and Packet technologies and for both classes of service (protected and unprotected). The chosen network is the Pan-European Research Network Geant2* [24] (shown in gure 2). The results
SE IE UK NL PL FR BE CH ES IT DE XX HU SI HR GR
Fig. 2. Geant2* network topology.

CZ AT

are depicted in gure 3. The plotted graphs show that the difference between the results obtained by SA and ILP is always beyond 10%. The static trafc is built starting from a matrix (taken from the project MUPBED [24]) representing

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

1e+3

Number of Deployed Interfaces

ILP - PACKET - NO PROT SA - PACK - NO PROT ILP - SDH - NO PROT SA - SDH - NO PROT ILP - PACK - PROT SA - PACK - PROT ILP - SDH - PROT SA - SDH - PROT

Ethernet interface is reputed to be less expensive than an SDH interface. Hence, the advantages of the Packet in terms of monetary CAPEX value are probably more remarkable than what predicted here. R EFERENCES
[1] R. Santitoro, Metro Ethernet Services - a technical overview, Metro Ethernet Forum, 2003. [2] IEEE P802.1Qay, IEEE Draft Standard for Local and Metropolitan Area NetworksVirtual Bridged Local Area Networks - Amendment: Provider Backbone Bridge Trafc Engineering, 2009. [3] ITU-T Standard G.8110.1/Y.1730.1, Architecture of Transport MPLS (T-MPLS) layer network, 2006. [4] Lucent-Technologies, Carrier Ethernet - dened, 2005. [5] TPACK, PBT: Carrier Grade Ethernet transport, 2006. [6] TPACK, Transport MPLS - a new route to Carrier Ethernet, 2006. [7] D. Allan, N. Bragg, and A. McGuire, Ethernet as Carrier Transport Infrastructure, IEEE Communications Magazine, vol. 44, no. 2, pp. 134 140, 2006. [8] A. Schmid-Egger and A. Kirst dter, Ethernet in core networks: A a technical and economical analysis, in HPSR 2006 Workshop, Poznan, Poland, 2006, pp. 135 140. [9] A. Kirst dter, C. Gruber, J. Riedl, and T. Bauschert, Carrier-Grade a Ethernet for packet core networks, in Proceedings of SPIE - The International Society for Optical Engineering, vol. 6354 I, Gwangju, Korea, Republic of, 2006. [10] M. Batayneh, D. A. Schupke, M. Hoffmann, A. Kirst dter, and a B. Mukherjee, Optical network design for a Multiline-Rate CarrierGrade Ethernet under transmission-range constraints, Journal of Lightwave Technology, vol. 26, no. 1, pp. 121 130, 2008. [11] M. Batayneh, D. A. Schupke, M. Hoffmann, A. Kirst dter, and a B. Mukherje, Carrier-Grade Ethernet over WDM under maximum transmission range (TR) constraints of signal, in OFC/NFOEC 2008, San Diego, CA, United states, 2008. [12] M. Batayneh, D. A. Schupke, M. Hoffmann, A. Kirst dter, and a B. Mukherjee, Cost-efcient routing in Mixed-Line-rate (MLR) optical networks for Carrier-Grade Ethernet, in OFC/NFOEC 2008, San Diego, CA, United states, 2008. [13] R. Fu, Y. Wang, and M. Berger, Carrier Ethernet network control plane based on the Next Generation Network, in Proceedings of the 1st ITUT Kaleidoscope Academic Conference, Innovations in NGN, K-INGN, Geneva, Switzerland, 2008. [14] A. Atie, Leveraging SDH infrastructure to deliver Ethernet Services, in IEEE LCW 2008, Beirut, Lebanon, 2008, pp. 51 57. [15] R. S nchez and L. Raptis, Ethernet as a Carrier Grade technology: Dea velopments and innovations, IEEE Communications Magazine, vol. 46, no. 9, pp. 88 94, 2008. [16] A. Reid, P. Willis, I. Hawkins, and C. Bilton, Carrier Ethernet, IEEE Communications Magazine, vol. 46, no. 9, pp. 96 103, 2008. [17] R. Vaishampayan, A. Gumaste, S. Rana, and N. Ghani, Application driven comparison of T-MPLS/MPLS-TP and PBB-TE - driver choices for Carrier Ethernet, in INFOCOM Workshops 2009, IEEE, April 2009, pp. 16. [18] T. Engel, A. Autenrieth, and J.-C. Bischoff, Packet layer topologies of cost optimized transport networks multi-layer netwok optimization, in ONDM 2009, Feb. 2009, pp. 17. [19] ITU-T Standard G.709/Y.1331, Interfaces for the Optical Transport Network (OTN), 2003. [20] R. Gu rin, H. Ahmadi, and M. Naghshineh, Equivalent capacity and e its application to bandwidth allocation in high-speed networks, IEEE Journal on selected areas in Communications, vol. 9, no. 7, pp. 968 981, 1991. [21] D. McDysan, QOS & trafc management in IP & ATM networks. McGraw Hill, 2000. [22] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness. W.H. Freeman and Company, 1979. [23] S. Kirkpatrick, C. D. Gerlatt, and M. P. Vecchi, Optimization by Simulated Annealing, Science, vol. 220, pp. 671680, 1983. [24] MUPBED - Multi-Partner European Test Beds for Research Networking, Deliverable D1.2 - Revision of the reference architecture according to the results of the project studies, 2005.

1e+2

1e+1 0 5 10 15 20 Traffic Scale Factor 25 30

Fig. 3.

Number of deployed interfaces in Geant2* network.

the estimation of actual trafc volumes between the nodes of the network. The total trafc volume has been varied over a limited range to ensure ILP convergence within a reasonable computational time. ILP has been performed on a standard pc using CPLEX software. As expected the computational time for SA is on average much lower (a few hours) than for ILP (a few days). The interfaces deployed in the Packet network are always less than those in the SDH one, not only because of the granularity constraint which affects SDH circuits, but also because the multipath effect is nearly absent. In fact at the highest trafc scale factor, the average requested bandwidth is 925 Mbit/s, which in most cases is too low to conveniently exploit VCAT multiple paths. Moreover, the Geant2* topology is characterized by a low average nodal degree, thus making few different paths available for VCAT. VII. C ONCLUDING REMARKS This paper deals with the problem of optimizing networks able to support Carrier Ethernet connections. The purpose is to evaluate the advantages of exible access to bandwidth, deriving from the adoption of layer-2 switching technologies (such as PBB-TE and T-MPLS), against the disadvantages of the xed bandwidth granularity of SDH networks. We have presented an integer linear programming formulation of the problem of minimizing the number of interfaces deployed in the network to setup a given static connection-request matrix, and proposed a less computationally-hard heuristic based on simulated annealing. This preliminary study allowed us to identify a set of fundamental parameters to evaluate the pros and cons of migrating from legacy technologies to newfangled ones. It should be nally commented that our study leads to comparisons based solely on the number of interfaces, but a 10 Gigabit