1. Introduction

It is widely believed that the current Internet has become ossified, having limited capability to support new networking technology innovations as well as emerging multimedia-based broadband services [1]. From the network service providers’ perspective, the future Internet is envisioned to transcend these limitations and evolve into a cost-effective network with a future-proof architecture that can quickly, flexibly, and dynamically adapt to new architectures, protocols, and services. Network virtualization has recently been proposed as a promising approach to realize this goal. In a virtualized network, multiple experiments or services can be run simultaneously over a shared substrate (a set of physical networking resources and facilities, e.g., routers, servers, optical links, and wireless devices) such that the overall system efficiency is improved by isolating different services into separate virtual containers and/or pseudointerfaces [1, 2] to form application-specific virtual networks in the Global Environment for Networks Innovation (GENI) [3]; virtual networks are also called “slices,” substrate-wide networks of computing and communication resources capable of running one or more experiments or a wide-area network service. An example of network slicing is shown Fig. 1, where each slice consists of a set of optical links, programmable routers, and servers. As illustrated in Fig. 1, slice 1 may be completely independent from slice 2 in terms of switching, routing, and packet format used, allowing researchers and/or service providers to perform many interesting experiments on future Internet designs in parallel, by creating multiple separate virtual networks over a single shared optical substrate. Creating such a high-speed, sliceable, and programmable substrate for multiple applications or virtual networks is thus both the main aim and the main challenge of network virtualization.

Compared with traditional virtualization technologies (e.g., operation system-level virtualization), network virtualization must also address networking interface heterogeneity, communication protocol complexity, and large-scale deployment scalability. Moreover, in order for the future Internet to support diverse future applications with various quality of service (QoS) requirements, network virtualization must provide highly adaptive connectivity, ranging from source-destination circuits of a given bandwidth to multinode connections with minimum delay and possibly different bandwidth granularities. Advanced optical networking techniques and devices have been recognized as fundamental in achieving these requirements [3, 4], as they are capable of supporting the new protocols from the link to the application layers. For example, in Fig. 1, a virtual node can be modeled as a packet router with a generic forwarding engine (GFE) associated with several virtual links, while a fluid queuing model with subwavelength connectivity (either in the time or frequency domains) can be assumed for the virtual link. Here the fluid queuing model is used to represent a service or experimental flow coming continuously over time at a random rate; it is actually not a precise description of traffic flow (e.g., packet-level description); it is just an envelope description of data flow (e.g., average data rate and burst size). For example, we usually use the fluid model to simulate TCP flows or label switched paths in IP/MPLS networks.

Programmable routers (or sliceable routers) are responsible for the emulation of switching-routing functionalities, whereas multidegree reconfigurable optical add–drop multiplexers with multiple transponders may be used for providing link virtualization at the wavelength level. We also note that such a virtualized network model (Fig. 1) is fundamentally different from existing models, such as generic grid mapping and the virtual private network (VPN), in three important aspects: (1) the virtualized nodes and links can flexibly allocate both optical and electrical resources; (2) the primary focus lies on L3/L2 networking applications; and (3) the forwarding rate, packet length, and link bandwidth are all relevant parameters for slice provisioning, unlike in grid or VPN techniques where only the computing resources (grid) and/or bandwidth requirements (VPN) are considered.

In order to efficiently provision a large number of such slices (Fig. 1) with diverse topologies, as well as dynamic bandwidth and packet forwarding rate requirements, the networking resources such as interfaces (the addressing space), forwarding engines, packet buffers, transponders, and control plane routing engines need to be isolated and partitioned in a programmable way. In this paper, we propose and study a novel programmable optical network architecture with a sliceable router structure designed to support the future Internet. The proposed approach is based on optical orthogonal frequency division multiple access (OFDMA) networking and relies on advanced digital signal processing for supporting virtual links with programmable bandwidth allocation. The rest of this paper is organized as follows. Section 2 gives a brief overview of relevant network virtualization technologies, including link and node virtualization, and generic packet formatting. Section 3 introduces an optical network substrate with a sliceable router structure based on optical OFDMA networking, along with its key benefits and transceiver design considerations. In Section 4, we conduct a comparative performance study of OFDMA versus time division multiple access (TDMA) in terms of their effectiveness in providing link virtualization. We also study an adaptive subcarrier allocation and assignment algorithm for various types of traffic flows in section 5. Section 6 concludes the paper.

2. Review of Network Virtualization

Network virtualization has become increasingly important for future Internet research in the networking community because of the capability to partition end-to-end network resources into parallel slices that allow multiple isolated services or experiments to run simultaneously. While the virtualization-slicing mechanisms for certain networking resources may be relatively simple, for others they are quite challenging. For example, while an interface buffer may be virtualized in a straightforward fashion by using hardware partitioning of queues and software threads/processes for processing of different protocols, virtualization of forwarding-routing tables (FIB/RIB) and link bandwidth sharing are significantly more complex, especially in real-time high-speed networking scenarios. Moreover, while several forms of virtualization, including virtual interfaces, virtual routing-forwarding), virtual output queuing (VoQ), and IP tunneling (that only provides connectivity without bandwidth guarantees), have been widely used in real virtualized networks, many more advanced issues remain open. These include QoS-guaranteed traffic flow separation, subwavelength (time-frequency domain) resource partitioning for higher experimental fidelity compared with IP tunneling, programmable framing, MAC-control-plane virtualization (signaling-routing), as well as protocol programmability to enable software-based node control. To illustrate the way the various software and hardware network virtualization methods for slice isolation interact at different layers (L1/L2/L3), planes (control or data), and granularities, we classify them into three major categories, as shown in Fig. 2:

Previous studies on network virtualization have relied mainly on conventional router-based overlay topologies for user level services [1]. Specifically, several methods for creating and mapping virtual nodes with virtual interfaces or virtual links have been proposed, with the majority falling into one of two categories: (1) container-based virtualization (used in Planetlab [1, 5]), which gives slivers access to a virtualized system call interface (VServer—one sliver is a virtual copy of the networking component’s resources (e.g., virtual machine or virtual link), or (2) paravirtualization, which gives slivers access to low-level hardware resources (Xen/OpenVZ). On the other hand, in the context of virtual private network design (VPN/VPLS), concepts including virtual forwarding interface (VRF/VFI), ATM VC/VP, Ethernet 802.1q trunking, provider backbone bridging, MPLS label switched paths, IP tunneling (e.g., GRE), and UDP tunneling have been extensively studied. For example, the early VINI prototype [6] used commodity computers to emulate programmable routers and relied on IP/UDP tunnels for link virtualization, whereas the Openflow architecture [7] exploited commercial hardware switch-router solutions with flow switching capability. The User Controlled LightPaths project (UCLPv2) [8] is another initiative to virtualize a physical optical WDM network into an arbitrary, user-specified topology.

The recently launched GENI initiative [3] focuses on the design and deployment of a shared, wide-area experimental facility to support a wide range of research into the future Internet. Fundamentally, GENI will consist of a global-scale wired network with programmable and virtualizable network elements (routers, switches, servers), as well as several wireless access networks, that will enable the simultaneous execution of multiple, independent experiments by providing each with an isolated slice created by end-to-end network resource virtualization. Particularly, in GENI, the optical networks will play the important role of providing programmability to any or all of the following: multidegree reconfigurable optical add–drop multiplexers for wavelength routing, virtual topology reconfiguration, optical multicasting, subwavelength burst and/or dynamic circuit switching, direct access to the optical spectrum, optical performance monitoring, adaptive modulation, optical control plane management, full service optical access, and programmable framing, as well as transport protocols and coding.

3. Optical OFDMA Networking Architecture

Virtualization of optical networks is more difficult than that of electrical networks due to the lack of fine-granularity (subwavelength) resource sharing mechanisms. In this section, we study a novel programmable mechanism based on optical OFDM/OFDMA that intrinsically enables flexible subwavelength bandwidth assignment, as well as other important benefits such as high spectral efficiency, adaptive transmission rates, and a simplified link scheduling model. The principle of OFDM, a digitized multicarrier modulation scheme, is to transmit a high-speed data stream by dividing it among a number of orthogonal subcarriers, each carrying a relatively low data rate, such that the per-subcarrier symbol duration is significantly longer compared with the temporal pulse spreading caused by optical dispersion. In this way, the transmission tolerance to both chromatic dispersion and polarization mode dispersion can be notably increased (e.g., the coherent optical OFDM presented in [9]). Optical OFDMA combines OFDM transmission with multiple access by assigning different OFDM subcarriers to different users [10] and in this way also lends itself to the creation of isolated, frequency-domain virtual links through advanced digital signal processing (FFT/IFFT). As shown in Fig. 3, each virtual link can be implemented in different ways such as lightpaths (WDMA), IP tunnels (TDMA), and subcarriers (OFDMA). Each subset of subcarriers is formed as one virtual link. Optical OFDM/OFDMA thus enables flexible, subwavelength provisioning of bandwidth in contrast to TDMA-based solutions [11, 12]. As an example, by using QPSK or 16-QAM modulation, 256 OFDM subcarriers can readily achieve an aggregate rate of $10\mathrm{Gbits}∕\mathrm{s}$ per wavelength in a 5 or $2.5\mathrm{GHz}$ bandwidth, respectively, and a bandwidth granularity of $40\mathrm{Mbits}∕\mathrm{s}$. We also note that this approach is capable of supporting different per-subcarrier bit rates by exploiting adaptive modulation–coding techniques on different sets of subcarriers and wavelengths. When compared with statistical or queuing-based approaches such as virtual Ethernet tunnels (802.1q), using the OFDMA-based method enables all nodes to transmit simultaneously on frequency-orthogonal subcarriers, which, in addition to the aforementioned flexible bandwidth assignment, can also provide full bandwidth utilization and perfect isolation between different subwavelength connections. This may be particularly valuable for mesh networking environments, as it could reduce the high hardware and software complexity and cost for subwavelength switching and grooming, as well as the time-synchronization-dependent protocol overhead associated with SONET or NG-SONET based approaches (such as VCAT and LCAS) [12]. Consequently, compared with legacy TDM/TDMA mechanisms, OFDM/OFDMA can lead to much more natural slicing of network resources in the GENI environment.

In order to support network virtualization for future Internet experimental research, resource programmability of the routing and forwarding engines and the interface buffer is needed. Link bandwidth virtualization in a virtualized router is likewise imperative. A proposed structure for a programmable router that helps realize these requirements is shown in Fig. 4, with a focus on its communication interfaces. The novel packet over optical OFDMA interface is designed to support multiple virtual interfaces through virtual links over a single physical link, where each virtual interface supports a Gigabit Ethernet (GE) or packet over SONET (POS) router interface. The second interface, composed of conventional router interfaces such as GE/POS, is mainly used for user access. The routing engine in the control plane supports multiple routing instances through virtual machines, while a GFE forwards generic packets from different slivers simultaneously. Specifically, on the ingress side, the GFE forwards incoming packets to the appropriate virtual interface based on a sliver ID or virtual interface ID in the packet header. We note that the incoming data rate is limited to the assigned bandwidth of the virtual interface. Moreover, each optical OFDMA link can carry multiple virtual links (e.g., M subsets of subcarriers), maintaining separate FIFO buffers for packet storage on each virtual link. On the egress side, the GFE maps each packet to the proper outgoing queue, which is configured to provide an appropriate encapsulation for different types of slivers.

To illustrate a complete optical OFDMA-based networking architecture, Fig. 5 shows a six-node virtualized optical substrate network based on the combination of the proposed programmable router shown in Fig. 4 and multidegree reconfigurable optical add–drop multiplexer technology. In the architecture of Fig. 5, it is assumed that a forwarding information database (FIB) stores entries for each virtual interface. We note that each OFDMA physical link supports parallel multiple bandwidth-programmable virtual links. Moreover, each virtual interface with a different IP address operates as an isolated physical transmission pipe. Bandwidth is guaranteed for each virtual link and can be changed in real time through signaling or routing protocols, such as the GMPLS extensions proposed in [4]. In addition, each node also provides wavelength-level switching and optical bypass functionality for virtual topology reconfiguration, enabling both strict isolation and dynamic reconfiguration of multigranularity virtual networks.

Figures 6(a), 6(b) present the packet over optical OFDMA interface transmitter and receiver structures, respectively. At the transmitter end [Fig. 6(a)] various types of packet coming from different slivers are first stored in separate FIFO buffers corresponding to different virtual interfaces. The virtual link controller is responsible for virtual link diagnosis (e.g., up–down detection) and packet encapsulation (as needed for different slivers). A line coding module (FEC encoder/decoder) may also be used to improve the bit error rate (BER) performance for each virtual link. The packet over OFDMA interface is assumed to support multiple virtual interfaces (VIFs), wherein each VIF is isolated and mapped to a dedicated subset of OFDM subcarriers. The choice of modulation–coding format may be done adaptively and on a per-subcarrier basis. An N-point IFFT converts the data symbols from the frequency to the time domain, followed by cyclic prefix insertion that enables simple frequency domain equalization. The remaining modules in Fig. 6(a), including the parallel–serial (P/S), digital–analog (D/A), and electrical–optical (E/O) converters are all used for generation of the physical optical OFDM signal. The receiver structure in Fig. 6(b) has nearly the same functionality as the OFDM transmitter, performed in reverse order, with the exception of the additional module for subcarrier alignment and channel synchronization.

In summary, in this section we have introduced the packet over optical OFDMA interface as a novel high-speed router transmission technique, which, compared with conventional POS/GE router interfaces, exhibits the following distinct advantages: (1) higher spectral efficiency due to the intrinsic features of OFDM [9], (2) adaptive data rate transmission by adaptive modulation–coding that also enables cross-layer optimization, (3) virtual link capabilities with programmable bandwidth allocation, (4) a simplified link scheduling–bandwidth sharing model that obviates the need for complex virtual clock or weighted fair queuing algorithms [13]. In addition, the proposed OFDMA-based interface can easily be extended to support multiple forwarding and routing engines and in this way adapt itself to various packet types and control plane protocols.

4. Comparative Study of Bandwidth Sharing Models

In this section, we quantitatively evaluate the performance of OFDMA- versus TDMA-based virtualization methods, considering factors such as number of slices supported and QoS performance of slices. As discussed above, the proposed optical OFDMA-based programmable router structure offers flexible provisioning of bandwidth at a subwavelength granularity. Theoretically, it provides better traffic flow separation than conventional temporal (TDMA)-based programmable routers, because it uses a different bandwidth sharing and link scheduling model. As shown in Fig. 7, in a conventional programmable router, all traffic flows share the bandwidth in a time domain, usually resulting in the problem of traffic pattern distortion [13], especially under higher link load, mainly due to the nonideal packet-level scheduler and nonequal packet length. On the other hand, the proposed programmable router shares the bandwidth in a dedicated frequency-domain subcarrier-based fashion, which, when combined with adaptive virtual link data rate transmission, can better maintain the fidelity of the slivers in terms of jitter, delay, and loss.

For generality, we evaluate the effectiveness of the proposed virtualized interface by modeling the system as an $M∕G∕1$ queue where packets of experimental flow i are generated by a Poisson process with arrival rate λ and the service time is arbitrary with mean μ and variance ${\sigma }^2$. We denote the physical link speed in Fig. 7(a) by C, the average packet length by $\mathcal{l}$, $\mu =\mathcal{l}∕C$, and the variance of the service time ${\sigma }^2=\alpha {(\mathcal{l}∕C)}^2$, $\alpha ⩾0$. We also define τ as the maximum allowable packet delay in each programmable router node and in this way impose a QoS constraint for each traffic flow. Letting S be the number of OFDM subcarriers in the link of Fig. 7(b), and $U_{\mathrm{BPSK}}$, $U_{\mathrm{QPSK}}$, and $U_{16\mathrm{QAM}}$ denote the data rates of each subcarrier under different modulation formats, we obtain $C=S\times U$, where $C=S\times U_{\mathrm{BPSK}}$ in practice. In order to guarantee the bandwidth of a virtual link, in Fig. 7(a) we use the effective bandwidth concept for sliver resource allocation defined in [14], where the effective bandwidth of an experimental flow i subject to the delay constraint τ was computed as

To compare the bandwidth sharing models for the two types of programmable router structures in Fig. 7, we conducted simulations in OPNET. The network model consists of two programmable routers having very large output buffers and physically interconnected by a single wavelength. The values for other relevant parameters are as follows: $\alpha =1$, $C=1\mathrm{Gbit}∕\mathrm{s}$, $S=256$, $\lambda =\mathrm{5,000}$, $\mathcal{l}=256$ bytes; according to the above formulas, this yields $N_{\text{OFDMA}}=85$ and $N_{\mathrm{TDMA}}⩽100$ depending on τ. For the sake of generality, two flow models are carried out in each sliver—one Poisson and one Pareto. The simulation results collected in Figs. 7(c), 7(d) compare the achievable performance of the two types of link virtualization approaches in a programmable router in terms of their packet delay and jitter. Figure 7(c) plots the average packet delay versus the number of supported slivers in a router interface, indicating that the proposed OFDMA method has a predictable and fixed delay (as each flow is isolated by different subcarriers) that can be considered as fixed propagation delay for the slice users. Moreover, as shown by Fig. 7(c), when the number of slivers increases, the delay under the conventional TDMA-based method increases dramatically. Figure 7(d) shows the relationship between jitter and the number of supported slivers in a router interface, revealing that the proposed OFDMA-based method has a predictable delay variation bound that depends mainly on the traffic flow itself. Under the conventional TDMA-based approach, jitter depends largely on the interference from other flows caused by imperfect time-domain bandwidth sharing [Fig. 7(a)]. As a result, the delay and jitter advantages provided by OFDMA can increase the fidelity of future Internet experimental research compared with TDMA by avoiding the flow coupling problem even under high-load conditions.

5. Adaptive Subcarrier Allocation and Assignment

In the previous section, bandwidth resources (OFDM subcarriers) were allocated according to the concept of effective bandwidth, which is not particularly well suited to highly bursty traffic flows. Consequently, in order to further improve the throughput and delay performance of the proposed optical OFDMA network, more efficient (adaptive) subcarrier resource allocation is vital, as it directly affects the slice performance. In this section, we propose and study an adaptive algorithm for OFDM subcarrier allocation, which combines the advantage of statistical multiplexing in packet switched systems with bandwidth guarantees of circuit switched networks. To do so, the proposed algorithm considers both the traffic characteristics as well as buffer occupancy/variance when determining the dynamic subcarrier allocation, such that the available subcarriers are always allocated to the nodes that require them most.

Based on the node model in Fig. 7(b), the key idea behind dynamic OFDM subcarrier allocation is to maximize the utility of each subcarrier in every short time period (scheduling interval). The utility function of each subcarrier is defined according to (1) the measured packet arrival rate in each flow, (2) the adaptive modulation format (e.g., BPSK/QPSK/8-QAM/16-QAM, but note that we only use QPSK here) based on transmission quality [15], and (3) the queuing length (delay). In addition, we assume the packet flows arriving randomly at a given node are buffered in several output queues (VoQ) such that there is one queue per flow, corresponding to each slice, and we monitor the flow load on each virtual link. We also determine the length of each scheduling interval adaptively by monitoring the queuing length variance among the queues to track the changing rate of the traffic flows. Generally, the scheduling interval falls into the range of $1\mathrm{ms}\text{to}1\mathrm{s}$, depending mainly on the traffic type. Several parameters are defined as follows: M denotes the number of queues (flows) for an optical OFDMA link, N is the total number of subcarriers in the optical OFDMA link, and ${\lambda }_i\left(\stackrel{⃖}{t}\right)$ is the measured arrival data rate for queue i during the previous scheduling interval $\stackrel{⃖}{t}$. The service rate is ${\mu }_i\left(t\right)={\sum }_j(x_{i,j}\times d_{i,j})$, where $x_{i,j}=1$ if subcarrier j is assigned to queue i, and 0 otherwise; $d_{i,j}$ is the corresponding data rate of each subcarrier when the adaptive modulation format based on the transmission quality (e.g., optical signal-to-noise ratio) is used. We define the buffer usage at queue i as $e^{{\overline{b}}_i\left(\stackrel{⃖}{t}\right)∕B_i}$, where ${\overline{b}}_i\left(\stackrel{⃖}{t}\right)$ is the measured average queuing length during the last scheduling interval $\stackrel{⃖}{t}$, and $B_i$ is the buffer size at queue i. Letting ${\Delta }_i\left(t\right)$ be the set of subcarriers assigned to queue i in the current scheduling interval t, the operation of the proposed algorithm is outlined in Fig. 8.

The effectiveness of the proposed dynamic subcarrier allocation algorithm is shown in Fig. 9. We assume that every queue carries different flows [as shown in Fig. 7(b)], with Queue 1 carrying real-time traffic at a nearly constant rate of $200\mathrm{Mbits}∕\mathrm{s}$, Queue 2 carrying a variable bit rate traffic flow with an average rate of $300\mathrm{Mbits}∕\mathrm{s}$, and Queue 3 carrying a bursty flow with a minimal rate of $0\mathrm{Mbits}∕\mathrm{s}$ and a peak rate of $600\mathrm{Mbits}∕\mathrm{s}$. Figures 9(a), 9(b), 9(c) show the real data rate and the allocated bandwidth for Queues 1, 2, and 3, respectively. We see that the dynamic subcarrier allocation algorithm performs very well with different types of real data flows, including highly bursty traffic. In addition, we compare the performance benefits of an adaptive scheduling interval over the fixed scheduling interval setup, here set to $100\mathrm{ms}$. Namely, we define a parameter known as the noneffective allocation ratio, which measures the ratio of total cases in which the allocated bandwidth is lower than the real data rate, and plot the obtained results versus the total scheduling counts during a given time period of $120\mathrm{s}$. Figure 9(d) shows the noneffective allocation ratio versus number of measurements subject to the fixed scheduling interval, indicating that the noneffective allocation ratio for highly bursty traffic flows is quite high (40%). We may thus conclude that the fixed scheduling interval setup is not appropriate for highly dynamic traffic flows, because it is not able to effectively track the variance of traffic flow. In contrast, Fig. 9(e) shows that using the adaptive method for bandwidth allocation significantly lowers the noneffective allocation ratios compared with the results of Fig. 9(d) and moreover makes them nearly independent of traffic flow type. The proposed adaptive OFDMA subcarrier allocation algorithm is thus an effective solution for both constant-rate and bursty traffic.

6. Conclusions

In this paper, we have discussed the virtualization of network resources, including nodes and links, as well the use of optical OFDMA technologies for facilitating the future Internet. Specifically, to meet the demands of future Internet experimental research and the needs of emerging applications, a bandwidth-programmable router structure and a novel packet over an optical OFDMA interface have been proposed to support heterogeneous services with various bandwidth granularities and QoS requirements. The unique features of the novel optical OFDMA-based network architecture include high spectral efficiency to support future $100\mathrm{Gbits}∕\mathrm{s}$ per channel transmission [16], as well as high flexibility to allow simultaneous bandwidth sharing at a subwavelength level. By using multiple orthogonal OFDM subcarriers for parallel optical transmissions in conjunction with postdetection digital signal processing, we have shown that the flexibility of subwavelength level resource sharing can be brought into the optical domain, while maintaining the transparency and exploiting the tremendous bandwidth of optics. Moreover, a comparative study of OFDMA versus legacy TDMA-based methods, as well as the strong performance of a novel OFDM-based adaptive bandwidth allocation algorithm, has shown that the proposed programmable optical network architecture is a promising candidate for enabling the future Internet.

 

Fig. 1 Illustration of network virtualization supporting heterogeneous services in the future Internet.

Download Full Size | PDF

 

Fig. 2 Multiplane, multilayer, and multigranularity network virtualization.

Download Full Size | PDF

 

Fig. 3 Optical OFDMA-based link virtualization.

Download Full Size | PDF

 

Fig. 4 Proposed sliceable router structure based on the novel packet over optical OFDMA interface.

Download Full Size | PDF

 

Fig. 5 Example of an optical OFDMA network.

Download Full Size | PDF

 

Fig. 6 Optical OFDMA (a) transmitter and (b) receiver structures.

Download Full Size | PDF

 

Fig. 7 (a) TDMA-based versus (b) OFDMA-based bandwidth sharing and link scheduling models for various traffic flows and (c), (d) the corresponding comparison results.

Download Full Size | PDF

 

Fig. 8 Adaptive OFDM subcarrier allocation and assignment algorithm.

Download Full Size | PDF

 

Fig. 9 Performance comparisons of adaptive subcarrier allocation under various types of traffic flows.

Download Full Size | PDF

1. T. Anderson, L. Peterson, S. Shenker, and J. Turner, “Overcoming the Internet impasse through virtualization,” Computer, vol. 38, no. 4, pp. 34–41, 2005. [CrossRef]  

2. A. T. Campbell, H. G. de Meer, M. E. Kounavis, K. Miki, J. Vicente, and D. Villela, “A survey of programmable networks,” ACM SIGCOMM Comput. Commun. Rev., vol. 29, no. 2, pp. 7–23, 1999. [CrossRef]  

3. Global Environment for Network Innovation (GENI), www.geni.org.

4. C. Qiao, W. Wei, and X. Liu, “Extending generalized multi-protocol label switching (GMPLS) for polymorphous, agile and transparent optical network (PATON),” IEEE Commun. Mag., vol. 44, no. 12, pp. 104–114, 2006. [CrossRef]  

5. J. Turner, P. Crowley, J. Dehart, A. Freestone, B. Heller, F. Kuhms, S. Kumar, J. Lockwood, J. Lu, M. Wilson, C. Wiseman, and D. Zar, “Supercharging PlanetLab—a high performance, multi-application, overlay network platform,” ACM SIGCOMM Comput. Commun. Rev., vol. 37, no. 4, pp. 85–96, Oct. 2007. [CrossRef]  

6. A. Bavier, N. Feamster, M. Huang, L. Peterson, and J. Rexford, “In VINI veritas: realistic and controlled network experimentation,” in Proc. 2006 Conf. on Applications, Technologies, Architectures, and Protocols for Computer Communications, Pisa, Italy, 2006, pp. 3–14.

7. N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson, J. Rexford, S. Shenker, and J. Turner, “OpenFlow: enabling innovation in campus networks,” ACM Comput. Commun. Rev., vol. 38, no. 2, pp. 69–74, 2008. [CrossRef]  

8. B. Arnaud, “${\mathrm{CA}}^{*}\text{net}$ 4 research program update—UCLP roadmap for creating user controlled and architected networks using service oriented architecture,” 2006.

9. W. Shieh and C. Athaudage, “Coherent optical orthogonal frequency division multiplexing,” Electron. Lett., vol. 42, pp. 587–589, 2006. [CrossRef]  

10. W. Wei, L. Zong, and D. Qian, “Wavelength-based sub-carrier multiplexing and grooming for optical networks bandwidth virtualization,” in Nat. Fiber Optic Engineers Conf., OSA Technical Digest (CD), Washington, DC: Optical Society of America, 2008, paper PDP35.

11. S. Melle, R. Dodd, S. Grubb, C. Liou, V. Vusirikala, and D. Welch, “Bandwidth virtualization enables long-haul WDM transport of $40\mathrm{Gb}∕\mathrm{s}$ and $100\mathrm{Gb}∕\mathrm{s}$ services,” IEEE Commun. Mag., vol. 46, no. 2, pp. 22–29, 2008. [CrossRef]  

12. D. Cavendish, K. Murakami, S.-H. Yun, O. Matsuda, and M. Nishihara, “New transport services for next-generation SONET/SDH systems,” IEEE Commun. Mag., vol. 40, no. 5, pp. 80–87, 2002. [CrossRef]  

13. H. Zhang, “Service disciplines for guaranteed performance service in packet-switching networks,” Proc. IEEE, vol. 83, no. 10, pp. 1374–1396, 1995. [CrossRef]  

14. F. Kelly, “Effective bandwidths at multi-class queues,” Queuing Syst., vol. 9, pp. 5–16, 1991. [CrossRef]  

15. A. Bocoi, M. Schuster, F. Rambach, D. A. Schupke, C. A. Bunge, and B. Spinnler, “Cost comparison of networks using traditional 10 and $40\mathrm{Gb}∕\mathrm{s}$ transponders versus OFDM transponders,” in Optical Fiber Communication Conf. and Expo. and Nat. Fiber Optic Engineers Conf., OSA Technical Digest (CD), Washington, DC: Optical Society of America, 2008, paper OThB4.

16. I. B. Djordjevic and B. Vasic, “Orthogonal frequency division multiplexing for high-speed optical transmission,” Opt. Express, vol. 14, no. 9, pp. 3767–3775, 2006. [CrossRef]   [PubMed]  

Wei Wei (M’05, SM’09) received the Ph.D. degree in communication and information systems from Shanghai Jiao Tong University, Shanghai, China, in 2004. He is currently with NEC Laboratories America as a Researcher with focuses on cross-layer optimization issues in optical/wireless integrated networks, network virtualization technologies for the future Internet, and next-generation full-service optical access networks.

Chonggang Wang received his Ph.D. degree from Beijing University of Posts and Telecommunications (BUPT). He is currently with NEC Laboratories America. His research focuses on hybrid optical and wireless networks, wireless networking, and future Internet design. He serves as a symposium co-chair for IEEE Globecom 2010 Communications QoS, Reliability and Modeling Symposium (CQRM). He is on the editorial board of ACM/Springer Journal of Wireless Networks and is an associate technical editor of IEEE Communications Magazine. He is a senior member of the IEEE.

Jianjun Yu received the B.S. degree in optics from Xiangtan University, Hunan, China, in June 1990 and the M.E. and Ph.D. degrees in electrical engineering from the Beijing University of Posts and Telecommunications, Beijing, China, in April 1996 and January 1999, respectively. From June 1999 to January 2001, he worked at the Research Center COM, Technical University of Denmark, Lyngby, Denmark, as an Assistant Research Professor. From February 2001 to December 2002, he worked for Lucent Technologies and Agere Systems, NJ, USA, as a member of the technical staff. He joined the Georgia Institute of Technology, Atlanta, GA, in January 2003, where he served on the research faculty and as the Director of the Optical Network Laboratory. He is currently a Senior Member of the Technical Staff with NEC Laboratories America, Princeton, NJ. He is also an Adjunct Professor and Ph.D. supervisor at the Georgia Institute of Technology and the Beijing University of Posts and Telecommunications. His current research interests include $100\mathrm{Gbit}∕\mathrm{s}$ high-speed transmission systems, new modulation-format techniques, radio-over-fiber systems and networks, wavelength-division-multiplexing passive optical networks, and optical-label switching in optical networks. As the first author, he has more than 100 publications in prestigious journals and conferences. He is the holder of 3 U.S. patents with 20 others pending. Dr. Yu is a Senior Member of the IEEE Lasers and Electro-Optics Society (LEOS). He served as a Guest Editor for a special issue “Convergence of optical and wireless networks” for the IEEE/OSA Journal of Lightwave Technology and a special issue “Radio-over-fiber-optical networking” for OSA’s Journal of Optical Networking. He was a Technical Committee Member (TPC) of the IEEE LEOS 2005–2007 annual meeting and now is serving as a TPC of OFC 2009–2010. He is an Associate Editor for the Journal of Lightwave Technology and the Journal of Optical Communications and Networking.

Neda Cvijetic (S’06, M’09) received the B.S. (summa cum laude), M.S., and Ph.D. degrees in electrical engineering from the University of Virginia, Charlottesville, in 2004, 2005, and 2008, respectively. She is currently a Research Staff Member in the Broadband and Mobile Networking Department at NEC Laboratories, Princeton, NJ. Her research interests include advanced modulation/detection techniques for high-speed optical transmission, optical-wireless convergence, and throughput optimization in heterogeneous networks.