Spectrum Efficient Super-Channels in Dynamic Flexible Grid Networks ...

3 downloads 25 Views 153KB Size Report
Ciena Corporation, 920 Elkridge Landing Rd, Linthicum, MD 21090, USA. Email: {sthiagar, mfrankel, dboertje}@ciena.com. Abstract: We analyze the blocking ...
OSA/OFC/NFOEC 2011

OTuI6.pdf OTuI6.pdf

Spectrum Efficient Super-Channels in Dynamic Flexible Grid Networks – A Blocking Analysis Sashisekaran Thiagarajan, Michael Frankel and David Boertjes Ciena Corporation, 920 Elkridge Landing Rd, Linthicum, MD 21090, USA Email: {sthiagar, mfrankel, dboertje}@ciena.com

Abstract: We analyze the blocking performance of spectrum efficient super-channels in dynamic flexible grid networks. Results demonstrate that increased spectral efficiency and flexible superchannel assignment do translate into network efficiency gains. OCIS codes: (060.4250) Networks; (060.4251) Networks, assignment and routing algorithms

1. Introduction Spectral efficiency has become the target parameter of choice for improvement to support the increasing demand for bandwidth in backbone networks. Multi-level modulation formats coupled with polarization division multiplexing technologies have evolved from simple OOK using direct detection to multi-level formats such as QPSK, 8-QAM and 16-QAM using coherent detection methods. For the future, scaling to higher multi-level formats, while being spectrally efficient, has a significant impact on required system OSNR, for example, 256-QAM requires 8.8dB higher OSNR as compared to 16-QAM. System designers must offset OSNR penalties accordingly with stronger FEC, deploying Raman amplifiers or take advantage of newer fiber types such as pure silica core fibers (PSCF). Recently, multi-carrier super-channels or super-wavelengths have been proposed [1] to achieve greater than 100G channel rates. In this approach, sub-carrier signals, each modulated at moderate multi-level rates, are aggregated and routed through the network as a single entity. Coherent detection allows these modulated subcarriers to be packed close to each other thereby realizing higher spectral efficiency while maintaining system reach [2]. A vital outcome of the super-wavelength approach is that channels greater than 100Gbps in capacity with equivalent reach will not fit within the standard ITU 50GHz channel spectrum. This calls for the introduction of an “elastic DWDM grid” or “flexible grid” concept [3] wherein the optic fiber spectrum is considered as a flexible, continuous resource rather than the rigid ITU-standard grid (henceforth referred to as “fixed grid”). In such a flexible grid approach, channels are allocated spectrum based on their capacity and reach requirements rather than as fixed chunks of spectrum. For e.g., a 200Gbps, 1500km reach super-channel may require, say, 75GHz of spectrum while a 400Gbps super-channel of similar reach may require only 135GHz. Further by employing different modulation formats, the 200Gbps channel may be packed into 60GHz while limiting its reach to, say 750km. The flexible grid approach imposes new requirements on WDM equipment such as ROADMs, tunable lasers and filters as these systems are conventionally designed to add/drop or lock onto channels that are anchored to center frequencies of the ITU grid. Nevertheless, new technologies such as flexible spectrum ROADMs [4] and programmable channel bandwidth filters are being introduced to make the flexible grid approach a reality. In this paper, we examine whether spectral efficiency gains obtained by employing super-channels translate to network efficiency gains in a dynamic flexible grid network scenario. Such a network scenario is particularly relevant to emerging data center interconnect networks where high capacity sessions are established and torn down to support multi-Tb machine-to-machine services such as cloud computing, real-time multi-media, server virtualization and virtual machine migration. A key difference between the fixed grid and flexible grid architectures is that the wavelength assignment problem in the former is transformed to a spectrum assignment problem in the latter. Recent studies [5-7] have considered the static network design problem for flexible grid networks and have proposed static ILP-based routing and spectrum assignment solutions incorporating distance-adaptive and capacity constraints. In this paper, we consider the dynamic networking case and employ Monte-Carlo simulations to compare the blocking performance of flexible grid and fixed grid networks of equivalent capacity. 2. Network Architecture Rather than consider the fiber spectrum as a continuous resource pool, we consider a fine granularity grid spectrum, say based on a 12.5GHz grid, in which channel spectrum is flexibly allocated in integer multiples of spectrum granularity based on channel capacity and guard band requirements. Indeed, recently introduced flexible-spectrum WSSs [4] allow dynamic selection of channel bandwidth from 50 GHz to 200 GHz in 12.5 GHz granular increments. On this basis, we assume a dynamic network scenario in which transponders are pre-deployed and high bandwidth traffic sessions arrive and depart from the network in a random manner. For the (1) fixed-grid baseline architecture, we assume 80-wavelength 50GHz-based fixed grid systems employing conventional ROADMs and conventional ITU-grid tunable transponders of 100Gbps wavelength capacity. In contrast, we assume (2) a

OSA/OFC/NFOEC 2011

OTuI6.pdf OTuI6.pdf

12.5GHz-based flexible grid architecture employing flexible spectrum ROADMs and tunable transponders, that can tune to center frequencies in 12.5GHz increments. In this case, channel width is flexibly assigned based on capacity requirements (in 12.5GHz multiples or spectrum units) over an overall spectrum of 4000GHz or 320 spectrum units. 3. Modeling Assumptions For simplicity, we assume ROADMs at all nodes for both scenarios are directionless and contention-less. We assume traffic sessions arrive randomly for a source-destination pair and can request bandwidth in four amounts, namely 100Gbps, 200Gbps, 300Gbps and 400Gbps. Each session is routed along a pre-specified path (shortest path in this case). If the session request cannot be accommodated in its entirety on this path then it is assumed to be blocked and is rejected. (1) In the fixed grid approach a traffic session is established based on the bandwidth and associated wavelength requirement shown in Table 1. In this case, the same set of wavelengths should be free on each link of the pre-specified path. (2) For the flexible grid scenario, we consider two options: (2a) Flexible grid without inverse multiplexing, in which a traffic session is established as a single super-channel if the same spectrum band is available on each link on the path. (2b) In the second option, if spectrum is not available for a single superchannel, then we attempt to inverse multiplex the session into smaller super-channels of the same aggregate bandwidth. The amount of spectrum needed for each super-channel is shown below in Table 1 and is specified in spectrum units, where each spectrum unit equals 12.5GHz granularity. Table 1. Bandwidth Requirements (1 spectrum unit = 12.5GHz) 100G

200G

300G

400G

Fixed Grid (wavelengths)









Flexible Grid (spectrum units)

4

6

8

10

We assume traffic session requests arrive at network nodes according to a Poisson process and are equally likely to be destined to any of the remaining nodes. To ensure fairness, the arrival rate of each bandwidth class is adjusted such that each bandwidth class requests the same combined capacity over the simulation period. The duration of each session is assumed to be exponentially distributed with unit mean. When the session finishes, the wavelength or spectrum allocated is released on all the links along the path. We employ the well-known First-Fit (FF) packing algorithm for wavelength and spectrum assignment. Specifically, in FF spectrum assignment, the spectrum units are numbered sequentially and the algorithm chooses that part of free spectrum with the smallest sequential spectrum unit number that fits a super-channel. 3. Numerical Results We present simulation results performed on two topologies, an 8-node ring and the 14-node NSFNET (as illustrated in [8]), under fixed grid and flexible grid network scenarios. As previously assumed, we have four classes of traffic sessions based on bandwidth requirements. The arrival rates of each of these classes are adjusted such that each class requests the same combined capacity over a simulation period. Even so, high bandwidth traffic sessions are blocked more often (over an order of magnitude more) than low bandwidth sessions unless algorithms that incorporate fairness and connection admission control are employed. Under such conditions, it is worthwhile to derive parameters that can be used to analyze the network-wide blocking performance and the fairness of session setup. We define the parameter network blocking probability per unit 100G capacity, P, (from [9]) as

P = ∑∀x pˆ x 4 where pˆ x is the blocking probability per unit 100G capacity of traffic class of bandwidth x,

ˆ x , is estimated as pˆ x = 1 − j 1 − p x , where p x is the where x is 100G, 200G, 300G or 400G. The parameter, p blocking probability for traffic class of bandwidth x. If x equals 100G, 200G, 300G or 400G, then j equals 1, 2, 3 or ˆ 400G pˆ 100G . If the fairness ratio is greater than 1, then 4 respectively. We define fairness ratio (from [9]) F = p high bandwidth sessions are penalized more compared to low bandwidth ones and vice versa. Figures 1 and 2 illustrate the network blocking probability per unit 100G capacity vs. load (in Erlangs) for the 8node ring and 14-node NSFNET respectively. Each figure compares fixed grid scenario, the flexible grid scenario with and without inverse muxing. The graphs show that at low-loads, the flexible grid approaches have lower blocking and better performance. This is likely the result of improved spectrum efficiency of super-channels and better packing of bandwidth across the flexible DWDM grid. Further, while there is a risk of fragmentation of the

OSA/OFC/NFOEC 2011

OTuI6.pdf OTuI6.pdf

spectrum into bands smaller than the smallest traffic session (i.e. 100G), the current application of the First-Fit packing algorithm for spectrum assignment does not appear to significantly affect performance in this respect. 1

10

100

1

10

100

1.E+00 Blocking Probability per unit 100G capacity

Blocking Probability per unit 100G capacity

1.E+00 1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E-06 1.E-07

1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E-06 1.E-07 1.E-08

1.E-08

Load (Erlangs)

Load (Erlangs) FIXED GRID

FLEX GRID

FIXED GRID

FLEX GRID ( with INV. MUX.)

Figure 1. Blocking performance vs. load for 8-node bidirectional ring

FLEX GRID

FLEX GRID ( with INV. MUX.)

Figure 2. Blocking performance vs. load for 14-node NSFNET

1.E-08

1.E-07

1.E-06

1.E-05

1.E-04

1.E-03

1.E-02

1.E-01

100

10

10

1 1.E+00

Fairness Ratio

100

1.E-08

1.E-07

1.E-06

1.E-05

1.E-04

1.E-03

1.E-02

1.E-01

0.1

0.1

Blocking Probability per unit 100G capacity FIXED GRID

FLEX GRID

FLEX GRID ( with INV. MUX.)

Figure 3. Fairness ratio vs. Blocking performance for 8-node bidirectional ring

1 1.E+00

Fairness Ratio

Figures 3 and 4 illustrate the scatter plot of fairness ratio vs. network blocking for the 8-node ring and 14-node NSFNET respectively. Ideally, we want the curves to be flat i.e. the fairness ratio to be as close to 1 as possible. We can readily observe that as the blocking probability increases (due to increase in network load) the ratio gets closer to one, indicating that all the low and high bandwidth connections are getting blocked on an equal basis because of lack of network capacity. More importantly, we observe that the fixed grid approach provides better fairness compared to the flexible grid approach with no inverse muxing. This indicates that high bandwidth sessions are less likely to find large, contiguous spectrum bands in the flexible grid approach and hence experience an unfair amount of blocking compared to the low bandwidth sessions. However, when we apply inverse muxing to the flexible grid approach we find that it provides equal if not better performance compared to the fixed grid option.

Blocking Probability per unit 100G Capacity FIXED GRID

FLEX GRID

FLEX GRID ( with INV. MUX.)

Figure 4. Fairness ratio vs. Blocking performance for 14-node NSFNET

5. Conclusions We have demonstrated that flexible grid networks provide tangible benefits in terms of network efficiency compared to fixed grid networks. While moving to a flexible grid certainly provides additional degrees of freedom in terms of improved spectrum efficiency and flexible channel allocation, it is important to consider features such as inverse multiplexing as part of flexible grid architecture to obtain the added benefits of fairness. 6. References [1] K. Roberts et. al., “Performance of Dual-Polarization QPSK for Optical Transport Systems,” Lightwave Technology., Journal of, vol. 27, no. 16, 3546-3559, Aug 2009. [2] K. Roberts et. al., “100G and Beyond with Digital Coherent Signal Processing,” IEEE Comm. Mag., 62-69, July 2010. [3] S. Gringeri et. al., “Flexible Architectures for Optical Transport Nodes and Networks,” IEEE Comm. Mag., 40-50, July 2010. [4] “Finisar to Demonstrate Flexgrid(TM) WSS Technology at ECOC 2010”, press release. [5] M. Jinno et. al., “Spectrum-Efficient and Scalable Elastic Optical Path Network: Architecture, Benefits and Enabling Technologies,” p. 66-73, IEEE Comm. Mag., Nov. 2009. [6] T. Takagi et. al., “Algorithms for Maximizing Spectrum Efficiency in Elastic Optical Path Networks that Adopt Distance Adaptive Modulation,” ECOC 2010. [7] K. Christodoulopoulos et. al., “Spectrally/Bitrate Flexible Optical Network Planning,” ECOC 2010. [8] V. M. Vokkarane and J. Jue, “Prioritized Routing and Burst Segmentation for QoS in Optical Burst-Switched Networks,” WG6, OFC2002. [9] S. Thiagarajan and A. Somani, “Capacity fairness of WDM networks with grooming capabilities,” Proc. SPIE, Vol. 4233, 2000.