Adaptive FEC-based lightpath routing and wavelength assignment in WDM optical networks

Author's Accepted Manuscript

Adaptive FEC-based lightpath routing and wavelength assignment in WDM optical networks Yongcheng Li, Hua Dai, Gangxiang Shen, Sanjay K. Bose

www.elsevier.com/locate/osn

PII: DOI: Reference:

S1573-4277(14)00065-4 http://dx.doi.org/10.1016/j.osn.2014.05.021 OSN317

To appear in:

Optical Switching and Networking

Received date: 6 March 2014 Revised date: 30 April 2014 Accepted date: 13 May 2014 Cite this article as: Yongcheng Li, Hua Dai, Gangxiang Shen, Sanjay K. Bose, Adaptive FEC-based lightpath routing and wavelength assignment in WDM optical networks, Optical Switching and Networking, http://dx.doi.org/10.1016/j. osn.2014.05.021 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

1

Adaptive FEC-Based Lightpath Routing and Wavelength Assignment in WDM Optical Networks Yongcheng Li, Hua Dai, Gangxiang Shen, and Sanjay K. Bose

Abstract—Forward error correction (FEC) has been widely used in optical communication systems to compensate for the degradation of the received optical signal to noise ratio (OSNR). Current optical networks tend to use the same type of FEC for all the lightpaths even though lightpaths with higher OSNRs can be established by FECs with lower overhead. This paper proposes an adaptive approach to choose the most efficient FECs for different lightpaths based on their individual OSNRs. An Integer Linear Programming (ILP) model and a simple waveplane-based heuristic algorithm considering shuffled lightpath demand sequences are developed to tackle the routing and wavelength assignment (RWA) problem. The simulation results indicate that compared to the non-adaptive case, using the proposed adaptive FEC selection scheme can significantly reduce the required FEC overhead. Apart from being far more tractable, the proposed heuristic approach performs almost as well as the ILP model. Index Terms—Adaptive forward error correction (FEC), optical signal to noise ratio (OSNR), shuffled demand sequences, waveplane-based heuristic algorithm

I. INTRODUCTION

W

ith the increase of traffic demand in backbone networks over the past few years, optical transmission systems are moving towards higher data rates over longer distances. However, in these systems, the physical-layer impairments such as fiber dispersion, nonlinear effects, channel noise, and other factors can significantly degrade the OSNRs of optical signals and thus impose a serious restriction on the transmission data rates and distances. The forward error correction (FEC) coding technique is considered as an effective way to compensate for the OSNR degradation. This technique has advantages of low hardware investment cost and high error-correction performance, and therefore has been widely used in fiber-optics transmission systems. The forward error correction techniques can be classified into three generations [2]. The first generation of FEC was based on the hard decision decoding technique that uses block codes, of which a typical example is the RS (255, 239) code with a Net Coding Gain (NCG) (@BER= 10 −13 ) of 5.8 dB. The second generation of FEC mainly focused on the concatenated codes, e.g., RS(239, 223)+RS(255, 239) and RS+BCH with an NCG (@10 ) of 7~9 dB. The third generation of FEC is based on more powerful soft-decision codes such as low-density parity-check code (LDPC) and Block Turbo codes. They have typical NCGs (@10 ) greater than 10 dB. In addition, the FEC coding techniques can be divided into the categories of in-band and out-band coding. The in-band coding technique uses the overhead bytes of a frame to store the FEC redundant bits. The SONET/SDH frame is a typical example of this type of in-band coding [3]. However, the limited overhead bytes in a frame (e.g., a SONET/SDH frame) restrict the highest coding performance of FEC. In contrast, as recommended in ITU-T G.975 [4] and G.709 [5], the out-band coding technique allows us to increase the line data rate using an extra overhead for FEC encoding, which can therefore maximize the performance benefit of FEC coding by increasing the number of redundant bits. Much effort has been made for studying FEC with different error correcting capabilities [6-13]. The experiment performed by Grover in 1988 was recognized as one of the first experimental FEC implementation in optical systems. By using the Hamming code, the coding gain of this FEC was only 2.5 dB (@10 ) [6]. As the wavelength division multiplexing (WDM) technique matured, researchers focused on constructing more complex codes. The most representative ones are Turbo and LDPC codes [7-9] where the LDPC codes can even come very close to the Shannon limit (i.e., within 0.04dB) [9]. Along with these more powerful FEC codes, related techniques such as interleaving, iterative decoding and soft-decision decoding based on multiple thresholds are employed to further improve the error correction performance [10-11]. In [12-13], the authors also present the future development Part of the paper was presented in ACP 2013 [1]. Manuscript received Feb. xx, 2014; revised xx, 2014. This work was jointly supported by the National 863 Plans Project of China (2012AA011302), National Natural Science Foundation of China (NSFC) (61172057, 61322109), Research Fund for the Doctoral Program of Higher Education of China (20113201110005), and Natural Science Foundation of Jiangsu Province (BK2012179, BK20130003). Y. Li, H. Dai, and G. Shen are with School of Electronic and Information Engineering, Soochow University, Suzhou, Jiangsu Province, P. R. China, 215006 (correspondence e-mail: [email protected]). S. K. Bose is with Department of Electrical and Electronic Engineering, Indian Institute Technology, Guwahati, India.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

2

trends in FECs so as to cater to further increases in transmission impairments. While these advances take the FEC coding techniques closer to the Shannon coding limit, there is a clear tradeoff between improved FEC coding performance and the growing coding overhead (OH); the increasing coding overhead shows up as additional line rate for the optical channel. The Net Coding Gain (NCG) is a typical measure for the coding efficiency of any particular type of FEC. Long-haul optical transmission systems must employ FEC types with higher NCGs in order to ensure a satisfactory BER with a lower OSNR requirement. The traditional FEC selection approach often chooses the best FEC (i.e., with the highest NCG such as the third generation of FEC) to cater for the lightpaths with the poorest OSNR in the whole network. This approach unnecessarily increases the overall network FEC coding overhead and therefore reduces network transmission capacity efficiency for lightpaths with higher OSNRs. Moreover, it also increases the network hardware cost because lightpaths tend to be over-engineered with more expensive FECs. To tackle the disadvantage of the uniform FEC selection strategy, this paper proposes an adaptive FEC selection approach for the optical channels. This approach chooses the most efficient FEC based on the actual OSNR of a lightpath. In the current stage, it is technically mature to implement the adaptive FEC selection strategy for static lightpath routing and optical channel establishment. With the further development of networking techniques such as software-defined optical networks (SDONs) [14-16] and real-time optical OSNR estimation [17], it is also feasible to implement the adaptive FEC selection in an online fashion. The software-defined optical networks which have been considered as a promising solution for a more flexible optical network can support the adaptive FEC selection approach. In such a type of network, optical transponders are controllable in many dimensions so that the related parameters such as the modulation formats and the types of FECs used by optical channels can be adjusted on demand by the control plane. In addition, with the support of optical performance monitoring, it is possible to estimate the OSNR of each lightpath online, which also provides the feasibility to choose different FEC types based on the actual OSNR of an optical channel. With the FEC selection for each optical channel decided in this fashion, we investigate the Routing and Wavelength Assignment (RWA) issues which would be important for the optical networks. Though there have been many studies on the RWA problem [18-24], this paper, for the first time, incorporates the FEC selection issue in lightpath routing and wavelength assignment. We propose an ILP optimization model and an efficient waveplane-based heuristic algorithm for the RWA problem which takes an adaptive FEC strategy into account. We evaluate the performance through simulations and find that compared to the non-adaptive scheme, using the proposed adaptive FEC selection approach can significantly reduce the required FEC overhead of lightpaths. This paper has focused on the WDM network because it is the most dominant in today’s transport networks. The adaptive FEC selection strategy uses the most efficient FEC type for each WDM optical channel, which can simplify the FEC coding algorithm, shorten the coding delay, as well as minimize the FEC overhead. As another study case, we can also apply this strategy to the elastic optical network. The rest of the paper is organized as follows. In Section II, we introduce the adaptive FEC selection strategy. In Section III, we explain in detail how to evaluate the OSNR of a lightpath. To solve the FEC-based RWA problem, we develop an ILP model and a waveplane-based heuristic algorithm in Sections IV and V, respectively. In Section VI, we present and analyze the simulation results. Section VII concludes the paper. II. ADAPTIVE FEC SELECTION In this section, we explain the proposed adaptive FEC selection strategy. For a 100-Gbps optical channel, we consider three FEC types, each one from the three FEC generations, including RS(255, 239) [12], RS(255, 239)+BCH(1023, 963) [10], and LDPC(4161, 3431, 0.825) [11]. The detailed information and performance of these FEC types is shown in Table I. It is clear that an FEC type with better performance (i.e., a higher NCG and lower OSNR limit) generally has a higher FEC overhead, and vice versa. For example, the first generation of FEC, RS(255, 239), has the lowest FEC overhead (i.e., 6.69%), but its FEC NCG is only 5.8 dB and the required OSNR limit is the highest at 14.5 dB. In contrast, the third generation of FEC, LDPC (4161, 3431, 0.825), can achieve the best NCG (i.e., 11.27 dB) and has the lowest OSNR limit (i.e., 9.1 dB). It however has the highest 21.2% overhead. TABLE I. INFORMATION ON THREE FEC TYPES

Assuming that the above three FEC types are available for use, we consider the six-node network of Fig. 1 as an example to illustrate the key idea behind our adaptive FEC selection strategy. In this network, there are three lightpaths, (0-1), (0-5-4), and (0-1-2-3). The OSNR values of the lightpaths are 15 dB, 13 dB, and 10 dB, respectively. Under the uniform FEC selection strategy which is typically prevalent, FEC type 3 with the highest NCG and highest overhead would be selected for all the lightpaths in order to ensure that the lightpath (0-1-2-3) with the lowest OSNR3=10 dB can be established. Though this FEC type 3 is indeed needed for lightpath (0-1-2-3), it is clearly wasteful to use it for lightpaths (0-1) and (0-5-4) as these only need FEC type 1 and type

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

3

2, respectively, to support their desired BERs. In contrast, if the new adaptive FEC selection strategy is adopted, then the most efficient FEC type can be adaptively used to support each of the lightpaths according to their individual OSNRs. Following this strategy, we should use FEC type 1 for lightpath (0-1) and FEC type 2 for lightpath (0-5-4) to meet their individual BER requirements. Under the uniform selection strategy, using FEC type 3 for all the lightpaths would require a total overhead of 21.2%×3=63.6%; in contrast, with the adaptive strategy, we can reduce the total overhead to 41.23% (6.69%+13.34%+21.2%) resulting in a savings of 22.37%. This example illustrates the benefit of using the proposed adaptive FEC selection strategy in reducing the FEC overhead of lightpaths.

Fig. 1. Example of adaptive FEC selection.

III. OSNR EVALUATION The proposed adaptive FEC selection strategy chooses FEC types based on the actual OSNR value of each lightpath. The detailed OSNR evaluation process for a lightpath is explained here. An optical transmission system would have several physical-layer impairments such as fiber dispersion, nonlinear effect, channel noise, etc., which would affect the receiving OSNR of a lightpath. In this paper, we consider the amplified spontaneous emission (ASE) noise of the Erbium Doped Fiber Amplifier (EDFA) [25] as the main impairment of a lightpath.

( ⎣⎢l

1

l1 (km) / 80 ⎦⎥ + 1)

n1 = ⎢⎣l1 / 80 ⎥⎦ + 2

( ⎣⎢l

2

l2 (km) / 80 ⎦⎥ + 1)

l1km 1 PAse

2 PAse

n1 −1 PAse

n1 PAse

1 PAse

2 PAse

n2 = ⎢⎣l2 / 80 ⎥⎦ + 2 l2 km n2 −1 PAse

n2 PAse

OSNRlink 1

OSNRlink 2 l1 G = 10dB G = (km) × 0.25(dB/ km) ( ⎢⎣l1 / 80⎥⎦ + 1)

l2 G = 15dB G = (km) × 0.25(dB/ km) ( ⎣⎢l2 / 80⎦⎥ + 1)

OSNRTotal Fig. 2. Calculation of OSNR for a lightpath

The other non-ASE-noise impairments such as chromatic dispersion, amplification ripple, polarization mode dispersion (PMD) etc., are accounted for as an overall 2.5-dB OSNR penalty as in [26]. The simple example of Fig. 2 which contains three optical cross-connect (OXC) nodes and two links with distances and km, is used to illustrate the OSNR calculation of a lightpath. Note that a pre-amplifier and a post-amplifier are placed before and after each OXC node. Line-amplifiers are evenly deployed in the middle of each fiber link if the distance of the link is longer than 80 km. Thus, for a fiber link of length l km, we deploy /80 2 EDFAs, where the last number 2 corresponds to a pre-amplifier and a post-amplifier respectively at the two end nodes of the link. With this deployment, the distance between any two neighboring km. For example, if l = 150 km, then we need to deploy 3 EDFAs with the equal distance of each EDFAs would be /

amplification span to be 75 km. We assume the fiber attenuation coefficient to be 0.25dB/km. If the amplification span distance (i.e., the distance between two neighboring EDFAs) is 80 km, then the required amplification gain would be 20 dB. To support different amplification gains, we consider two types of amplifiers whose maximum gains are 15 dB and 22 dB, respectively. Fig. 3 shows the numerical curves of the noise figures (NFs) of the two types of amplifiers under different amplification gains. Because of its lower maximum gain, the 15-dB EDFA shows better NF performance than the 22-dB EDFA in the range in which their amplification gains overlap.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

4

30 15-dB EDFA 25

22-dB EDFA

20 15 10 5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

Amplification Gains

Fig. 3. Noise Figures (NF) of the two amplifier types under different amplification gains [26].

The amplifier type is chosen by using the following approach. If the required gain is lower than 15 dB, the 15-dB EDFA is used because of its lower NF; otherwise, the 22-dB EDFA is used. The actual gain of each selected amplifier can compensate for all the losses accumulated before the amplifier. For example, if the amplification span is 60 km, then the working gain of a line-amplifier connected to the span would be 60×0.25=15dB. The gain setting for the post-amplifiers is treated differently. If it is a post-amplifier on the first link of a lightpath, its working gain is set to be 10 dB; otherwise, it is set to be 15 dB as shown in Fig. 2. This is because a lightpath traversing an intermediate node has to pass through more optical component stages than when it is added at a source node. We employ the following formulas to calculate the OSNR of a lightpath. 58 ∑

∑

⁄ 1

(1) (2)

10

1/ 1

(3) (4)

∑

denotes the power of the ASE noise of an optical amplifier within 0.1-nm The definitions of the notations are as follows. spectrum range in units of dBm. denotes the output power of a lightpath. (We assume a constant 0-dBm channel output power, i.e., =1(mw)). G is the optical amplification gain where we use Newton’s interpolation to find the corresponding NF for each G based on the curves in Fig. 3. is the corresponding noise figure value when the amplification gain is G. S is the set of links traversed by the lightpath. is the OSNR of the sth link on the lightpath. The explanations of the equations are as follows. Equation (1) finds the ASE noise of an EDFA. Equation (2) finds the accumulated ASE noise on a fiber link which contains n EDFAs. Equation (3) calculates the OSNR of the fiber link. Equation (4) finds the OSNR of a lightpath that traverses multiple links, based on the individual link OSNRs. It should be noted that in our calculations, all the OSNRs have their linear values (not in dB). However, the final lightpath linear OSNR is converted to a dB scale. IV. ILP MODEL FOR ADAPTIVE FEC SELECTION In this section, we focus on the FEC-based RWA (routing and wavelength assignment) problem for the WDM optical network. According to different lightpath traffic demand types, the RWA problem can be classified into two categories, i.e., static and dynamic RWAs. In this study, we consider the static case which assumes that the lightpath traffic demand matrix is given a priori. In addition, we also assume that the wavelength resources (i.e., the maximal number of wavelengths) are limited on each fiber link and the constraint of wavelength continuity is followed when establishing lightpaths. Based on the above assumptions, we develop an ILP optimization model to tackle the FEC-based RWA problem. Due to the limited network resources on each fiber link, the objective of the model is to maximize the number of successfully served lightpath demands while minimizing the FEC overhead of the lightpaths which are successfully established. The sets, parameters, and variables of the ILP model are defined as follows. Sets: the set of network links; L the set of node pairs; R W

the set of wavelengths on the fiber links; the set of candidate routes between node pair r, which are used for lightpath establishment Parameters:

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < ,

5

a binary variable, which takes the value of 1 if link i is traversed by path p of node pair r; 0, otherwise; the number of lightpath demand units between node pair r;

,

the minimum FEC overhead required for candidate path p between node pair r. We use the OSNR estimation model in Section III to find the OSNR for the candidate path and then compare the found OSNR with the lowest OSNR limits supported by different FEC types to choose an FEC type that requires the lowest overhead and whose OSNR limit is smaller than the path OSNR. a weight factor.

Variables: , the number of lightpath demand units served by candidate path p of node pair r; ,

a binary variable, which takes the value of 1 if wavelength on candidate path p of node pair r is used to establish a lightpath; 0, otherwise. , , ∑ , Objective: Minimize ∑ , · , Subject to: , ∑ ,

∑

(5) ,

∑ ,

,

, ,

(6) ,

(7)

The first objective is to maximize the total number of served lightpath demand units and the second objective is to minimize the total overhead of all the successfully established lightpaths. The weight factor α is set to be a small value so as to ensure that successfully serving the traffic demand units has the highest priority. Constraint (5) ensures that the number of served demand units is smaller than the required demand units. Constraint (6) sums the total served demand units on all the wavelengths of the paths between the node pair. Constraint (7) ensures that any wavelength on a fiber link can only be assigned to a single lightpath. The computational complexity of the ILP model is decided by the dominant numbers of variables and constraints. In the above | |· model, the dominant number of variables is | | · | | · | | , and the dominant number of constraints is | | , | | · | | , where | | and | | are respectively the total numbers of node pairs and links in the network, | | is the number of wavelengths on each fiber link, and | | is the number of candidate routes between each node pair. V. HEURISTIC ALGORITHM FOR ADAPTIVE FEC SELECTION The ILP model will find an optimal solution to the FEC-based RWA problem, which can however be proven to be NP-complete. Its computational complexity would therefore be high for large networks. In order to reduce the computational requirements, it would be desirable to develop an efficient heuristic algorithm for the FEC-based RWA problem. We extend the traditional waveplane-based RWA algorithm to tackle the FEC-based RWA problem. In addition, to account for the performance uncertainty due to different sequences of lightpath establishment, we employ as in [23] a multi-iteration process to consider multiple shuffled demand sequences and choose the demand sequence with the best performance. Before describing the heuristic algorithm, we first explain the concepts of waveplane and shuffle demand sequence. A “waveplane” is defined as a virtual topology which is copied from the physical topology, in which each virtual link corresponds to a single wavelength on the corresponding fiber link and each node is split from a physical node [24]. All the virtual links in a waveplane have a common wavelength index which is the same as that of the waveplane. Thus, for a network with W wavelengths on each fiber link, we can split it into W waveplanes (or wavelength layered graphs). A “shuffled demand sequence” is referred to as a list of lightpath demands which is obtained by randomly shuffling the initial demand list. We call the proposed heuristic algorithm Adaptive FEC Selection-based Waveplane (AFS-W) algorithm, whose major steps are as follows.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

6

Adaptive FEC Selection-based Waveplane Algorithm: 1: Read network topology and place amplifiers on each link following the approach described in Section III; 2: Shuffle the initial lightpath demand list a few times and add all these shuffled demand lists into a list S; 3: Get a demand list D from list S; 4: For every lightpath demand unit in D do 5: For every used waveplane w do 6: Search the shortest route by the physical distance for the demand unit and calculate its OSNR on waveplane w; 7: If successful and OSNR >=9.1 dB then 8: Record the route and waveplane index w; 9: Find a route with the shortest hops from all the routes recorded in the previous for-loop; 10: If a shortest-hop route is found then 11: Establish a lightpath along the shortest-hop route and assign the corresponding wavelength to the lightpath; Based on the calculated OSNR of the lightpath, adaptively select an FEC type for the lightpath; 12: else if no routes can be found then 13: Search the shortest route (by the physical distance) on the first unused waveplane; Establish a lightpath along the shortest route and assign the corresponding wavelength to the lightpath; Based on the calculated OSNR of the lightpath, adaptively select an FEC type for the lightpath; 14: else 15: The request of lightpath demand unit is blocked; 16: Count all the served lightpath demand units and also calculate the total FEC overhead; Remove demand list D from list S; 17: If there is a next shuffled demand sequence in S then 18: Return to Step 3; 19: else 20: Select the shuffled demand sequence that has the largest number of satisfied lightpath demand units (with the first priority) and the lowest total overhead from all the evaluated demand sequences in S as the best result to output. In the above algorithm, we check if “OSNR >=9.1 dB.” This is because FEC type LDPC(4161,3431,0.825) that can provide the highest OSNR tolerance requires an OSNR at least 9.1 dB. Thus, it is impossible to establish a direct lightpath between a pair of nodes if the shortest route between the node pair has an OSNR even lower than 9.1 dB. VI. RESULTS AND DISCUSSIONS We evaluated the proposed FEC selection strategy by running simulations on two test networks. They were (1) the 10-node 22-link SmallNet network, and (2) the 24-node 43-link US backbone network (USNET) which have been shown in Fig. 4. The number close to each link is its physical distance in km. Note that, as case studies, the link distances are not actual ones, but are scaled up based on a certain ratio. To consider the situations of different numbers of wavelengths in each fiber link, we assume that the maximum numbers of wavelengths on each fiber link are 16 and 80 for the SmallNet and USNET networks, respectively. For our simulations, we also assume that we have the same number of demand units between each pair of nodes in the networks.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

7

For the ILP model, we search for the three shortest routes between each pair of nodes using the k-shortest path algorithm and store these routes in the candidate route set if their calculated OSNRs are at least higher than the OSNR limit (i.e., 9.1 dB) required by LDPC(4161, 3431, 0.825). We employed the AMPL/Gurobi [27] software package (version 5.0.0) to solve the ILP model on a 64-bit server with 2.4-GHz CPU and 8-G memory. The MIPGAP of the ILP model is set as 0.1%. For the heuristic algorithm, we shuffle the initial lightpath demand list 1000 times and select the demand sequence with the best design performance (i.e., the most served requests as the first criterion and the lowest coding overhead of the served demands as the next one) as our final result. In addition, to verify the efficiency of the proposed waveplane-based algorithm, we compare it with the approach based on fixed shortest path routing and first-fit wavelength assignment, under which we always choose the shortest route (distance-based) between a pair of nodes for lightpath establishment and assign the first available wavelength to the lightpath. 2230

0

1810 19 90

4

19 60

30 18

2

2000

3

24 10

1990 2180

1 2010

2400

1930

20 40

5

2120 80 18

1990

6

2070

7 1950

20 22

2400

2360 1990

8

9

(a) 10-node 22-link SmallNet network 1

19 1000

800

2600 1200 1900

950

2

11

6

1300

1300

15

20

1400 1000 1100 1000 3

1000

7 250

1000 9

600 1000

12

16

850 1000

4

1000

900

1200

1150

800

700 1000

900

300

1000 850

5

900 1200

8

21

800

950 10

13

1100

850

800

650 14

17

1200

22 600 1000

23

900 900

24

18

(b) 24-node 43-link US backbone network (USNET) Fig. 4. Test networks

A. Satisfied demand units Fig. 5 compares the satisfied demand units by different design approaches. Fig. 5(a) shows the results of SmallNet and Fig. 5(b) shows the results of USNET, in which the x-axis is the number of demand units between each node pair and the y-axis shows the number of satisfied demand units. Three design approaches are compared, including (1) “Shortest-path_adaptive_FEC,” which applies the fixed shortest path routing algorithm and the first-fit wavelength assignment strategy with adaptive FEC selection strategy; (2) “AFS-W,” which applies the AFS-W heuristic algorithm described in Section V; (3) “ILP_adaptive_FEC,” which is based on the ILP optimization model considering the adaptive FEC selection strategy. In addition, as an upper bound for the satisfied demands, we provide a curve that counts the total number of demands to be served (i.e., “Total” in legend). Based on the results in Fig. 5, we can see that the ILP model achieves the best result in terms of the number of satisfied demand units. The AFS-W algorithm is also very efficient as it is only slightly worse than the results of the ILP model. However, compared to the fixed shortest-path scheme, the AFS-W algorithm can achieve much better performance with much more satisfied demands. This is because the fixed shortest-path algorithm always uses the fixed shortest paths for lightpath establishment and would lose many opportunities of using longer routes for lightpath establishment when there are no free wavelengths available on the first shortest route. Finally, with an increasing number of demand units between each pair of nodes, compared to the total requested demands (i.e., the “Total” curve), there are some demand units that cannot be satisfied even with the ILP model. This is attributed to a limited wavelength number on each fiber link.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

8

(a) 10-node and 22-link SmallNet network

(b) 24-node 43-link US backbone network (USNET) Fig. 5. Satisfied demand units (16 wavelengths in each fiber link for SmallNet and 80 wavelengths in each fiber link for USNET

B. Distribution of FEC types We also show the distribution of each FEC type for all the satisfied lightpath demand units in Fig. 6. We compare the three design approaches as in Fig. 5. We find that the highest percentage (about 45%-60%) of lightpaths use FEC type RS (255, 239), while the percentage of the lightpaths using LDPC (4161, 3431, 0.825) is the lowest. This is related to the distribution of the physical distances of the lightpath demands. In general, there is a high percentage of node pairs that have short distances and only a small percentage of node pairs with long distances. Consequently, most of the lightpaths can be served with a lower-level FEC type RS (255, 239) and only few lightpaths need a higher-level FEC type LDPC (4161, 3431, 0.825). In addition, with an increasing number of demand units between each pair of nodes, there are more demand units coded with RS (255, 239). This is because the networks have limited number of wavelengths or resources on each fiber link and the design approaches tend to be biased towards shorter lightpaths in order to satisfy more lightpath demand units.

(a) 10-node and 22-link SmallNet network

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

9

(b) 24-node 43-link US backbone network (USNET) Fig. 6. Distribution of FEC types of served lightpath demand units

Average overhead

C. Average overhead per lightpath unit Another important performance comparison criterion is the average FEC overhead per optical channel. Fig. 7 shows this performance for the four design scenarios, including the earlier-mentioned three cases along with a non-adaptive FEC selection strategy with the shortest path routing and first-fit wavelength assignment (i.e., “Shortest-path_LDPC-FEC” in legend). In the non-adaptive FEC selection strategy, in order to support the lightpath with the worst OSNR, we choose the most advanced LDPC FEC type for all the lightpath demand units. We can see that the average overhead of the three adaptive approaches can reduce the encoding overhead by about 40%~55% compared to the fixed FEC selection strategy. In addition, all the three adaptive approaches have similar average overheads. Although the ILP model and the AFS-W heuristic algorithm perform slightly worse than the other two because they may use longer paths to satisfy more lightpath demand units, the AFS-W heuristic algorithm can serve much more demands units than the shortest-path adaptive strategy. We also observe that the average overheads of all the schemes tend to decrease with an increasing number of demand units per node pair. Once again, this happens because the optimization approaches are biased towards shorter paths (correspondingly, simpler FECs) given the limited link wavelength resources.

(a) 10-node and 22-link SmallNet network

(b) 24-node 43-link US backbone network (USNET) Fig. 7. Average overhead per served lightpath unit

D. Impact of demand shuffles For the AFS-W algorithm, we evaluate how the total number of shuffled demand sequences (or iterations) can affect the performance of satisfied lightpath demand units. We considered the USNET network with 1, 10, 100, and 1000 shuffled demand sequences. Table II shows the result of different numbers of shuffled demand sequences. Note that the column of “Improvement (%)” shows the performance improvement by “1,000 shuffles” over “1 shuffle.” We can observe that the order of the served demand units does influence the number of satisfied demands and moreover there seems to be a trend towards saturation between

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

10

the number of satisfied demand units and the number of shuffled demand sequences. In addition, with an increasing number of demand units per node pair, we see that the percentage of performance improvement between “1,000 shuffles” and “1 shuffle” becomes larger. This means that the shuffle of demand sequences can provide more benefit to the design cases with a larger number of demand units. This is reasonable since a larger number of demand units per node pair correspond to more combinations of possible shuffled demand sequences, which requires more shuffled demand sequence samples to obtain good performance. Table II. IMPACT OF NUMBER OF SHUFFLED DEMAND SEQUENCES

VII. CONCLUSION To reduce lightpath FEC overhead, we proposed an adaptive FEC selection strategy to choose the FEC type based on the actual OSNR of each lightpath. With reasonable assumptions on the optical amplifier deployment on each fiber link and for solving the RWA problem, we developed an ILP model to maximize total served demand units while minimizing the total required FEC overhead of served lightpaths. Considering the high computational complexity of the ILP model, we also developed a waveplane-based heuristic algorithm which performs close to the ILP model. The simulation studies show that the proposed adaptive FEC selection strategy is effective in reducing the FEC overhead costs by about 40%~55% compared to a non-adaptive scheme. Moreover, the ILP model and the proposed waveplane-based heuristic algorithm can satisfy more lightpath demands and the corresponding required FEC overheads are also much lower than the simple shortest-path scheme with a non-adaptive FEC type. In addition, the simulation results indicate that the design performance tends to improve by increasing the number of shuffled demand sequences though this also exhibits a trend towards saturation. VIII. ACKNOWLEDGMENT We would like to thank the reviewers for their valuable comments that helped improve the paper. We would also like to thank the guest editors of this special issue for their great effort on coordinating the review process. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19]

Y. Li, H. Dai, G. Shen, and S. K. Bose, “Adaptive lightpath FEC Selection in an Optical Network,” in Proc. ACP 2013. T. Mizuochi, “Recent progress in forward error correction and its interplay with transmission impairments,” IEEE Journal of Selected Topics in Quantum Electronics, vol. 12, no. 4, pp. 544-554, Jul./Aug. 2006. A. Tychopoulos, O. Koufopaulou, and I. Tomkos, “FEC in optical communications-A tutorial overview on the evolution of architectures and the future prospects of outband and inband FEC for optical communications,” IEEE Circuits and Devices Magazine, vol. 22, no. 6, pp. 79-86, Nov./Dec. 2006. ITU-T G.975, Forward error correction for submarine systems [s], 1996. ITU-T G.709, Interfaces for the optical transport network [J], 2001. W. D. Grover, “Forward error correction in dispersion limited lightwave system,” IEEE/OSA Journal of Lightwave Technology, vol. 6, no. 5, pp. 643-645, May. 1988. T. Mizuochi et al., “Forward error correction based on block turbo code with 3-bit soft decision for 10-Gb/s optical communication systems,” IEEE Journal of Selected Topics in Quantum Electronics, vol. 10, pp. 376-386, Mar./Apr. 2004. B. Vasic and I. B. Djordjevic, “Low-density parity check codes for long-haul optical communication systems,” IEEE Photonics Technology Letters, vol. 14, no. 8, pp. 1208–1210, Aug. 2002. S. Y. Chung, J. G. D. Forney, T. Richardson, and R. Urbanke, “On the design of low-density parity-check codes within 0.0045 dB of the Shannon limit,” IEEE Communication Letters, vol. 5, no. 2, pp. 58-60, Feb. 2001. J. Yuan, W. Ye, Z. Jiang et al., “A novel super-FEC code based on concatenated code for high-speed long-haul optical transmission systems,” Optics Communications, vol. 273, no. 2, pp. 421-427, May. 2007. I. B. Djordjevic and B. Vasic, “Iteratively decodable codes from orthogonal arrays for optical communication systems,” IEEE Communications Letters, vol. 9, no. 10, pp. 924–926, Oct. 2005. T. Mizuochi, “Next generation FEC for optical transmission systems,” in Proc. OFC/NFOEC 2008. F. Chang, K. Onohara, and T. Mizuochi, “Forward error correction for 100G transport networks,” IEEE Communication Magazine, vol. 48, no. 3, pp. S48-S55, Mar. 2010. N. McKeown, “Software-defined networking,” in Proc. INFOCOM 2009. O. Gerstel, M. Jinno, A. Lord, and S. J. B. Yoo, “Elastic optical networking: a new dawn for the optical layer?,” IEEE Communications Magazine, vol. 50, no. 2, pp. s12-s20, Feb. 2012. S. Gringeri, N. Bitar, and T. J. Xia, “Extending software defined network principles to include optical transport,” IEEE Communications Magazine, vol. 51, no. 3, pp. 32-40, Mar. 2013. J. H. Lee, N. Yoshikane, T. Tsuritani et al., “In-band OSNR monitoring technique based on link-by-link estimation for dynamic transparent optical networks,” IEEE/OSA Journal of Lightwave Technology, vol. 26, no. 10, pp. 1217-1225, May. 2008. H. Zang, J. P. Jue, and B. Mukherjee, “A review of routing and wavelength assignment approaches for wavelength-routed optical WDM networks,” Optical Networks Magazine, vol. 1, no. 1, pp. 47-60, Jan. 2000. D. Banerjee and Mukherjee B, “A practical approach for routing and wavelength assignment in large wavelength-routed optical networks,” IEEE Journal on Selected Areas in Communications, vol. 14, no. 5, pp. 903-908, Jun. 1996.

> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <

11

[20] S. Subramaniam and R. A. Barry, “Wavelength assignment in fixed routing WDM networks,” in Proc. ICC 1997. [21] Z. Zhang and A. S. Acampora, “A heuristic wavelength assignment algorithm for multihop WDM networks with wavelength routing and wavelength re-use,” IEEE/ACM Transactions on Networking, vol. 3, no. 3, pp. 281-288, Jun. 1995. [22] T. Lee, K. Lee, and S. Park, “Optimal routing and wavelength assignment in WDM ring networks,” IEEE Journal on Selected Areas in Communications, vol. 18, no. 10, pp. 2146-2154, Oct. 2000. [23] G. Shen, C. Wu, and J. Dong, “An almost-optimal approach for minimizing the number of required wavelengths for large-scale optical networks,” in Proc. OFC/NFOEC 2013. [24] G. Shen, S. K. Bose, T. H. Cheng, C. Lu, and T. Y. Chai, “Efficient heuristic algorithms for light-path routing and wavelength assignment in WDM networks under dynamically varying loads,” Computer Communications, vol. 24, no. 3, pp. 364-373, Feb. 2001 [25] G. Shen, W. V. Sorin, and R. S. Tucker, “Cross-layer design of ASE-noise-limited island-based translucent optical networks,” IEEE/OSA Journal of Lightwave Technology, vol. 27, no. 11, pp. 1434-1442, Jun. 2009. [26] G. Shen, Y. Shen, and H. P. Sardesai, “Impairment-aware lightpath routing and regenerator placement in optical networks with physical-layer heterogeneity,” IEEE/OSA Journal of Lightwave Technology, vol. 29, no. 18, pp. 2853–2860, Sep. 2011. [27] AMPL+Gurobi, Linear programming optimization software package [Online]. Available: http://www.gurobi.com/doc/30/ampl/.

Table I. Information on three FEC types FEC Type RS(255,239) RS(255,239)/ BCH(1023,963) LDPC(4161,3431,0. 825)

6.69%

Data Rate (Gbps) 106.69

13.34%

113.34

21.2%

121.2

Overhead

CG

Q Limit

5.8dB@

6.08dB

11.2dB

OSNR Limit 14.5dB

7.3dB@

7.92dB

9.0dB

12.6dB

11.27dB@

12.1dB

5.2dB

9.1dB

NCG

Table II. Impact of Number of shuffled demand Sequences Demand units per node pair 1 2 3 4 5

1 276 552 759 868 944

Number of shuffled demand sequences 10 100 276 276 552 552 763 764 871 878 960 960

1000 276 552 764 882 966

Improvement (%) 0 0 0.659 1.613 2.334