Provisioning Virtualized Cloud Services in IP/MPLS-over-EON Networks

Transcription

1 Provisioning Virtualize Clou Services in IP/MPLS-over-EON Networks Pan Yi an Byrav Ramamurthy Department of Computer Science an Engineering, University of Nebraska-Lincoln Lincoln, Nebraska USA Abstract Clou computing enables the provisioning of resources in a reliable an on-eman manner. With the increasing importance of the network banwith in the Clou environment, the networking relate resources nee to be optimally allocate together with the traitional Clou computing resources. In aition, the significant growth of the global ata center traffic poses the challenge of supporting emans with large banwith requirement for the Clou provier. In our paper, we consier the network-efficient virtualize clou infrastructure provisioning (NE-VCIP) problem in IP-over-EON network base on the ata center as a service moel. The elastic optical network (EON) is aopte to provie spectrum an cost efficient networking resources for large banwith requests in our work. We evelop ILP formulations to construct the mathematical moel for this problem an propose a cost-optimize heuristic to solve this problem. To investigate the cost an blocking rate for the serve emans, ifferent moulation formats are compare in the EON layer. The experimental results show that ifferent moulation formats that are aopte in the EON layer will have ifferent impacts on the total cost an eman blocking rate for the same ata set size. In orer to minimize the total cost an obtain a better system performance as well (e.g. low blocking rate), a trae-off nees to be consiere between the two. I. INTRODUCTION Clou computing offers computing resources to a large amount of on-eman service applications. Customers can reserve the require resources through the infrastructure as a service (IaaS) to complete their computing intensive tasks. In the future, customers may not only want to reserve computing resources, such as virtual machines (VMs) an storage, but woul also want to reserve their own Clou environment. A new architecture propose in [1] as ata center as a service (DCaaS) for the future Clou computing coul satisfy such requirements from customers. DCaaS allows customers to create their own Clou platform without constructing the physical DCs. The virtual ata center (VDC) service which falls within IaaS enables users to quickly access Clou infrastructure from a service provier such as vclou Suite by VMware [2], VMDC by Cisco [3], etc. The reserve virtual Clou environment consists of geographically istribute virtual ata centers (VDCs) an backbone networks that connect these VDCs. The Cisco Global Clou Inex (GCI) [4] is an ongoing effort to forecast the growth of global ata center an cloubase IP traffic. GCI inicates in the forecast an methoology report for that, the global ata center traffic an global Clou traffic will increase significantly in the future years [4]. To support the large amount of traffic in the Clou environment an satisfy the requirement of non-blocking bisection banwith among servers, huge banwith capacity shoul be provie by an efficient interconnection architecture. Therefore, optical networking with scalable banwith capacity, low cost an low latency woul be esirable [5]. In this paper, we investigate the network-efficient virtualize Clou infrastructure provisioning (NE-VCIP) in multilayer network architecture. In our problem, a virtualize clou infrastructure (VCI) eman submitte by a user consists of the VDC infrastructures an the virtualize network (VN) connectivity. Each VDC is provie with require amount of computing resources. The VNs are specifie with certain amount of banwith for ata transmission. The banwith requirement is an essential aition which provies the significant benefit of performance preictability for istribute computing [6]. The centralize controller nees to map the VDCs an VN to the geographically istribute physical DCs an backbone networks that both have enough relate resources. To guarantee the banwith requirements by VN, optical circuits are establishe. In this paper, we consier the backbone network with IP-over-EON (elastic optical network) architecture. So one important task for the controller is to complete the routing an spectrum assignment (RSA) in the multi-layer network when oing VN mapping. The elastic optical network (EON) has become a promising approach for flexible banwith provisioning in optical networks. EON can provie high capacity banwith for the emans that cannot be well supporte in current WDM networks. II. RELATED WORK The optimal resource provisioning in Clou has been a challenge in Clou computing. Various investigations have been conucte for the resource provisioning problems in the Clou. The work in [7] stuies the virtual resource allocation problem for networke clou environments, incorporating heterogeneous substrate resources, an provies an approximation approach to aress the problem. The work in [8] presents a system that makes use of virtualization technology to allocation DC resources ynamically an targets to optimizing the number of servers in use. A set of heuristics are evelope to prevent overloa in the system while saving energy use. Moreover, to get the maximum benefit from a istribute clou system, efficient algorithms are neee for resource allocation which minimize communication costs an latency. The work in [9] evelops efficient resource allocation algorithms to aress such problems in istribute clous. The target of this work is to minimize the maximum istance, or latency, between the selecte DCs. In aition, VDC networks has been consiere as a feasible alternative to satisfy the requirements of avance Clou

2 infrastructure services. Proper mapping of VDC resources to their physical counterparts can impact the revenue of clou proviers [10]. In aition to the VM resources, the work in [10] proposes a new embeing solution for DCs that consiers the relation between switches an links, an allows multiple resources to be mappe to a single physical DC. Also a recent stuy in [11] about VDC embeing proposes VDC Planner, a migration-aware ynamic virtual ata center embeing framework that aims at achieving high revenue while minimizing the total energy cost over-time. This work focuses on embeing VDCs into one physical ata center. The work in [12] about VDC embeing proposes Greenhea, a holistic resource management framework for embeing VDCs across geographically istribute ata centers connecte through a backbone network. The goal of Greenhea is to maximize the clou provier s revenue while ensuring that the infrastructure is as environment-frienly as possible. This work focuses on embeing VDCs into istribute infrastructures which is ifferent from the work in [11] an [10]. Different from other works, we make use of the flexible optical network as the backbone network in the Clou to investigate the virtual clou resource provisioning problems. Our objective is to minimize the total cost (CapEx an OpEx) for resource provisioning in the clou environment. To the best of our knowlege, this is the first work that investigates costoptimize virtual clou resource provisioning while utilizing the IP-over-EON network architecture. III. NE-VCIP PROBLEM As we mentione above, in the NE-VCIP problem, a VCI eman submitte by a customer consists of VDC infrastructures an the VN interconnecting VDCs. Each VDC requires a certain amount of computing resources (e.g. CPU an storage) an IT resources (e.g. ports for infrastructure connections within a VDC). The VN that connects VDCs requires a certain amount of banwith for ata transmission. The centralize scheuler nees to map the VDCs an VN to the geographically istribute DCs an backbone networks such that both have enough resources. To guarantee the banwith requirements for the VN, optical circuits are establishe an the spectrum is assigne to the emans. In this paper, we consier a backbone network which uses an IP-over- EON architecture as shown in Fig. 1. At the starting point of the ata transmission path, the ata traffic goes across the IP/MPLS layer noe to the connecte EON layer noe (banwith-variable wavelength cross-connects (BV-WXCs)) through banwith-variable transponers (BVTs). Then the ata traffic travels along the light path in the EON layer, arrives at the EON layer estination noe an finally reaches the en point of IP/MPLS layer. Therefore, to perform the VN mapping, an important task is to complete the routing an spectrum assignment (RSA) in the multi-layer network. EON is one of the most exciting future irections for optical networks an also an efficient an cost-effective solution for provisioning of Clou traffic [13]. A. VDC mapping For the VDC mapping, we suppose that no two VDCs in a same eman will be mappe to the same physical DC since we woul like to avoi the scenario of a isaster at one DC IP/MPLS Layer EON Layer Fig. 1: The IP/MPLS-over-EON architecture IP router BV-WXC Amplifier affecting multiple VDCs of a VCI eman. The geographically istribute DCs have ifferent amount of resources with ifferent rental prices. We assume that the DCs in the central region of the Clou network have lower rental price compare those in west/east regions, because of the richer resources an lower construction costs. B. RSA in EON layer The RSA problem in flexible gri optical networks contains both the routing ecision for traffic emans an the subcarrier assignment to satisfy the requirements by corresponing traffic emans [14]. The VN mapping in the EONs is actually a RSA problem, which is NP -har [15]. For the VN requirement of a eman, the central scheuler nees to fin the path between two geographically istribute DCs that has the lowest cost an make sure that all the fiber links along this path have enough spectrum resources. Then the scheuler assigns the relate frequency slots (FSs) from each fiber link along the path for the emans. The require number of FSs must be contiguous in frequency omain an temporal omain for each link on the path. In aition, the links along the routing path must use the same FSs, which is calle spectrum continuity. In our moel the require banwith of a virtual link (VL) in VN is given in bit rate (Gbps). In orer to estimate the number of FSs that each VL requires, we convert the require bit rate banwith into frequency (GHz) first accoring to the theoretical banwith efficiency limits for the main moulation formats [16]. Formula F = B/M is use in the conversion. Here F is frequency; B is bit rate; M is log-2 value of moulation formats, M = 1, 2, 3, 4 for Binary Phase Shift Keying (BPSK), Quarature Phase Shift Keying (QPSK), 8 Quarature Amplitue Moulation (8- QAM) an 16 Quarature Amplitue Moulation (16-QAM) respectively. The assume transmission reaches of moulation formats BPSK, QPSK, 8-QAM an 16-QAM are 5000, 3000, 1500 an 700 km respectively [17]. Then accoring to the frequency gri of the EON, the require number of FSs of a VL can be obtaine. In EON, the available optical spectrum is ivie into a set of FSs of a fixe finer spectrum with (frequency gri), such as 25GHz, 12.5GHz an 6.25GHz. We use 12.5 GHz as the frequency gri for the operation of the EON in this work with a total of 320 FSs in the C-ban on each fiber link. IV. A. NE-VCIP Problem Setting MATHEMATICAL FORMULATION The objective of the NE-VCIP problem is to achieve the minimal cost while satisfying the input emans.

3 Given: A physical Clou computing infrastructure, moele as a weighte irecte graph G(V, E), V is the set of DCs with a set of computing resources an their unit costs, E is the set of network links. Each DC is escribe as a tuple ata center = (DC v, C v, S v, P v, α v, β v, γ v, i v, e v ) with the capacity an unit cost of types of resources in this ata center, the meaning of each item in the tuple is escribe in Table I. Each ege in E is escribe as a tuple e = (u, v, (u,v) ), which inicates the link between DCs (u, v), an the link istance. The network in the moele Clou computing infrastructure is in multi-layer; A VCI eman, moele as a weighte irecte graph G (V, E ), in which V is a set of VDCs with specifie computing, storage an switch port requirements, E is a set of weighte VLs that inicate the require banwith. Each VDC of V is escribe as a tuple VDC = (, v, RC v t, t RSv, RP v t ), in which inicates the eman ID an v inicates the VDC ID of eman ; the meaning of other items in the tuple is escribe in Table I. Each VL in E is escribe as V L = (, u, v, RB (u,v )t ), in which u, v inicate the two en VDCs of current virtual link. The cost for each optical amplifier (OA) an regenerator to be installe in the use fiber links an EON noes respectively; an the cost for banwith-variable transponer (BVT) at each IP/MPLS noe for connecting optical layer noe. Both will be escribe in Section IV. B in etail. The moulation format for optical signals in EON layer. Output: The mapping for the VDCs in each VCI eman to the physical DCs; The routing path for mappe VN with allocate FSs; The total cost for satisfying all emans. Objective: Minimize the total cost for all the VCI emans (Full Fit), or for satisfying the maximum number of VCI emans (Best Fit) if accepting all emans is not possible. C v, S v, P v α v, β v, γ v i v, e v b e ST, ET V E TABLE I: Parameters CPU, storage an switch port capacities in DC v, v V The unit cost of CPU, storage an switch port in DC v The cost of IP/MPLS, EON layer terminals in DC v The unit cost (per Gbps) of banwith resource The istance of link e, e E The start an en time of eman The set of VDCs by eman The set of VN-links by eman Bu The buget of eman RC v t, RSv t, The require amount of CPU, storage an switch port for VDC RP v t v by eman in time slot t, v V RB (u,v )t T (u,v )t eg v B. Network Moel The require amount of banwith of virtual link between VDC (u, v ) by eman The require number of frequency slots by eman between VDC (u, v ) The egree of VDC v in VCI topology by eman In this work we aopt the IP/MPLS-over-EON architecture for the clou network as shown in Fig.1. In the IP/MPLS-over- EON network, the intermeiate noe along the routing path coul be (1) a multi-layer noe with both IP/MPLS an EON capability; (2) only a EON layer noe if the transmitte optical signal is not neee to be processe by the IP/MPLS layer; (3) a patch fiel that only connecting the optical fibers such as optical amplifier if the transmitte optical signal is not neee to be processe by neither IP/MPLS layer nor EON layer. In the IP/MPLS layer, an electrical noe which can be seen as an IP/MPLS router, consists of the main builing blocks: the basic noe (incluing switching matrix, power supply an mechanics), line cars (LC), with a ifferent number of ports for transceivers an the transceivers [18]. In the EON layer, a flexible EON noe can be seen as a banwith variable wavelength cross-connect (BV-WXC), which is use to establish optical cross-connections with various frequency slot with. The BV-WXC which mainly consists of BVT an banwith-variable wavelength selective switch (BV-WSS) can provie both sub-wavelength an super-wavelength for the flexible optical network. The EON can provie a granularity of 12.5GHz instea of 50 GHz in current WDM systems. BVTs can ajust the optical signal transmission rate to the actual traffic eman, by expaning or contracting the banwith of an optical path (i.e. varying the number of sub-carriers) an by moifying the moulation format [19]. C. Cost Moel In this work, the cost we consiere for the NE-VCIP problem comes from the renting cost for computing resources such as CPU an storage (note as operating expeniture (OpEx) in this work), an the fixe cost for network equipments an fibers (note as capital expeniture (CapEx) in this work). For the OpEx, we refer to the Amazon EC2 cost moel to give the unit rental cost (cost per resource unit per time slot) of CPU, storage an banwith. For the CapEx, we refer to the cost moel in [20] for the cost of IP/MPLS noes, BVTs, optical amplifier, etc. We assume that the metro noe in our topologies aopt a single-chassis router, which consists of a single shelf with 10 line-car slots. All the cost values in our work are normalize. D. MILP Moel Next we will escribe the MILP formulations of VCI mapping while consiering RSA in EON problem. The MILP moels we constructe for the VCIP problem is Full-Fit, in which we restrict that all VCI emans in the eman set must be accepte an the total cost of resource provisioning for all emans must be minimize. In general the physical frequency filtering requires that various spectrum paths are separate in the spectrum omain by guar frequencies when two spectrum paths share one or more common fiber links. In our problem, to simplify the moel, we assume that the size of guar frequencies is zero. To construct MILP formulations, we efine some variables as shown in the following. - x v v, 1 if require VDC v by eman is mappe to DC v; 0 otherwise - y (u,v) f(u,v ), 1 if the FS f is use on physical link (u, v), which is on the mapping path for virtual link (u, v ) of eman ; 0 otherwise. (u, v) E, (u, v ) E - COST, the cost for eman

4 Objective: COST = (RC v t t,v,v + v,v + e,e Minimize D COST (1) α v + RS v t β v + RP v t (i v + e v ) x (v,v) eg v + RB e t b t,e y (e,e) e comcost RB e t /10 γ v ) x (v,v) where comcost integrates the OA, regenerator, etc. along the fiber links, t [ST, ET ], v V, u, v V Computing Resource Capacity Constraints: RC v t x (v,v) C v (3),v RS v t x (v,v) S v (4),v RP v t x (v,v) P v (5),v where t [ST, ET ]. Equations (3)-(5) ensure that the assigne computing resources to the eman o not excee the resource capacity of each noe. Resource Allocation Region Constraints: x (v,v) = 1, D, v V. (6) v x (v,v) v (2) 1, D, v V. (7) Equation (6) guarantees that one VDC of a VCI eman can only obtain resources from a single physical DC. Equation (7) guarantees that a physical DC can only have at most one VDC of a eman to be assigne to itself. Spectrum Continuity Constraint: y (u,o) f(u,v ) u xu T (u,v )t = 0, y (i,u) f(u,v ) = 0 (8) f f y (i,v) f(u,v ) xv v T (u,v )t = 0, y (v,o) f(u,v ) = 0 (9) f,j v y (i,j) f(u,v ) = f,j u y (j,o) f(u,v ) (10) where i, o, j V, t [ST, ET ]. We inicate that u, v are the source an estination noes of the mapping route for VL (u, v ). Equations in (8) guarantee that the number of output frequency slots from the source noe equals the require number of frequency slots, an no input flows to the source noe. Equations in (9) guarantee that the number of input frequency slots to the estination noe equals the require input frequency slots, an no output flows from the estination noe. Equation (10) ensures that the spectrum route uses the same spectrum along the routing path. Frequency Slot Consecutiveness Constraint: (y (u,v) f(u,v ) y(u,v) (f+1)(u,v ) 1) ( N) f y (u,v) f (u,v ) where f [1, F 1], f [f + 2, F ], u, v V, u, v V. (11) Equation (11) ensures that the employe frequency slots are consecutive in the frequency omain. The FS consecutiveness constraint requires that, for a spectrum route, the allocate FSs are consecutive in frequency omain. This constraint can be equivalently converte to: if y (i,o) f(u,v ) = 1 an y(i,o) (f+1)(u,v ) = 0, all FSs with inex higher than f + 1 will not be allocate to the VL (u, v ) from fiber link (i, o). We introuce a large number N in this constraint. Frequency Slot Capacity Constraint: 1, f, i, o (12),u,v y (i,o) f(u,v ) Equation (12) ensures that one frequency slot on a fiber link can only be use by one route in a time slot.,f,u,v y (i,o) f(u,v ) F N, i, o V (13) Equation (13) ensures that the use frequency slots cannot excee the spectrum capacity (note as F N) of each fiber link. V. HEURISTIC ALGORITHM We propose a cost-optimize greey heuristic for the NE- VCIP problem. The heuristic works for both Full-Fit an Best- Fit scenarios. In full-fit, all emans will be accepte an the total cost for all emans will be minimize. In bestfit, we accept as many emans as possible an minimize the total cost of accepte emans. Every VCI eman is generate ranomly with start time, finish time in [0,24], with require banwith an computing resources. In our propose heuristic, we o not separate the joint resource allocation into two phases: computing resource phase an banwith resource phase, but combine them together. The reason for this is to utilize the banwith resource more efficiently since we foun from previous experiments that the network resource is the bottleneck (compare to computing resources in each DC) in completing the joint resource allocation for emans. The general ieas of the propose cost-optimize greey heuristic are: (1) Map the first VDC (e.g. v1) of a eman, map it to the DC (e.g. u) which has enough computing resources an has lowest resource unit cost; (2) check if v1 has connections with other VDCs (e.g. v2) in the VCI eman graph; (3) if yes, for each connection, the Dijkstra algorithm is aopte to fin the shortest path p(u, es) between u an every other DC, an sort the paths in ascening orer of istance; if no, go to (5); (4) map v2 to es which is the estination DC of the shortest path if es DC has enough computing resources an all links along path p have require number of FSs; (5) continues until a VCI eman is processe, then go for the next eman. The algorithm etails are shown in Algorithm 1. We call the Dijkstra algorithm

5 whose time complexity is O( E + V log V ) in our propose heuristic. The total time complexity of the propose heuristic is O(( E + V log V + V 2 ) D V ). Algorithm 1 Cost-optimize Greey Algorithm Input an Initializations: G(V, E) // network topology D // eman set G (V, E ) // virtual topology of eman Output: Minimize Cost. 1: Sort V in ascening orer of computing resources cost 2: for all D o 3: for all v V o 4: if v is not been processe then 5: Map v on noe v with enough resources for v ; 6: Allocate computing resources from v for v ; 7: Upate Cost ; 8: en if 9: Construct set P v ; 10: for all u V, u v o 11: Fin shortest path p(v, u), a p(v, u) into P v ; 12: en for 13: Sort paths in P v in ascening orer of istance; 14: for all u Ajacent(v ) o 15: if u is not been mappe then 16: for all p(v, u) P v o 17: if enough resources on noe/path then 18: Map u on noe u; 19: Allocate computing resources for u ; 20: Allocate spectrums an upate Cost ; 21: en if 22: en for 23: else {// suppose u is alreay mappe to x V } 24: if p(v, x) has enough spectrums then 25: Allocate spectrums an Upate Cost ; 26: else 27: Drop eman, release resources; 28: en if 29: en if 30: en for 31: en for 32: en for 33: return Cost VI. EXPERIMENTAL RESULTS AND ANALYSIS We carry out the simulations for ILP moel (using IBM ILOG CPLEX Optimization stuio) an cost-optimize greey heuristic on a cluster noe which has 2 CPUs/16 cores an 64GB memory with Linux system. Two network topologies are teste for the simulations: a 6-noe topology (Google DC locations) an NSFNET topology shown in Fig. 2 with the istance in km. In the simulations, each VCI eman is generate ranomly, which means the value of each item in a VCI eman tuple (as escribe in Section IV. A) is generate ranomly. In general, in our simulation, we assume that each VCI eman consists of no more than 5 VDCs. The correctness of the propose greey heuristic is verifie by comparing its results with the ILP results for the small ata set for the 6-noe topology as shown in Table II. When given one eman, the heuristic can give the near optimal solution compare with that of CPLEX an the computing time is much more less than that of CPLEX. When given two or more emans, the CPLEX converges much slower to generate optimal solution compare to the heuristic metho. In this case, in the later experiments, larger ata sets are only teste by the heuristic on two topologies ue to the slowness of ILP solution by CPLEX Fig. 2: NSFNET network topology TABLE II: Cost an time comparison between CPLEX solver an heuristic # of emans Total Cost (normalize) Running Time CPLEX Heuristic CPLEX Heuristic hours s (gap 18.56%) hours s (gap 97.34%) hours s s s We compare the total cost an eman blocking rate for ifferent ata sets with ifferent moulation formats. Due to the space limitation, here we only list the comparison results for the NSFNET topology; similar results are also obtaine on the 6-noe topology. In Fig. 3, all the emans can be accepte (full-fit) an allocate resources from the Clou by the resource allocator. By observing Fig. 3 we can see that for ifferent sizes of the eman set, the total costs ecrease with the moulation format orer of BPSK, QPSK an 8-QAM, since the require number of FSs of each eman is reuce. But the total cost with 16- QAM increases compare to that with 8-QAM, although each eman has the least number of require FSs with 16-QAM. We note that the require number of FSs for a given bit rate is reuce sequentially with the moulation format orers of BPSK, QPSK, 8-QAM an 16-QAM; an at the same time the optical signal reaches are reuce along the same moulation format orer. In this case, more regenerators are neee along the optical path to regenerate the signals an the total cost increases instea. It seems that it is a better choice to aopt 8-QAM moulation format to get a lower total cost. With the limite resource capacities, the resource scheuler will rop some emans that cannot be satisfie when the number of emans increases. During the experiment we foun that the network resource is a bottleneck compare with other computing resources. Almost every eman that is roppe is

6 ue to lack of continuous spectrum resource along its optical path. We observe the blocking rate of ifferent sizes of eman set with four types of moulation formats as shown in Fig. 4. It is obvious that for the moulation format orer of BPSK, QPSK, 8-QAM an 16-QAM, the require number of FSs for each eman reuces significantly, so that the resource scheuler can accept many more emans. When the input number of emans reaches 1800, the blocking rates are nearly 13.6%, 0.7%, 0.022% an 0 with BPSK, QPSK, 8-QAM an 16-QAM respectively in Fig. 4. Thus, while consiering the total cost an blocking rate together, we foun that 8-QAM in our experiment performs best, which has the lowest total cost an has the blocking rate close to 0 for larger ata sets. cost an also obtain a better system performance (e.g., low blocking rate, high resource utilization), a trae-off nees to be consiere between the two. In our experiments, we conclue that aopting 8-QAM in EON layer woul be a suitable choice for the resource scheuler to obtain the lowest total cost an also obtain an acceptable low blocking rate. In this work, we only consier using 10 Gbps BVT in EON layer, an we will investigate the effects of BVTs with ifferent bit rates in our future work. In aition, we will consier the ynamic traffic for the emans an also the impact on the total cost by traffic grooming in EON. ACKNOWLEDGMENT This work is supporte by the U. S. National Science Founation uner Grant No. CNS Total cost (c.u.) Blocking Rate BPSK QPSK 8-QAM 16-QAM Number of emans Fig. 3: Total cost comparison (NSFNET) BPSK QPSK 8-QAM 16-QAM Number of emans Fig. 4: Blocking rate comparison (NSFNET) VII. CONCLUSION In this work, we propose an investigate the NE-VCIP problem in IP-over-EON network architecture. An ILP mathematical moel is constructe an a cost-optimize greey heuristic is evelope to solve the NE-VCIP problem. Different moulation formats that are aopte in the EON layer will have ifferent results for the total cost an the eman blocking rate for the same ata set size. So in orer to minimize the total REFERENCES [1] S. Peng, R. Nejabati, an D. Simeoniou, Role of optical network virtualization in clou computing, IEEE/OSA JOCN, vol. 5, Oct [2] vclou Suite. [Online]. Available: [3] Cisco Virtualize Multiservice Data Center. [Online]. Available: vmc.html [4] Cisco, Cisco global clou inex: forecast an methoology, , White Paper, Cisco, [5] C. Kachris an I. Tomkos, Optical interconnection networks for ata centers, in ONDM, April 2013, pp [6] C. Guo et al., Seconnet: A ata center network virtualization architecture with banwith guarantees, in Co-NEXT 10, New York, NY, USA, 2010, pp. 15:1 15:12. [7] C. Papagianni et al., On the optimal allocation of virtual resources in clou computing networks, IEEE Transactions on Computers, [8] Z. Xiao, W. Song, an Q. Chen, Dynamic resource allocation using virtual machines for clou computing environment, TPDS, [9] M. Alicherry an T. V. Lakshman, Network aware resource allocation in istribute clous, in INFOCOM, March 2012, pp [10] M. G. Rabbani et al., On tackling virtual ata center embeing problem, in IFIP/IEEE International Symposium on IM, May 2013, pp [11] M. F. Zhani et al., VDC planner: Dynamic migration-aware virtual ata center embeing for clous, in IFIP/IEEE International Symposium on IM, May 2013, pp [12] A. Amokrane et al., Greenhea: Virtual ata center embeing across istribute infrastructures, IEEE Transactions on Clou Computing, [13] M. Klinkowski an K. Walkowiak, On the avantages of elastic optical networks for provisioning of clou computing traffic,, IEEE Networks, [14] C. Politi et al., Routing in ynamic future flexi-gri optical networks, in ONDM, April 2012, pp [15] Y. Wang, X. Cao, an Y. Pan, A stuy of the routing an spectrum allocation in spectrum-slice elastic optical path networks, in INFOCOM, April 2011, pp [16] F. Xiong, Digital Moulation Techniques. Artech House, Inc., [17] Infinera, Super-channels: DWDM transmission at 100Gb/s an beyon, White Paper, Infinera, [18] F. Rambach et al., A multilayer cost moel for metro/core networks, IEEE/OSA JOCN, pp , March [19] V. Lopez et al., Target cost for sliceable banwith variable transponers in a real core network, in Future Network an Mobile Summit, July 2013, pp [20] O. Perola et al., CAPEX stuy for a multilayer ip/mpls-over-flexgri optical network, IEEE/OSA JOCN, pp , Aug 2012.