QoS aware Virtual Network Embedding in SDN-based

QoS aware Virtual Network Embedding in SDN-based Metro Optical Network Faisal Zaman1 , Abdallah Jarray2 , and Ahmed Karmouch

3

1

University of Ottawa, Ottawa, Canada, [email protected] University of Ottawa, Ottawa, Canada, [email protected] SITE, University of Ottawa, Ottawa, Canada, [email protected] 2

3

Abstract. A framework using Software Defined Networking (SDN) architecture to map the Virtual Network (VN) requests onto Edge Clouds interconnected by Metropolitan Optical Network (MON) is presented in this paper. As the proposed mathematical model relies on SDN for making an informed decision, the framework is termed as, Virtual Network Embedding on Metro Optical Network (VNE-MON). The proposed optimizer considers distinct characteristics of Optical Network and Multimedia applications to ensure Quality of Service (QoS) for the accepted VN requests. SDN enabled ensuring of proper coordination between different modules of the proposed framework. The proof of concept was demonstrated on a SDN and Generalized Multi-Protocol Label Switching (GMPLS) testbed. The two control planes were evaluated on different QoS metrics after VN embedding phase. Numerical result reveals that the proposed SDN-based model outperformed the current state of the art control plane GMPLS for optical network.

1

Introduction

Due to the increasing demand and nature of new classes of multimedia applications traffics, Dense Wavelength Division Multiplexing (DWDM) is now beginning to expand from a network core technology towards the Metropolitan area network. The focus is beginning to shift towards Metro Optical Networks (MONs), with the goal of bringing the benefits (cost and networks efficiency) of the optical networking revolution to the end-users. MON plays a crucial role in giving seamless experience to the end consumers of the multimedia services. To further enhance the multimedia application experience to the users, a new class of cloud architecture is introduced. This new class of cloud called Media Edge Cloud Data Centres (MEC-DCs) [1] are kept at the last mile, closer to the users. It aids in reducing the latency and cost for communication. To increase the utilization of MEC-DCs connected by MON, virtualization of network is introduced. Network Virtualization helps to host multiple users on the same substrate network without affecting the agreed Service Level Agreement (SLA). One of the challenge faced by Metro Carrier Service Providers (CSP) is that, for a given service they not only have to consider the availability of network resources, but also take the account of SLA for different applications and services

[2]. Traditionally managing of cloud resources and network resources were done by separate teams. Moreover in a MON the provisioning of network resources were performed manually. The confluence of Software Defined Networking and Virtualization along with the introduction of Reconfigurable Add Drop Multiplexers (ROADMs) has created the perfect opportunity for the Metro CSP to enhance the utilization and guarantee the Quality of Service (QoS) through automation. SDN brings flexibility for Network Virtualization, as each VN’s control logic runs on a centralized controller rather than on a physical switch [3]. One of the long investigated challenge associated with Virtualization is the placement of Virtual Networks (VN). This challenge is called as VN embedding problem. Prior to SDN several proposals were made to solve the VN embedding problem, but they were either highly theoretical or not compatible with SDN-based network [4]. More recently, [6]- [9] have proposed SDN-aware VN embedding solutions. These approaches are proposed for either packet switching-based network (also known as Core Network) or a generic VN embedding solution using SDN paradigm. Since aforementioned proposals do not consider the DWDM-enabled Optical Nature of the substrate link and instead consider the availability of link as aggregated bandwidth can result in solution which will ultimately be non-feasible. In other words, for a packet switching domain considering the link resource as a total available bandwidth can be used, but in case of optical network each optical wavelength has a certain capacity, if ignored to consider these factors might result in higher acceptance but ultimately an in-feasible solution. In this paper, the multimedia requirement and substrate requirement to guarantee the QoS for the accepted service are considered. While doing so, the profit to the service provider is ensured in this competitive environment of cloud service providers. In order to address these challenges, a framework of VN embedding for an SDN controller is designed. VN Embedding on SDN-based MON (VNEMON), considers the distinct characteristics of the substrate network along with specific requirements for Multimedia VN requests. VNE-MON uses a combination of Integer Linear Programming (ILP) and Column Generation (CG) technique for optimization and orchestration of substrate resources to the incoming VN requests. This ensures that an optimized solution for VN embedding is obtained in least number of CPU cycles. VNE-MON’s feasibility is demonstrated by means of simulation through standard tools and topologies. Performance results are compared to the current state of the art control plane for the optical network Generalized Multi-Protocol Label Switching (GMPLS). The rest of the paper is structured as follows. Section II provides an overview of the related works. Section III explains our proposed SDN based model for VN embedding, i.e., VNE-MON. Section IV evaluates the performance of VNE-MON and is compared with the distributed GMPLS-based VN embedding approach. Finally, Section V concludes the paper and provides the perspective for future work.

2

Related Work

The abstract global view provided by SDN/OpenFlow architecture is leveraged by many researchers to optimally utilize the available network resources. Authors in [5] have proposed VN mapping on optical network using the reactive approach, where the allocation is triggered during a network failure. Authors in paper [6] have made use of failure probability factor in their mathematical model to decide on the Virtual Optical Network Location. [7] proposes a dynamic VN mapping based on switch load, and switch factors like memory and storage. [7] uses ILP technique and focused on optimizing the network performance. Authors in [8] considers the computational attributes of VN request along with link and switch requirement, the main objective of their proposal is to increase the acceptance ratio and infrastructure service provider’s revenue. [9] has used Mixed integer programming for VN mapping to a substrate network. This model allows for partial embedding of VN requests, and least preference was given to QoS with the objective of increasing the profit. In the light of aforementioned works, the main contribution from this paper are: (a) A Virtual Network embedding approach on a Metro Optical Network consisting of Media Edge Clouds (MEC- DC) to provide guaranteed QoS. (b) Proposed an SDN framework for integrating VN embedding and to ensure proper coordination between different embedding modules. Finally, (d) The performance of VNE-MON compared to distributed architecture GMPLS-based VN embedding is evaluated with regards to QoS metrics.

3 3.1

Resource Allocation Framework for SDN-enabled MON Overview of the Model

The proposed VN embedding model for SDN-based network is shown in Fig 1. The network resource allocation mechanism is built as an application on top of SDN controller. The SDN controller ONOS is modified. Modified ONOS controller is used to collect link utilization information of optical links. Whereas, for the rest of the communication REST APIs are used. The application acts as a brain to the ONOS controller. The important modules and their functions of the application are explained briefly: 1. Available Network Resources: This is a simple Database which has live updates of the available network resources. The network resources include bandwidth and wavelength availability on each link, the degree of freedom, number of available ports, and memory available at each Reconfigurable Add Drop Multiplexer (ROADM), and available datacenter resources i.e. Storage, Memory, RAM, CPU and GPU. 2. Topology Database (DB): Any change in network configuration (adjustment of the network elements) is updated with the assistance of the topology database. This DB stores and maintains the nodes and links along with their characteristics.

Virtual Network Requests VN 2 VN 1 Modules of Resource Allocation Application

QoS module

CG Link Optimizer

ILP Node Optimizer

Buffered VN request batch

Available Network resources

Topology DataBase

Network Resource Allocation Application as a SDN application Northbound APIs – REST API, JAVA API

Abstraction Layer SDN Controller for transport Network

Network Recovery Module

OpenFlow Ver. 1.3.0

Substrate Network

Fig. 1. Proposed SDN-based Virtual Network Embedding in Metro Optical Network

3. Buffered VN requests DB: This buffer keeps the incoming VN requests to which the resources are allocated in the next time period. The size of the buffer is decided based on the VN requests incoming rate. Higher the buffer size better the network utilization, but at the cost of delayed embedding of VN requests. In contrast, if the buffer size is small there is faster embedding of VN requests with inefficient utilization of substrate resources. 4. ILP Node Optimizer: This module performs optimization according to the defined Integer Linear Programming (ILP) mathematical model. The input for this module is taken from the Topology DB and Buffered VN requests DB. Proposed ILP model for VN requests embedding is discussed in detail in the next Section. 5. CG Link Optimizer: The output from the ILP Node Optimizer is fed to Column Generation (CG) optimizer. This facilitates in selecting the best mapping configurations for a given set of VN requests. CG technique of optimization is known for using least number of CPU cycles to arrive at optimal solution. 6. QoS Module: This module is utilized to sort the incoming VN requests into different classes based on QoS requirements. The QoS are classified based on the resource requirement of the VN requests. For the purpose of evaluating the model we define a five different QoS classes. This module can be configured based on the service providers need and the traffic associated with the selected metro network.

7. Network Recovery Module: This module is an integral part of SDN controller, separate from the optimizer. This ensures faster recovery and less CPU cycles utilization during the link failure. The strategy considered in this proposal consists of bundle approach [11]. Bundle approach aids in avoiding spectrum and node contentions, which is a very common challenge faced by GMPLS based network. Due to the expensive and complex nature of Shared Risk Link Group (SRLG) approach considered in GMPLS-based network, evaluation of Network Recovery Module is beyond the scope of this paper. 3.2

CG-ILP Based Formulation for VN request mapping on MEC-DC and WDM Network

The presented formulation is an extension of our previous proposals [16] and [17]. The major changes from the work to support SDN and ROADM technology considered here are: 1. Inclusion of ROADM technology and Optical Transport Network. 2. A max-min formulation based approach to ensure profit to CSP as well as least cost to end user. 3. A combination of Linear programming and Column-Generation to reduce complexity of the formulation and to bring agility in solving the problem. 4. Multimedia aware requirements. ILP based Node Embedding for MEC-DC The Substrate network can be defined by an undirected graph Ks = (Hs , Ls ). Where, Hs represents set of substrate nodes and Ls set of links. The substrate node u ∈ Hs , can be defined by, u = {Ru , Su , Gu , Cu } where, Ru , is the Memory, Su is the storage capacity, Gu is the GPU, and Cu is the CPU capacity of the node u. The residual CPU, r/s/g/c (t). Hence, Memory, Storage and GPU capacity at time t is given by, Pu at any given time the available node capacity can be defined by a set Qu = {Pur (t), Pus (t), Puc (t), Pug (t)}. A VN request can be represented as Kv = (Hv , Lv ) where, Hv represents set of virtual nodes and Lv represents directional Links. The QoS requirement of a virtual node a ∈ Kv can be defined as:(QoS)a = r/s/g/c represents required RAM, Storage, GPU and CPU at a (pra , psa , pga , pca ) Pa selected substrate node for processing the selected VN request. Similarly, for a virtual link e ∈ Lv , can be defined by a set: (QoS)e = (be , de ), be and de define the minimum required bandwidth and delay for the selected virtual request respectively. xua and zn are binary variables. xua is 1 if a virtual node a is assigned to substrate node u, 0 otherwise. Similarly, zn is a binary variable which is 1 if a virtual network request is accepted, 0 otherwise. Objective F unction : To increase the revenue by maintaining QoS requirement for media requests, i.e., maximizing profit. Consider, Pen ∗ zn as the total revenue generated by use of substrate resources, π(uv) is the shortest path for (e = sd) between source u and destination v, cl is the unit cost of using a substrate

link l and fu + fv is the cost of using datacenter resources at u and v. Then objective function can be defined as: ( fobj =

)

X

X

n∈N

e∈Lv ;e=(sd)

X

Pen ∗zn −

(u,v)∈Hs ∗Hs

xus ∗xvd ∗(fu +fv +

X

cl ∗be )

l∈π(uv)

(1) The Resource allocation of VN request on a DC is governed by the following constraints: 1. All accepted virtual requests should be embedded on the substrate network, which is given by X xus ∗ xvd ; (sd) = e ∈ Ls zn 6 (2) (u,v)∈Hs ∗Hs

2. No more than one virtual node for a given VN request can be assigned to the same substrate node (Eq. 3). tc (a) is a subset of Hs , represents the potential locations that can satisfy the VN’s datacenter requirement. DC requirement to be satisfied for a given virtual node a are governed by the Eqs. (4) - (7). X xus 6 zn ; a ∈ Hv , n ∈ N (3) u∈tc(a)

X X

xua ∗ pca 6 pcu ; u ∈ Hs

(CP U Constraint)

(4)

xua ∗ pga 6 pgu ; u ∈ Hs

(GP U constraint)

(5)

xua ∗ psa 6 psu ; u ∈ Hs

(Storage Constraint)

(6)

xua ∗ pra 6 pru ; u ∈ Hs

(RAM Constraint)

(7)

n∈N a∈Hv

X X n∈N a∈Hv

X X n∈N a∈Hv

X X n∈N a∈Hv

Light path embedding for the virtual request using Column Generation Technique: Column Generation makes use of duality nature of ILP to solve large optimization problems. Mapping of VN link based on CG approach e can be defined as Ml : Lv 7→ Πuv . Where, Πuv is the shortest path for the virtual link e = (sd) ∈ Lv between substrate source u and destination v. The Πuv ∈ π(uv) where π(uv) is calculated using K-shortest path algorithm [19]. CG approach is decomposed to master problem and pricing problem. Master Problem Objective Function is to reduce the cost of using a substrate network for a given VN request and to maintain the QoS. X M in (Costc ∗ λc ) (8) c∈C

Where, c ∈ C is defined as independent mapping configuration for each VN request, this ensures that each mapping of a VN request is independent of each other in a WDM based optical Network. λc is a binary variable for master problem, which is 1 if the VN request is satisfied by the configuration c, 0 otherwise. Costc defines the unit cost of configuration c: X X B c (l) ∗ cl (b) (9) Costc = T c (u) ∗ cROADM + c∈C

l∈LS

Where, cROADM is the unit cost of using a ROADM, cl (b) is the unit cost of bandwidth, T c (u) is the total number of ROADMs used in selected configuration c and B c (l) is the used bandwidth by the link l in configuration c. Master problem is governed by the following constraints: X λc 6 w (αo ) (10) c∈C

X

λc ∗ T c (u) 6 NROADM s (u)

(µu )

(11)

X

(βu )

(12)

c∈C

λc ∗ anc 6 1; u ∈ Hs ; n ∈ N.

c∈C

Eq. (10) defines that configuration used by a VN request should be unique. The maximum number of configuration supported by the system is equal to total number of wavelengths w. Eq. (11) assures that the signal quality is not degraded due to Optical - Electrical - Optical (OEO) conversions. Eq (12) ensures that a maximum of one configuration can be used for mapping of each VN request. Pricing Problem It is used to generate columns. Column is generated based on the constraints (10) - (12). The objective f unction is defined as: X X Costc = Costc + αo + µu ∗ T c (u) − anc ∗ βu (13) u∈Hs

n∈N

The relation between master and pricing problem is defined by the following equations: zn = anc

(14)

T (u) = 2 ∗ yu

(15)

c

Where, zn and yu are the binary variables for pricing problem. zn is 1 if VN request is served by configuration c, 0 otherwise. Similarly, yu is 1 if ROADM is installed in node u, 0 otherwise. Pricing problem is governed by following Constraints: 1. All the virtual links have to be accepted for a given VN request, otherwise VN request will be rejected. 2. ROADM requirement to satisfy the VN request without disrespecting the QoS agreement should be maintained.

3. The assigned substrate link should satisfy the bandwidth and wavelength requirement of a given VN path. 4. Grooming factor is defined as the total amount bandwidth that can be pushed into a given wavelength. It is dependent on Optical Transport Unit (OTU) scheme used. For the sake of simulation the grooming factor is kept constant. OTU-1 is used for each VN request and each substrate link has a capacity defined by OTU-3 standards.

4 4.1

Performance Evaluation Simulation Setup

Performance evaluation of the VNE-MON is done by comparing with GMPLSbased network. For generating traffic an in-house built traffic generator in C++ is used. Generated traffic at each time period consists of n number of requests, wherein each request consists of a set of DCs and Virtual links between them. Several Topologies from topology zoo [18] are considered for the evaluation which are fibre and closest to be considered as a metro network. The whole setup was run on a computer with intel core i7 processor, clock speed of 3.1 Ghz and 16 GB RAM. The simulation of VNE-MON and GMPLS based models are explained briefly in the subsequent Sections. 1. Test Bed For VNE-MON: CPLEX is used to calculate the optimal solution based on the proposed CG-ILP model. The Optimization algorithm is written in C++. It is communicated to the SDN/OpenFlow Network through REST APIs. ONOS as a SDN controller is used. ONOS is modified to support the CG-ILP model. Topologies were created using Mininet with the help of LINC-OE switches. These switches emulate the behaviour of CDC (colourless directionless and contentionsless) ROADMs. 2. Test Bed for GMPLS-based Network Resource Allocation: In case of GMPLS, which is responsible for the performance of the link, only the end nodes are provided and Open Shortest Path First (OSPF) calculates the best paths for the selected nodes. The GMPLS testbed is simulated by using GMPLS Light Agile Switching Simulator (GLASS) [15]. GLASS has all the necessary protocols and dependencies required for this implementation. 4.2

Evaluation Metrics

This paper attempts to compare QoS metrics for VNE-MON and distributed GMPLS testbed. The QoS metrics used for measurement are (a) Resilience factor, (b) Topology Discovery time, (c) Throughput, (d) Jitter and (d) Acceptance of VNs ratio versus cost of wavelength conversion. 4.3

Performance Measurement

In the figure (2) Case - A corresponds to the scenario where we consider the pricing problem constrains as well as Eqs. (11) - (12) whereas, Case B- represents the situation where we ignore aforementioned Equations. The simulation

100

0.8

90

0.7

80

0.6

70 60

0.5

50

0.4

40

0.3

30

0.2

20

Drop Ratio of Requests

Percentage utilization of Link Bandwidth

was performed for several number of rounds with various types of VN requests having different QoS requirement such as Bandwidth, link latency and node requirement. Considering the wavelength does reduce the acceptance ratio and utilization but that does not degrade the QoS and guarantees the QoS for the users. In contrast, the Case B failed to assure QoS for the users.

0.1

10 0

0 1

2

3

4

5

6 7 Time Period

8

9

10

11

12

Case A - Utilization

Case B - Utilization

Case A - Drop Ratio

Case B - Drop Ratio

Fig. 2. Snapshot of jitter for VNE-MON vs GMPLS based Model

The round trip time (RTT) taken by an ICMP packet during a link failure for the VNE-MON and GMPLS based network is shown in Fig (3)(a). The link failure event is simulated by triggering port down event in the test bed. The results illustrates that the maximum RTT for the VNE-MON reached 5.3 ms, whereas in case of GMPLS it was 12 ms. GMPLS’s poor performance is because of OSPF’s convergence issue and overhead related to Resource Reservation Protocol (RSVP) for the bidirectional link. The convergence time of OSPF is mainly contributed by the hello timer interval (say ht ), the dead timer interval by industry standard is 4 ∗ ht plus the path recalculation time. The path re-calculation is a CPU intensive work, where network topology will be exchanged among nodes. Label Distribution Protocol (LDP) and RSVP have a role in setting up the new connections by brute force method. VNE-MON, is based on SDN architecture. When the controller senses failure notification it updates the path with available network data, thus saving time in collecting the network information. The reinstallation of flows is performed on the affected nodes, unlike in case of GMPLS where a complete new path is established. These factors play a crucial role in improving link failure recovery time. The Topology discovery time is directly proportional to the scalability factor of control plane. GMPLS was found to be stable for smaller networks, and as the network size increases the time taken to form the initial topology was found to increase exponentially (when the number of nodes increases for more than 12 nodes in case of our simulation). This is evident from the Fig (3)(b). In contrast, for the VNE-MON model, topology discovery time remained almost constant even after increasing the topology size. This is an important parameter when considering a network which has many edge DCs to be managed.

Topology Discovery Time

CONVERGENCE TIME VNE-MON

VNE-MON

14 12 10

TIme in ms

8 6 4 2 0 0

1

2 Time Period

3

10 15 Number of Nodes

20

25

Jitter for UDP packets 4

100

3.5

80

3

60

2.5

40 20

GMPLS enabled MON VNE-MON

2 1.5 1

0

1

11

21 31 41 51 61 71 81 Snapshot of Simulation Runtime in sec

91

0.5 0 0

VNE-MON

10

6 4 2 0

60

70

80

90 100 110 120 130 140

100

90

90

80

80

70

70

60

60

50

50

40

40

30

Single Two Three Three Hop Hop Hop Hop path path path path

30

20

20

10

10

0

0 1

VNE-MON

Times more than average RTT to Recover

Fig.(3) (e): Recovery Time

50

100

Acceptance Ratio

8

GMPLS

40

Snapshot of Simulation Runtime in sec

10

Single Two Three Three Hop Hop Hop Hop path path path path

30

3 (d) : Jitter

Resilence for GMPLS based Network and VNE-MON based Network

Average RTT

20

GMPLS enabled MON

3 (c): Throughput

Round trip time in ms

5

3 (b): Topology Discovery Time

Throughput for SDN Enabled network vs. GMPLS Enabled Network

Jitter in ms

Percentage available utilization of Link

0

4

3 (a) : Convergence Time

GMPLS

160 140 120 100 80 60 40 20 0

Percentage of Cost due to Wavelength Conversion

Round Trip Time delay in ms

GMPLS

2

3

4

5

6

7

8

Time Period % Cost for conversion in VNE-MON % cost of conversion in GMPLS Acceptance Ratio in VNE-MON

Acceptance Ratio in GMPLS Based MON

Fig.(3) (f) : Acceptance Ratio

Fig. 3. Performance Evaluation between GMPLS-based network and VNE-MON

It is evident from the obtained results depicted in Fig. (3)(c), GMPLS was highly unstable compared to the VNE-MON. Reason for the instability of throughput in GMPLS network can be labeled to it’s protocols used for traffic engineering. Routing and Wavelength Assignment uses best fit and OSPF algorithm, which aims at reducing the use of the number of wavelengths to as much as possible. The best fit leads to increased congestion of links, resulting in instability of the control plane for processing the incoming packets. Hence, due to link congestion, link utilization in GMPLS reduces drastically when all the labels in a link are active. In contrast, VNE-MON not only considers different factors like link cost and number of hops, it also uses the first fit algorithm for wavelength assignment along with ILP and Column generation. This assures that the

links are not overly crowded and avoids link congestion. Hence, evident from the obtained results, for the VNE-MON achieved a constant throughput. The obtained jitter values are depicted in Fig (3)(d). The reason for maximum jitter at around time period 100 sec is due to the simulation of link failure. It is clear that the maximum jitter of VNE-MON is more than maximum jitter of GMPLS-enabled MON. Maximum jitter in VNE-MON is because of the use of single SDN controller and its placement. When a link failure event occurs the packet in message has to reach the SDN controller and then the SDN controller updates the required nodes. But, in GMPLS when link failure event is initiated the overall jitter values changed from tightly-coupled to moderately-coupled representing instability. This effect is due to the fact that the GMPLS controller overloads the neighbouring link for faster recovery. To generalize the effect of link failure from different controller in use that is GMPLS-based system and VNE-MON system, we simulated link failures on different lengths of paths between the DCs. The length of path is defined by the number of hops in the considered path. As it is clearly evident from the results obtained, in case of GMPLS-based network, convergence time increased exponentially with the increase in the number of nodes, whereas, in case of VNEMON the convergence time was almost static. Thus even though in VNE-MON a remote controller is used, it did not have adverse effect on the convergence compared to GMPLS-based network. This plays a crucial role while ensuring path QoS for an accepted VN request. Fig (3)(e) illustrates the same. Finally, acceptance ratio of VN requests between GMPLS and VNE-MON is depicted in Fig (3)(f). At first the acceptance ratio of both GMPLS and SDN based system looks similar from Fig 6, but due to the lack of global view GMPLS performed more number of wavelength conversion for a given path compared to SDN based system. This not only increases the cost of communication, but also degrades the performance. Hence, it can be concluded that even though GMPLS can be used for MON network it will be very expensive in terms of OPEX and complexity for providing VN embedding and resilience on MON.

5

Conclusion

In this paper, we have proposed an architecture for VN embedding on SDNbased MON. The proposed architecture uses Optimizer as an integral part of the SDN controller. A combination of ILP and CG technique is used for embedding the VN requests. Distinctive characteristics of optical network for MON and Multimedia VN requests are considered while performing the orchestration of substrate resource. Performance evaluation proves that SDN-based VN embedding for MON network i.e., VNE-MON is more stable and achieved better acceptance ratio of VN requests than its counter part (GMPLS). The global view provided by SDN controller also facilitated in a better load-balancing of wavelength. Which in-turn reduced the cost of wavelength conversion by reducing the O-E-O conversions. In short, although GMPLS can be used for MON network, due to lack of global view, it’s efficacy becomes highly unpredictable

with increase in network size. Future direction includes performing an in-depth evaluation on the recovery approaches in GMPLS-based network and SDN-based approach to provide resiliency to the accepted VN request.

References 1. W. Zhu and C. Luo and J. Wang and S. Li, “Multimedia Cloud Computing,” in IEEE Signal Processing Magazine, vol.28, no.3, 2011, pp. 59-69. 2. J. Reich et.al., “Leveraging SDN to streamline metro network operations,” in IEEE Communications Magazine, vol. 54, no. 10, October 2016, pp. 109-115. 3. McKeown, Nick et.al., “OpenFlow: Enabling Innovation in Campus Networks,” in SIGCOMM Comput. Commun. Rev, vol. 38, no. 2, April 2008, pp. 69-74. 4. N.M. Mosharaf Kabir Chowdhury, Raouf Boutaba “A survey of network virtualization, Computer Networks,” Comput. Netw.,Volume 54, April 2010, pp 862-876. 5. F. Gu et.al., “Virtual network reconfiguration in optical substrate networks,” in Optical Fiber Communication Conference and Exposition and the National Fiber Optic Engineers Conference (OFC/NFOEC), 2013, Anaheim, CA, 2013, pp. 1-3. 6. Jiachen Ma et.al,“Virtual optical network embedding with sdn architecture based on memetic algorithm,” 4th International Conference on Computer Science and Network Technology (ICCSNT), Harbin, 2015, pp. 1050-1053. 7. M. Capelle et.al, “Online virtual links resource allocation in Software-Defined Networks,” IFIP Networking Conference, 2015, Toulouse, 2015, pp. 1-9. 8. X. Wen et.al., “An Efficient Resource Embedding Algorithm in Software Defined Virtualized Data Center,” IEEE Globecom, San Diego, CA, 2015, pp. 1-7. 9. R. Guerzoni et al., “A novel approach to virtual networks embedding for SDN management and orchestration,” in IEEE Network Operations and Management Symposium (NOMS), Krakow, 2014, pp. 1-7. 10. Kai Chen et.al.,“OSA: an optical switching architecture for data center networks with unprecedented flexibility” IEEE/ACM Trans. Netw. 22, 2014, 498-511. 11. A. Giorgetti et.al., “Dynamic restoration with GMPLS and SDN control plane in elastic optical networks [Invited],” in IEEE/OSA Journal of Optical Communications and Networking, vol.7, no.2, Feb. 2015, pp. A174-A182. 12. P. Berde et al., “ONOS: Towards an Open, Distributed SDN OS,” in Proceedings of the Third Workshop on Hot Topics in Software Defined Networking, HotSDN ’14, Chicago, Illinois, USA, 2014, ACM, pp.1-6. 13. “FlowForwarding/LINC-Switch”, GitHub, 2017. [Online]. Available: https://github.com/FlowForwarding/LINC-Switch. [Accessed: 23- Jan- 2017]. 14. Mannie, E. et.al., “Generalized Multiprotocol Label Switching Architecture(GMPLS),” IETF RFC 3945, 2004. 15. Youngtak Kim et.al., “GLASS (GMPLS Lightwave Agile Switching Simulator): A Scalable Discrete Event Network Simulator for GMPLS-based Optical Internet,” White-paper, NIST-Gaithersburg, USA, 2002. 16. A. Jarray and A. Karmouch, “Periodical auctioning for QoS aware virtual network embedding,” in Quality of Service (IWQoS), 2012, pp. 1-4. 17. A. Jarray et.al, “Column generation based-approach for VN aware Networked Edge Data-Centers”, in 2014 IEEE Globecom Workshops, Austin, TX, 2014, pp. 93-98. 18. S. Knight et.al., “The Internet Topology Zoo,” in IEEE Journal on Selected Areas in Communications, vol. 29, no. 9, October 2011, pp. 1765-1775. 19. D.Eppstein, “Finding the K Shortest Paths”, in SIAM J. Comput., Philadelphia, PA, USA, 1999, pp. 652-673.