Cross-layer RWA in Translucent Optical NetworksJuzi Zhao, Suresh Subramaniam

Department of Electrical and

Computer Engineering

The George Washington University, Washington, DC 20052

Email: juzizhao@gwu.edu, suresh@gwu.edu

Maıte Brandt-PearceCharles L. Brown Department of

Electrical and Computer Engineering

University of Virginia, Charlottesville, Virginia 22904

Email: mb-p@virginia.edu

Abstract—Wavelength Division Multiplexing (WDM)-based op-tical networks are ideal candidates for core backbone networksbecause of their ability to carry large amounts of traffic. Opticalamplification has increased the reach of long-haul optical links.Nevertheless, it is impossible today to construct a truly opticalnetwork without converting optical signals to electrical signalsand regenerating them, because of the deleterious effects ofphysical impairments such as amplifier noise, dispersion, andnon-linear effects such as four-wave mixing and cross-phasemodulation. Being expensive devices, these regenerators areexpected to be sparsely located and used in such a networkcalled as a translucent optical network. In this paper, we considerthe routing and wavelength assignment (RWA) problem so thatthe Quality of Transmission (QoT) for connections is satisfied,and the network-level performance metric of blocking probabilityis minimized. Cross-layer heuristics that are based on dynamicprogramming to effectively allocate the sparse regenerators aredeveloped, and extensive simulation results are presented todemonstrate their effectiveness.

Index Terms—Translucent optical networks, cross-layer RWA,QoT-awareness, physical impairments, dynamic programming.

I. INTRODUCTION

Optical amplification has increased the reach of long-haul

optical links in Wavelength Division Multiplexing (WDM)-

based optical networks; yet, physical impairments make it

impossible to construct truly optical core networks. To combat

these impairments, optical signals are intermittently converted

to electrical signals and regenerated, thus ridding the optical

signals of accumulated impairments. Networks with sparse re-

generation capabilities are called translucent optical networks.In physically-impaired networks, the specified Quality of

Transmission (QoT) of connections must be satisfied. At the

same time, network operators are interested in utilizing the

network effectively, or in other words, minimizing the blocking

probability of connections. In this paper, several cross-layer

heuristics that are based on dynamic programming to effec-

tively allocate the Optical-Electronic-Optical (OEO) conver-

sion devices in sparse regenerators (the placements of regener-

ators are given) are developed, and extensive simulation results

are presented to demonstrate their effectiveness in reducing the

blocking probability while ensuring connections’ QoT. These

heuristic algorithms are QoT-aware (or impairment-aware, IA)

in the sense that the algorithms incorporate connections’ QoTs

in their selection process, as opposed to selecting routes and

wavelengths and then checking if the QoTs of connections are

met (QoT-guaranteed in [1]).

The authors of [2] investigate the RWA and regeneration al-

location problem by proposing an exhaustive search algorithm

with effective domination relationship for candidate paths, but

its worst-case time complexity is exponential in the number of

wavelengths and nodes. In [3], the authors propose dynamic

routing algorithms considering only static (i.e., network-state-

independent) physical impairments (Amplified Spontaneous

Emission (ASE), Polarization Mode Dispersion and worst-case

first-order crosstalk), and assume that at most one regenerator

can be used by each connection. The authors of [4] consider

the IA-RWA path selection and regenerator allocation problem,

but their algorithm is suitable for link-additive impairments

(also used in [5], [6] and [7]). Wavelength conversion is not

considered in that work, and it is assumed that the cost of a

path is reset to zero when a regenerator is used (which in this

context means that the BER becomes zero if a regenerator

is used). However, Q-factors of links in reality are non-

additive, and paths do have non-zero BERs even if regenerators

are used. IA-RWA algorithm and corresponding Generalized

Multi-Protocol Label Switching extensions are presented in

[8], considering the QoT of each transparent segment on

the path (also assumed in [9]). In our case, even though

each transparent segment has a low BER, the total path may

have unacceptable BER. An IA-RWA algorithm is proposed

in [10], considering both the path quality (state-independent

static impairments) and regenerator allocation. We propose and

implement an improvement of this algorithm in this paper, and

compare it with our dynamic programming method.

This paper addresses the fundamental problem of QoT-

aware routing and wavelength assignment in translucent op-

tical networks. The paper is organized as follows. In Section II,

we present the network model and the physical layer impair-

ments considered in our study. Section III describes the RWA

algorithms. Section IV presents and discusses the simulation

results. We conclude the paper and present possible extensions

to this work in Section V.

II. NETWORK MODEL AND DEFINITIONS

There are two kinds of nodes in the network: transparentnodes and regeneration (3R) nodes (with given locations).

Each 3R node has a given limited number of OEO conversion

devices, which can be used for signal regeneration and/or

wavelength conversion. The OEOs are assumed to be available

for any connection passing through the 3R node. An OEO

that is allocated to a connection cannot be used by another

connection until the first connection is torn down. The wave-

length conversion afforded by an OEO use can be from any

wavelength to any other. Each link consists of two fibers with

W wavelengths each, oriented in opposite directions. A 3R

node will be referred as an OEO node if it has at least one

IEEE ICC 2012 - Optical Networks and Systems

978-1-4577-2053-6/12/$31.00 ©2012 IEEE 3079

Fig. 1. An end-to-end path.

OEO converter available for use. Further, the transparent path

between two OEO nodes is called an OEO segment. In thisdefinition, the intermediate nodes may have OEO converters,

but they are not allocated to the path in question. A consecutiveOEO segment is the segment connecting two consecutive OEOnodes.

A. Physical Impairment Model

We adopt the physical layer model in [1]. Fig. 1 shows an

end-to-end transmission path, where OXC denotes an Optical

Crossconnect, and AG represents an optical amplifier with

gain G. Three kinds of physical impairments are considered:Inter Symbol Inference (ISI), ASE noise, and non-linear

crosstalk, consisting of cross phase modulation (XPM) and

Four Wave Mixing (FWM). Because of the good port isolation

of Wavelength Selective Switch-based Reconfigurable Optical

Add/Drop Multiplexers ([11]), node crosstalk considered in

[1] is ignored in this paper. (However, our RWA algorithms

are independent of this assumption.) The Q factor is used as

a metric for the QoT of a transparent connection. For on-

off keying transmission, assumed in this paper, Q = μ1−μ0

σ0+σ1,

where μ1 and μ0 are the means of the samples if a ‘1’ is sentand if a ‘0’ is sent, respectively; σ1 and σ0 are the standarddeviations if a ‘1’ or a ‘0’ is sent, respectively. σ1 can becalculated by σ21 = σ2isi+σ2n+σ2nl, where σ

2isi, σ

2n and σ

2nl are

the variances due to ISI, ASE noise and non-linear crosstalk,

respectively and σ20 = σ2isi. μ1, μ0, σ0, σisi, and σn dependon path length; they can be pre-computed and stored in a

lookup table. σnl is a function of the wavelength usage andthe length of the path (i.e., it is state-dependent, and therefore

cannot be computed offline). Therefore, ASE noise and ISI are

called static impairments, and FWM and XMP are considered

dynamic impairments. The interested reader is referred to [1]

for further details.

B. Traffic Model

Bidirectional connection requests (of full wavelength ca-

pacity) between a pair of nodes (s, d) arrive to and departfrom the network randomly. All connections require a spec-

ified BER. An RWA algorithm returns a path and one or

more wavelengths and OEO allocations for a connection. A

connection can be blocked due to any of three reasons: (a)

the RWA algorithm cannot find a path with a wavelength on

an OEO segment of the path (path blocking), (b) the BERrequirement cannot be satisfied for the new connection by

the selected path and wavelength(s) (QoT1 blocking), or (c)the BER requirement of one or more existing connections

would be violated if the new connection is admitted using

the path and wavelength(s) selected by the RWA algorithm

(QoT2 blocking).

The BER for an OEO segment is computed by

BER =1

2erfc

(Q√2

). (1)

If a path has multiple OEO segments, the end-to-end BER is

calculated by assuming independent bit errors as:

BER = 1−∏i

(1−BERi), (2)

where BERi is the BER for OEO segment i.

III. RWA ALGORITHMS

In this section, we present our algorithms, as well as the

algorithms used for comparison and benchmarking.

The general flowchart for all algorithms is shown in Fig. 2.

Suppose a connection request between source s and destinationd with a specified BER requirement arrives. An ordered path

set P ′ consisting of (up to)K ′ paths (whereK ′ is a parameter)is pre-computed offline for each s-d pair. (For simplicity, weomit the subscript s-d from all variables.) These are the paths

from which every algorithm draws its candidate paths; they

may be computed in a variety of ways, but in this paper, it

is assumed that these are the K ′ shortest paths (based on

actual link lengths) computed using Yen’s algorithm (sorted

by increasing length). Note that there may not exist K ′ pathsbetween s and d, which is why we say “up toK ′” paths above.Given the set P ′, a candidate ordered set of paths P ⊆ P ′

consisting of (up to) K paths (K is a parameter specifying

the number of alternate paths that can be tried) is selected

by the algorithm according to its own criteria, which will be

detailed hereafter. For each path pi ∈ P , i = 1, 2, . . . ,K,the algorithm checks pi according to its own satisfactory

performance criteria (specified in Section III.B below). If pi isdeemed not satisfactory by the algorithm, the next path pi+1is checked, and so on, until all paths in P are tried. If a

satisfactory path is found by the algorithm, the path and the

associated wavelength(s) are passed to the QoT checker. TheQoT checker computes the end-to-end BER of the incoming

connection and the BERs of all of the existing connections

affected by this proposed path and wavelength(s). If all BERs

are adequate, the incoming connection is set up; otherwise it

is blocked. The algorithms differ in: (a) the computation of

the candidate path set P , and (b) the selection of the path andwavelength(s) to be passed to the QoT checker.

A. Selection of Candidate Path Set

We use the following methods for selecting the candidate

path set P in the various RWA algorithms.

• Plain: Here, the first K paths in P ′ are selected to beincluded in P ; simply called the K shortest paths.

• Seg: The static physical impairments in the network

place an upper limit on the length of a transparent

path, (or equivalently, the number of spans, if every

link has an integer number of spans), often called the

reach. Let the maximum number of spans allowed be

L. In this method, all paths with more than L spans

between any pair of consecutive 3R nodes on the path are

3080

Given: Path set P’ with (up to) K’ paths(computed offline)

Select candidate path set P (subset of P’)with (up to) K paths

P empty?

Accept connection

Is path “good”? All pathschecked?

QoT checker(not part of algorithms)

Pass?

Block connection

Y

N

N

N

N

Y

Y

Y

Select an untried path in P

Fig. 2. General flowchart for RWA algorithms.

discarded. The remaining paths are sorted by increasing

length, and the first K of these form P . (Note that thelimit on the number of spans and the path set P can

be computed offline, since only static impairments are

considered here.)

• Min: This method is used by the algorithm in [10]. Here,

the first path in P is the shortest path (i.e., first path

in P ′); the other paths are sorted by increasing value of(1+S)×D, where S is the number of shared links with

the first path in P , and D is the path length. Once again,

P can be computed offline.

• Online: Discard the paths in which any consecutive OEOsegment has no free wavelength and/or the segment has

more than L spans. Then sort the remaining paths by

increasing length. Note that this is an online method,

because it considers the availability of OEO converters

at 3R nodes and wavelengths of links (see definitions of

OEO node and consecutive OEO segment).

Each of the RWA algorithms could use any of the four path

set selection methods above, but we restrict the method used

to what makes sense for each algorithm.

B. Routing and Wavelength Assignment Algorithms

1) QoT-Guaranteed algorithm (QoT-G): This is a non-QoT-aware algorithm that completely ignores physical impairments

in selecting the path, and is used as a reference algorithm. The

candidate path set is formed using the Plain method. Givena candidate path pi ∈ P , OEO allocation and wavelength

assignment on pi are done as follows. Starting from node

s, the longest wavelength-continuous OEO segment that is

possible on pi is selected (nodes s and d are assumed to bedummy OEO nodes). If there is more than one wavelength

available on the segment, then the first available wavelength

is chosen (First-Fit WA). Suppose the segment terminates at

node

TABLE IPHYSICAL PARAMETERS

Description ValueBit rate 10 Gbps

Signal peak power 2 mWPulse shape NRZ

WDM grid spacing 50 GHzFiber loss 0.22 dB/km

Nonlinear coefficient 2.2 (W km)−1

Chromatic dispersion 17 ps/nm/kmDispersion compensation 100% post-DC

Noise factor 2Receiver electrical bandwidth 7 GHzNumber of wavelengths per fiber 16

converters, with nodem > i being the first converter node (i.e.,lowest index converter node). Then, we can write a recursive

function as follows (using (2)):

1−B′(i, j,m, k) = [1−B(i,m, 0)] · [1−B(m, j, k − 1)],where i+ 1 ≤ m ≤ j − k, k = 1, 2, . . . , j − i− 1, and

B(i, j, k) = minm

B′(i, j,m, k).

The base of the recursion is B(i, j, 0), 0 ≤ i < j ≤ a + 1,which is simply the BER on the transparent path from i toj, and is obtained using (1). When computing B(i, j, 0), weuse the wavelength that gives the best BER, and keep track

of which OEO converters are allocated to obtain B(i, j, k).The final k of B(s, d, k) for the connection is chosen as

the smallest one (i.e., minimum number of OEOs) which

makes all the connections’ (this new connection and ongoing

connections) BERs satisfactory.

IV. SIMULATION RESULTS

Connections are assumed to arrive to the network according

to a Poisson process. Each connection is assumed to have a

BER requirement of 10−9. For each data point in the graphs,we simulated several instances of 10000 to 100000 connection

arrivals, and obtained 95% confidence intervals. Table I shows

the physical parameters that were used in the simulation.

We present results for the European Network (EON) shown

in Fig. 3. In order to select the 3R nodes, the networks are

divided into two domains (solid line) or three domains (dashed

lines); the 3R nodes are selected as the border nodes (those

that are directly connected to nodes in other domains). For

example, the 3R nodes for 2 domains in the EON are 8, 13,

19, 9, 15, 20.1 The L in the Seg and Online path selectionmethods is set to 12, since for our assumed physical layer

parameters, the maximum number of allowed spans on a path

without OEO regeneration is 12 (if only the static impairments

of ASE noise and ISI are considered) for a BER requirement

of 10−9. We assume that K ′ = 40, i.e., up to 40 paths (notnecessarily disjoint) are computed offline for each node pair.

1We emphasize that algorithms for selecting the 3R nodes are outside thispaper’s scope. This paper’s focus is on developing RWA algorithms for agiven translucent network and comparing their performance. The domains areused only for the purpose of selecting the locations of the 3R nodes, andthe RWA algorithms are assumed to have full information about the entirenetwork, regardless of domains.

Figs. 4 and 5 show the blocking probability versus the

network Erlang load for the various algorithms and vari-

ous parameters. The loads were selected so that blocking

probabilities fall in the 10−4 to 10−1 range. As expected,QoT-G performs much worse than the QoT-aware algorithms.

Interestingly, the blocking probability stays essentially flat (in

fact, it decreases a tiny bit with increasing load). This can be

explained as follows. Recall that QoT-G selects the path and

wavelength(s) considering only wavelength availability and not

QoT. When the selected path is passed to the QoT checker,

there is a very high probability that the BER requirement is

not satisfied (either for this connection or for one or more of

the ongoing connections), and hence the connection is blocked.

When the load increases, a wavelength-continuous path from sto d becomes harder to find, and therefore, OEO converters areallocated to the connection for doing wavelength conversion.

Since these OEO nodes also perform 3R regeneration, the

probability that the selected path will be rejected by the QoT

checker does not increase.

In general, all versions of DP (except DP-Plain) outperformMINCOD-Q-REG, with DP-Online being the best. As can beseen, DP-Online outperforms MINCOD-Q-REG by up to 3

orders of magnitude, particularly at lower loads. This points

out that both candidate path selection and OEO allocation

are important. Another interesting observation is that DP-Minperforms much better than DP-Seg in the 3-domain case,

which is the opposite of what happens in the 2-domain cases.

What is happening here is that when there are more 3R nodes,

a candidate path is more likely to be selected by Seg becauseit is easier to find a path where each segment between two

consecutive 3R nodes is less than L spans. Many of these pathsend up not having an available wavelength, thereby causing

large path blocking. This does not happen in the 2-domain

case because many of these paths are not likely to be selected

as candidates in the first place because a segment between two

consecutive 3R nodes on the path has length greater than Lspans (due to fewer 3R nodes).

The blocking probability is plotted as a function of the num-

ber of OEO converters per 3R node in Fig. 6 for MINCOD-Q-

REG, DP-Seg, and DP-Online (as these are the best DP algo-rithms for the 2-domain case). For DP-Seg and DP-Online,there is a significant decrease in blocking as the number of

converters increases, and then the law of diminishing returns

kicks in and the curves tend to flatten out. The performance

improvement for DP-Online is especially sharp, pointing toits excellent ability to wisely allocate OEO converters.

We next show in Fig. 7 how the number of alternate paths

K affects the performance. It is remarkable that while DP-

Seg and MINCOD-Q-REG show improved performance as Kincreases until the performance plateaus, DP-Online is ableto achieve almost the same performance even with K = 1.The performance of DP-Seg and DP-Online is eventually thesame; it is a question of complexity. DP-Seg needs to checkabout K = 7 alternate paths to achieve the same performanceas DP-Online with K = 1.Finally, we show the importance of the dynamic (i.e., state-

3082

1

2

34

56

7

89 10

1112 13

1415

16

17 18

19

20

2122

23

24 25

2627

28

2

5

3

3

2

2

1

3 2

33

3

6

4

2

3

4

21

3

2

3

2 3

4

4

4

4

4

3

2

1

2

2

3

24

3

2

6

7

Fig. 3. 28-node EON. The number on eachlink corresponds to the number of spans, 70 kmeach.

0 20 40 60 8010−6

10−5

10−4

10−3

10−2

10−1

100

Load

Blo

ckin

g pr

obab

ility

QoTGDP−PlainMINCOD−Q−REGDP−MinDP−SegDP−Online

Fig. 4. Blocking vs. load in the EON for 2domains, 10 OEO converters per 3R node,K =2.

0 20 40 60 80 10010−6

10−5

10−4

10−3

10−2

10−1

100

Load

Blo

ckin

g pr

obab

ility

QoTGDP−PlainMINCOD−Q−REGDP−MinDP−SegDP−Online

Fig. 5. Blocking vs. load in the EON for 3domains, 10 OEO converters per 3R node,K =2.

5 10 15 20 2510−5

10−4

10−3

10−2

10−1

Number of OEO converters per 3R node

Blo

ckin

g pr

obab

ility

MINCOD−Q−REGDP−SegDP−Online

Fig. 6. Blocking vs. number of OEO convertersper 3R node for the EON; 2 domains, load =30 Erlangs, K = 2.

0 10 20 30 40 5010−4

10−3

10−2

10−1

Number of alternate paths K

Blo

ckin

g pr

obab

ility

MINCOD−Q−REGDP−SegDP−Online

Fig. 7. Blocking vs. number of alternate pathsK in the EON for 2 domains, load = 30, 10OEO converters per 3R node.

0 20 40 60 8010−6

10−5

10−4

10−3

10−2

10−1

Load

Blo

ckin

g pr

obab

ility

DPnoNL−Online 10OEO/3RDP−Online 10OEO/3RDPnoNL−Online 20OEO/3RDP−Online 20OEO/3R

Fig. 8. DPnoNL-Online vs. DP-Online in theEON for 2 domains, K = 2.

dependent) impairments, FWM and XPM, in the algorithms’

consideration. For this purpose, we implement a version of the

DP-Online algorithm (called DPnoNL-Online) that ignoresthe dynamic impairments in the computation of the Q-factor,

which is in turn used in the RWA process. Note, however,

that the QoT checker always computes the BER considering

all impairments, before admitting or rejecting a path. Fig. 8

shows the performance of DP-Online and DPnoNL-Online inthe EON with 2 domains, for two different values of number

of OEO converters per 3R node. The significant difference in

performance between DP-Online and DPnoNL-Online at lowloads points to the importance of considering dynamic non-

linear impairments in the RWA process.

V. CONCLUSIONS

In this paper, we investigated the QoT-aware routing and

wavelength assignment problem in translucent optical net-

works. We proposed an effective polynomial-time heuris-

tic based on dynamic programming, and simulation results

showed that this algorithm can allocate OEO converters to

connections judiciously, and significantly improve the block-

ing performance over current methods in the literature. Future

work may include the grooming of sub-wavelength connec-

tions, mixed line rates, and more sophisticated error coding

and modulation schemes.

ACKNOWLEDGMENT

This work was supported in part by NSF grants CNS-

0915795 and CNS-0916890.

REFERENCES

[1] Y. Pointurier, M. Brandt-Pearce, S. Subramaniam, and B. Xu, “Cross-layer adaptive routing and wavelength assignment in all-optical net-works,” IEEE J. Select. Areas Commun., vol. 26, no. 6, pp. 32– 44,Aug. 2008.

[2] K. Manousakis, P. Kokkinos, K. Christodoulopoulos, and E.Varvarigos,“Joint online routing, wavelength assignment and regenerator allocationin translucent optical networks,” IEEE/OSA J. Lightwave Technol.,vol. 28, no. 8, pp. 1152– 1163, Apr. 2010.

[3] X. Yang and B. Ramamurthy, “Dynamic routing in translucent WDMoptical networks: the intradomain case,” IEEE/OSA J. Lightwave Tech-nol., vol. 23, no. 3, pp. 955–971, Mar. 2005.

[4] F. Kuipers, A. Beshir, A. Orda, and P. V. Mieghem, “Impairment-aware path selection and regenerator placement in translucent opticalnetworks,” in Proceedings of the 18th IEEE International Conferenceon Network Protocols (ICNP), Kyoto, Japan, Oct. 2010, pp. 5–8.

[5] S. Rai, S. Ching-Fong, and B. Mukherjee, “On provisioning in all-optical networks: An impairment-aware approach,” IEEE/ACM Trans.Networking, vol. 17, no. 6, pp. 1989 – 2001, Dec. 2009.

[6] S. Pachnicke, N. Luck, and P. M. Krummrich, “Online physical-layerimpairment-aware routing with quality of transmission constraints intranslucent optical networks,” in IEEE International Conference onTransparent Optical Networks (ICTON), June/July 2009, pp. 1–4.

[7] V. Chava, E. Salvadori, A. Zanardi, S. Dalsass, G. Galimberti, A. Tanzi,G. Martinelli, and O. Gerstel, “Impairment and regenerator awarelightpath setup using distributed reachability graphs,” in Proceedingsof the IEEE Conference on Computer Communications (INFOCOM),2011, pp. 162 – 167.

[8] N. Sambo, N. Andriolli, A. Giorgetti, L. Valcarenghi, F. Cugini, andP. Castoldi, “Accounting for shared regenerators in GMPLS-controlledtranslucent optical networks,” IEEE/OSA J. Lightwave Technol., vol. 27,no. 19, pp. 4338 – 4347, Oct. 2009.

[9] R. Martınez, R. Casellas, R. Munoz, and T. Tsuritani, “Experimen-tal translucent-oriented routing for dynamic lightpath provisioning inGMPLS-enabled wavelength switched optical networks,” IEEE/OSA J.Lightwave Technol., vol. 28, no. 8, pp. 1241 – 1255, 2010.

[10] E. M. Tordera, R. Martınez, R. Munoz, R. Casellas, and J. Sole-Pareta,“Improving IA-RWA algorithms in translucent networks by regeneratorallocation,” in IEEE International Conference on Transparent OpticalNetworks (ICTON), June/July 2009, pp. 1–4.

[11] S. Tibuleac and M. Filer, “Transmission impairments in DWDM net-works with reconfigurable optical add-drop multiplexers,” IEEE/OSA J.Lightwave Technol., vol. 28, no. 4, pp. 557–568, Feb. 2010.

3083