Impact of Wavelength Assignment Strategies on. Hybrid WDM Network Planning. Sawsan Al Zahr, Maurice Gagnaire, and Nicolas Puech. Telecom Paris - LTCI ...
Impact of Wavelength Assignment Strategies on Hybrid WDM Network Planning Sawsan Al Zahr, Maurice Gagnaire, and Nicolas Puech Telecom Paris - LTCI - UMR 5141 CNRS 46, rue Barrault F 75634 Paris Cedex 13 - France Email: {sawsan.alzahr|maurice.gagnaire|nicolas.puech}@enst.fr
Abstract— Most studies carried out in the field of RWA in WDM networks assume an ideal optical medium. However, throughout its route an optical signal undergoes many transmission impairments introduced by long-haul optical components. In all-optical WDM networks, where no electrical 3R regeneration is performed at intermediate nodes, transmission impairments are accumulating and may result in high BER values at the receiver’s side. Hybrid WDM networks use sparse regeneration to overcome transmission impairments. Such networks achieve performance measures close to those obtained by fully opaque networks at a much lesser cost. In previous studies, we addressed the problem of hybrid WDM network design by proposing a tool for routing, wavelength assignment and regenerator placement so that the physical constraints are taken into account. The proposed tool uses a standard First-Fit wavelength assignment strategy. In this paper, we propose two new wavelength assignment strategies: Min-BER-Fit and Best-BER-Fit. The Min-BER-Fit strategy chooses the best available wavelength w.r.t. the BER to set up a lightpath whereas the Best-BER-Fit strategy chooses the first available wavelength that guarantees the quality of transmission required to establish the lightpath. The studied strategies are compared to the First-Fit strategy in terms of the number of required regenerators.
I. I NTRODUCTION Wavelength division multiplexing (WDM) networks provide huge bandwidth to keep up with the explosive growth of the traffic demand. All-optical – or transparent – WDM networks are nowadays achievable thanks to the development of optical cross-connects. In such networks, data is transmitted from its source to its destination in optical form. Switching and routing operations are performed in the optical domain without undergoing any optical-to-electrical conversion. In the absence of any wavelength conversion, the wavelength continuity constraint is imposed: a connection, also called a lightpath, is supposed to use the same wavelength on all links along the chosen route. In addition, these networks must face the technical difficulties in overcoming the transmission impairments introduced by long-haul optical components. Most of studies reported in the literature assume perfect physical layer conditions and thus the transmission impairments can be neglected. However, the optical signal undergoes various transmission impairments such as attenuation, dispersion, nonlinear effects, etc. In fully opaque networks, the optiThis work is supported by the e-photon One+ Network of Excellence.
cal signal quality is always considered to be acceptable since electrical 3R regeneration1 (Re-amplifying, Re-shaping, and Re-timing) is provided at each node. Meanwhile, providing regeneration at each node is costly. In all-optical networks, as no electrical 3R regeneration is performed at intermediate nodes, the quality of transmission (QoT) of the optical signal may be unacceptable at the destination node. Since optical 3R regenerators are still very expensive, hybrid WDM networks become today a promising solution to meet the QoT requirements. Electrical regenerators are used at intermediates nodes only when it is necessary improve the signal budget. In this context, we propose a tool called QWP (Quality of transmission dependent WDM network Planning), to deal with the problem of routing, wavelength assignment, and regenerator placement considering physical layer constraints. QWP consists of two modules, namely BERPredictor (Bit Error Rate-Predictor) and LERP (Lightpath Establishment and Regenerator Placement). Given a network topology, BER-Predictor provides an estimation of the BER on any lightpath taking into account the simultaneous effect of four transmission impairments, namely chromatic dispersion (CD), polarization mode dispersion (PMD), nonlinear phase shift (φNL ), and amplified simultaneous emission (ASE). For a specific lightpath, the estimated value of the BER gives an indication of the intermediate nodes at which the signal has to be regenerated. A signal regeneration is required whenever the BER value exceeds a certain threshold. In practice, several regenerators may be needed along a lightpath. LERP aims both at optimizing the network resources utilization and at minimizing the number of regenerators required to establish lightpaths under acceptable QoT conditions. It first solves the routing and wavelength assignment (RWA) problem associated with the traffic demands and then checks for the signal quality and places regenerators when necessary. In this paper, we investigate two new wavelength assignment (WA) strategies: Min-BER-Fit and Best-BER-Fit. The Min-BERFit strategy chooses the best available wavelength w.r.t. the BER to set up a lightpath whilst the Best-BER-Fit strategy chooses the first available wavelength that guarantees the required QoT. 1 In the rest of the paper, regeneration will refer to electrical 3R regeneration.
The paper is organized as follows. In section II, we provide a brief review of related studies on hybrid WDM networks and RWA issues considering physical layer constraints. In section III, we describe the QWP tool and the different WA strategies. The studied WA strategies are compared through simulation results in section IV. Our conclusions are drawn in section V. II. R EVIEW OF R ELATED S TUDIES Recent research work in the field of optical network planning focuses on the problem of RWA considering physical layer constraints. Various algorithms taking into account the impact of transmission impairments have been proposed. A. Regenerator Placement Several studies, reported in the literature, support the idea of sparse regeneration in large-scale WDM networks. As it has been surveyed in [1], the 600 km reach distance usually used is far from the average connection range for Internet traffic. In average, for a large set of possible network topologies and traffic scenarios, advanced transmission systems enabling a reach distance of 3000 km may only satisfy up to 60% of all Internet connections. Nevertheless, current technologies have difficulties in extending the reach distance to more than 2000 km. In [2], the authors suggest to establish sub-connections between regeneration sites using “islands of transparency” so that the QoT requirements are met. In [3] and [4], the authors propose an architecture of the regeneration capable node and demonstrate its feasibility. Regenerator placement algorithms are carried out, in [5], at the network planning stage based on the prediction of the future traffic demands. Simulations results show the tradeoff between the blocking probability and the total number of used regenerators under the light and heavy traffic loads. In each experiment case, the total number of used regenerators is normalized to the number of regenerators in a fully opaque network (e.g. the fully opaque 53-node, 68link, 16-wavelength USA network needs 2176 regenerators(see [5])). It has been showed that when the number of used regenerators exceeds 20% of the number of regenerators in a fully opaque network, adding extra regenerators only provides a little additional improvement in the blocking probability. B. QoT Aware Wavelength Assignment In [6], the authors study the impact of crosstalk on blocking performance in all-optical WDM networks. They propose four crosstalk-aware WA algorithms as variants of the wellknown First-Fit, Random-Pick, Most-Used, and Least-Used WA strategies so that the crosstalk factor is taken into consideration. The proposed strategies aim at minimizing the crosstalk effect in the network: they choose the available wavelength that minimizes the crosstalk between the new and existing lightpaths in order to reduce the blocking probability. Simulation results show that, compared to their traditional counterparts, the proposed algorithms can significantly reduce blocking caused by poor QoT. Nonlinear effects have been also considered in [7]. The authors propose an algorithm, called B-OSNR (Best-OSNR), which aims at minimizing the effect
Network topology G(V,E)
Eppstein algorithm
Permanent traffic demands
K-shortest paths Available resources {W}
Predict the QoT between any sourcedestination nodes for any wavelength along a shortest path (BER-Predictor)
BER on each shortest path for any wavelength
RWA + Electrical regenerators placement (LERP)
Routed demands satisfying the QoT requirement (Max) Number (Min) and placement of the required regenerators
Fig. 1.
The QWP tool.
of transmission impairments when solving the RWA problem. B-OSNR chooses the wavelength that provides the maximum OSNR. Simulations results showed that, when transmission impairments come into play, the B-OSNR outperforms traditional algorithms (for instance First-Fit) in terms of blocking probability. In our studies [8][9][10], we propose methods and algorithms to tackle the RWA problem while meeting the QoT requirements for the established lightpaths. In our approach, we assume that it is possible to set up a regenerator for a demand at any intermediate node if necessary. We deal with static traffic (permanent demands) and aim at minimizing the number of rejected demands. Our algorithms take into account four parameters describing the transmission impairments in order to estimate the signal quality (see below). In this paper, we propose some improvements of these algorithms to also take the signal quality into account within the WA strategy. III. T HE QWP T OOL The QWP tool deals with the problem of routing, wavelength assignment, and regenerator placement. It consists of a BER prediction tool and a dimensioning tool called LERP (Figure 1). A. BER-Predictor Given a lightpath (route, λ), BER-Predictor provides an estimate of the optical signal quality at the destination node. This is done by computing the physical parameters we have chosen to estimate the Q -factor. The Q -factor is a quantitative description of the optical signal quality and is related to the BER according to the following equation: µ ¶ Q 1 (1) BER = er f c √ 2 2 where: Z +∞ 2 2 er f c(x) = √ e−t dt (2) π x
Various techniques have been proposed to model the signal degradations considering one or more physical impairments such as chromatic dispersion, polarization mode dispersion, amplified spontaneous emission, crosstalk and nonlinear effects [7][6][11]. One originality of our approach is to take into account interactions between four physical parameters considered as relevant to describe the signal quality, namely chromatic dispersion (CD), polarization mode dispersion (PMD), nonlinear phase shift (ΦNL ), and amplified spontaneous emission (ASE). In our study, the Q -factor estimation is computed as a function of these four parameters, i.e. Q = f (CD, PMD, OSNR, ΦNL ) where f has been derived both analytically, from equations describing the physical phenomena, and experimentally ([8] and [9]). Chromatic dispersion is a consequence of the dependence of the group velocity of light with respect to the optical frequency. It yields pulse broadening which accumulates along the lightpath and results in errors at the receiver’s side due to pulse overlapping. The deployed fibers experience mechanical and thermal stresses that result in an asymmetrical refractive index. This leads to two orthogonal states of polarization whose traveling speeds through the fiber differ. This effect, known as polarization mode dispersion, also has an impact on data pulse distortion. Optical amplification generates a noise, called amplified spontaneous emission, with a random polarization state and a random wavelength (inside the amplification band). This noise accumulates along the amplifier cascades. When detecting the signal, the bits are distorted because of the beatings between the signal and noise contained in the electrical bandwidth of the detector. The noise degradation is directly linked to the optical signal to noise ratio (OSNR). The refractive index of a fiber also depends on the power of the whole optical field. Hence the power variation observed on a pulse edge can slightly disrupt the propagation condition of this pulse but also of the other pulses carried by other channels at different wavelengths. The repetition, and so the accumulation, of this detrimental effect is related to the amplification distribution and the chromatic dispersion management along the lightpath. Nonlinear effects produce a nonlinear phase shift. B. LERP The LERP algorithm is twofold. First, it solves the RWA problem associated to the traffic demands. Second, it verifies the QoT requirement and places regenerators when necessary. The RWA problem is solved using a sequential algorithm based on a random search (RS) method [10]. The RS method randomly considers various orderings according to which the demands will be routed, and chooses that ordering which minimizes the number of rejected demands. The routes are chosen among the k-shortest paths computed according to Epstein’s algorithm [12] whilst the wavelengths are assigned according to the First-Fit strategy.
K-shortest paths
Wavelengths {W} Established Lightpaths
Network Topology RWA
QoTTest
Traffic Matrix {D}
No
Empty traffic matrix?
RWA
Rejected Demands
Yes Free Ressources
RWA
QoTTest
No
Output no. 1 Rejected Demands
Fig. 2.
Empty traffic matrix?
Yes
Output no. 2 Routed Demands (Lightpath, wavelegth,regenerators)
Synopsis of the LERP algorithm.
The RWA problem being solved, the LERP algorithm addresses regenerator placement via a second step called the QoT-Test. The QoT-Test places regenerators so that the QoT requirements are met on all established lightpaths. Lightpaths composed of at least two hops are tested one by one. Considering a lightpath connecting node s to node d, the test begins at the third node. If the Q -factor is less than a certain threshold, the signal has to be regenerated at the previous node of the lightpath. Otherwise, we proceed to the same test at the next node. For each lightpath, the test ends if the destination node is reached or if a regenerator is placed at some intermediate node i. In the latter case, we define a new sub-demand whose source is i and whose destination is d. The route of the first segment of the lightpath (s, i) is stored and the used wavelength is reserved. The relative complementary sub-demand is added to a new traffic matrix to be routed afterwards. Once all the lightpaths have been tested, a new traffic matrix has been defined: it contains the sub-demands relative to demands that undergo signal regeneration. An RWA solution is computed for this new traffic matrix just as in the first step, taking into account the already allocated resources for the lightpaths established in the first step. Thus, we obtain a new routing scheme. The QoT-Test is performed for these newly established lightpaths. This procedure is repeated until no more signal regeneration is necessary, i.e. the QoT-Test does not return any sub-demand. Processing new RWA steps after the regenerator placement steps makes it possible to find shorter paths for the routed demands. Thus, the LERP algorithm minimizes the number of required regenerators [10]. At this stage, we try to route the initially rejected demands. Indeed, once a regenerator is placed, the wavelength continuity constraint is relaxed. Thus, the sub-demand (i, d) can be routed using a different wavelength from the one used by the (s, i) segment. This may free several WDM channels, and some demands, initially rejected because of the lack of resources, may now be routed [10]. Figure 2 gives a synopsis of the LERP algorithm.
1 15
12 2
16 18
TABLE II Q- FACTOR VALUES FOR THE LIGHTPATHS COMPUTED IN E XAMPLE 1 Demand
13
Route
Wavelength (λnm )
P1 : 2 − 3 − 4
1561.41 1553.32 1545.32 1540.55
10
6 8
17
3
9
5
δ1 : 2 → 4
14
Q -factor
dB
11.42 14.95 15.77 17.49
11
4 7
Fig. 3.
The American NSF backbone network topology.
δ2 : 1 → 6
P2 : 1 − 2 − 3 − 6
TABLE I T HREE TRAFFIC DEMANDS AND THEIR ROUTES (E XAMPLE 1) Demand (δi )
Source (si )
Destination (di )
Route (Pi )
δ1 δ2 δ3
2 1 3
4 6 7
2−3−4 1−2−3−6 3−4−7
C. Wavelength Assignment Strategies The aim of this study is to propose and investigate two new QoT-aware WA strategies, namely Min-BER-Fit and Best-BERFit. In the following, we describe these strategies as well as the standard First-Fit which will serve as a basis for comparisons. 1) First-Fit: All the available wavelengths are indexed according to their position in the frequency spectrum. The wavelength with the lowest index is selected from the set of available wavelengths. 2) Min-BER-Fit: All the available wavelengths are sorted according to their Q -factor value. The wavelength with the highest value for their Q -factor is selected from the set of available wavelengths. 3) Best-BER-Fit: All the available wavelengths whose Q factor value is higher than the threshold are sorted in ascending order according to their Q -factor value. The first wavelength that guarantees the required QoT is chosen. This approach enables to keep wavelengths having higher Q -factor values for longer lightpaths. Example 1 A numerical example may clarify how these different strategies work. Let us consider the 18-node north American backbone network (Figure 3) and a set of traffic demands as described in Table I. We assume that four wavelengths are available on each fiber-link, namely λ1 = 1561.41 nm, λ2 = 1553.32 nm, λ3 = 1545.32 nm, λ4 = 1540.55 nm. For each route Pi , we compute the Q -factor corresponding to each wavelength as shown in Table II. The BER threshold is assumed to be 10−5 which corresponds to a Q -factor value of about 12.6 dB. Solutions provided by the three strategies are given in Figure 4. By using the First-Fit strategy, 3 regenerators are needed, two of them being placed at node 3 (Figure 4(a)). Min-BER-Fit uses 2 regenerators whereas Best-BER-Fit only uses one regenerator (Figures 4(b) and 4(c) respectively). This result is due to the
δ3 : 3 → 7
P3 : 3 − 4 − 7
1561.41
5.71
1553.32 1545.32 1540.55
11.49 12.19 13.19
1561.41 1553.32 1545.32
−0.24 6.92 9.39
1540.55
12.69
fact that Best-BER-Fit first uses λ2 , so it keeps λ4 for the longest route. IV. N UMERICAL R ESULTS Several numerical simulations have been carried out in order to assess the improvement provided by the proposed new WA strategies. These simulations have been achieved considering the 18-node north American backbone network (NSF) shown in Figure 3. The network is assumed to be deployed using standard single-mode fibers (SMF) covering the C-band with 100 GHz spacing (providing 40 wavelengths on each fiber-link). In order to recover the fiber losses, double-stage EDFA (ErbiumDoped Fiber Amplifier) amplifiers are deployed every 80 km. The amplifiers characteristics, namely the gain and the noise figure are assumed to be wavelength dependent. Chromatic dispersion is dealt with dispersion compensating fibers (DCF) which are deployed at the amplification sites. Further details about the transmission system’s assumptions are given in [9]. Simulation results have been obtained considering static (also called permanent) traffic matrices generated randomly according to a uniform distribution. In our simulation scenarios, we consider various traffic loads where matrices of 100 to 700 demands are used. For each traffic load, we deal with 10 different matrices. Hence, each result represented in this paper is the mean value of 10 experiments. We first compare the WA strategies in terms of the number of required regenerators. Figure 5 shows the mean values of the number of regenerators required to establish lightpaths for various traffic loads. Vertical lines refer to the confidence intervals, i.e. mean value ± standard deviation. Regenerators are placed considering a typical BER threshold value of 10−5 . We assume that the system uses a forward error code (FEC), therefore the system can achieve an end-to-end BER of about 10−20 whilst the effective BER is of about 10−5 . Figure 5 shows that for low traffic loads, Min-BER-Fit is the best strategy whereas Best-BER-Fit is the best one for high traffic loads. As shown by Example 1, since the Min-BER-Fit strategy first
(a) First-Fit
(b) Min-BER-Fit Fig. 4.
Solutions to Example 1 computed by the three WA strategies.
300
70 First−Fit Min−BER−Fit Best−BER−Fit
Min−BER−Fit Best−BER−Fit 60
200
50 Gain %
Number of regenerators
250
150
40
100
30
50
20
0 0
100
Fig. 5.
200
300 400 500 Number of demands
600
700
10 100
800
Number of regenerators w.r.t. traffic load.
200
Fig. 6.
300 400 500 Number of demands
600
700
Gain in the number of regenerators w.r.t. traffic load.
35
260
30 Number of regenerators
240 Number of regenerators
(c) Best-BER-Fit
220 200 180 160 First−Fit Min−BER−Fit Best−BER−Fit
140
Number of regenerators w.r.t. BER threshold.
25 20 15 10 5 0
120 −4 −5 −6 −7 −8 −9 −10 −11 −12 −13 −14 10 10 10 10 10 10 10 10 10 10 10 BER threshold
Fig. 7.
First−Fit Min−BER−Fit Best−BER−Fit
Fig. 8.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Node number
Geographical distribution of the regenerators w.r.t. WA strategies.
consumes the wavelengths achieving the best performances, it no longer has the possibility to make better choices for heavy traffic loads as opposed to Best-BER-Fit. In any case, both the Min-BER-Fit and the Best-BER-Fit strategies always outperform the First-Fit strategy. In Figure 6, we plot the gain in the mean value of the number of regenerators achieved by both the Min-BER-Fit and the Best-BER-Fit strategies w.r.t. the First-Fit strategy. We notice an average benefit of 66% and 50% achieved by MinBER-Fit and Best-BER-Fit respectively for low traffic loads. Nevertheless, for heavy traffic load, Best-BER-Fit provides up to a 15% gain in the number of regenerators whereas MinBER-Fit only provides a gain of 11.5%. 25
Number of regenerators
20
15
10
5
0
Fig. 9.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Node number
Location distribution of regenerators for different traffic matrices.
Figure 7 shows the mean values of the number of regenerators required to satisfy 400 demands for various values of the BER threshold for the three WA strategies. As expected, the number of regenerators increases for all strategies as the BER requirement becomes stronger. Once again, Min-BER-Fit and Best-BER-Fit outperform First-Fit. Considering a low threshold (below 10−6 ) the Min-BER-Fit strategy outperforms the BestBER-Fit strategy which confirms the results shown in Figure 5. In Figure 8, we draw the geographical distribution of the regenerators’ locations for the three WA strategies considering a traffic load of 400 demands and a BER threshold of 10−5 . We notice that Nodes 6 and 7 contain the highest number of regenerators. These nodes also have the highest physical degree (5). Moreover, the links connecting nodes 6 and 7 to their neighbors have an important length, especially the links 7 − 14, 7 − 11 and 6 − 12. Crossing these links, the signal has a high probability to be regenerated. In all cases, the demands use the same routes. By changing the WA strategy, only the used wavelengths are changed. This can explain the limited impact of the used WA strategy on the regenerators’ distribution. The Min-BER-Fit strategy leads to the lowest number of used regenerators but the Best-BER-Fit leads to a better-balanced regenerator distribution. A balanced
distribution is worthwhile considering that in practical node implementations the number of regenerators is limited (the regenerators are often provided in racks). Actually, other factors, such as traffic distribution, also have an impact on the regenerators’ location. In Figure 9, we show the geographical distribution of regenerators for 10 different traffic matrices with 400 demands. Actually, we provide the mean values of the number of regenerators at each node as well as the confidence intervals. These results have been obtained using the Best-BER-Fit strategy. We notice that the traffic distribution fluctuations result in little fluctuations in the regenerators’s locations while the shape of the distribution remains the same. Thus, we may suggest that the network topology actually is the key factor explaining the regenerators’ locations. V. C ONCLUSION In this paper, we have proposed two new QoT-aware wavelength assignment strategies, namely Min-BER-Fit and BestBER-Fit. Simulations applied to the 18-node north American backbone compare the proposed strategies to the First-Fit which we used in previous work. Both Min-BER-Fit and BestBER-Fit have been shown to outperform the First-Fit strategy for various traffic loads and various BER threshold values. The numerical results show that the Min-BER-Fit strategy is the best one for low traffic loads. However, for heavy traffic loads, the Best-BER-Fit strategy is the more efficient by achieving an average benefit of 15% whereas Min-BER-Fit only achieved a benefit of 11.5%. We have also investigated the geographical distribution of the regenerators under the different WA strategies. The obtained results show that Best-BERFit leads to a better balance in the regenerators’ distribution. Nevertheless, since the routes are not changed, the impact of the WA strategies is limited. Therefore, the network topology can be considered as one of the key factors explaining the regenerators’ distribution. This conclusion suggests a different approach to deal with the problem of routing, wavelength assignment and regenerator placement, namely it may be more profitable to deal with regenerator placement before solving the RWA problem. Such an approach is the subject of our current work. R EFERENCES [1] R. E. Wagner, L. Nederlof, S. De Maesschalck, “The potential of optical layer networks,” in Proc. OFC’01, vol. 2, 2001, pp. TuT3–1 – TuT3–3. [2] A. Saleh, “Islands of transparency: an emerging reality in multiwavelength optical networking,” presented at IEEE/LEOS Summer Toptical Meeting Broadband Optical Networks Technologies, 1998. [3] B. Ramamurthy, H. Feng, D. Datta, J. P. Heritage, B. Mukherjee, “Transparent vs. opaque vs. translucent wavelength-routed optical networks,” in Proc. OFC’99, vol. 1, 1999, pp. 59–61. [4] B. Ramamurthy, S. Yaragorla, X. Yang, “Translucent optical WDM networks for the next-generation backbone networks,” in Proc. GLOBECOM’01, vol. 1, 2001, pp. 60–64. [5] X. Yang, B. Ramamurthy, “Sparse regeneration in translucent wavelength-routed optical networks: architecture, network design and wavelength routing,” Photonic Network Communications, vol. 10, no. 1, pp. 39–50, 2005. [6] T. Deng, S. Subramaniam, J. Xu, “Crosstalk-aware wavelength assignment in dynamic wavelength-routed optical networks,” in Proc. Broadnets’04, 2004, pp. 140–149.
[7] R. Cardillo, V. Curri, M. Mellia, “Considering transmission impairments in wavelength routed networks,” in Proc. ONDM’05, 2005, pp. 421–429. [8] N. Puech, S. Al Zahr, M. Gagnaire, “Wavelength-dependent quality of transmission in WDM transparent optical networks,” in Proc. ICT’06, 2006, pp. 335–340. [9] S. Al Zahr, M. Gagnaire, N. Puech, M. Koub`aa, “Physical layer impairments in WDM core networks: a comparison between a north-American backbone and a pan-European backbone,” in Proc. Broadnets’05, 2005, pp. 335–340.
[10] M.A. Ezzahdi, S. Al Zahr, M. Koub`aa, N. Puech, M. Gagnaire, “LERP: a quality of transmission dependent heuristic for routing and wavelength assignment in hybrid WDM networks,” in Proc. ICCCN’06, 2006. [11] Y.R. Zhou, A. Lord, S. Santoni, D. Setti, T. Fischer, G. Lehmann, H. Bulow, H. Haunstein, A. Schinabeck, “PMD rules for physical constraint-based routing in all optical networks,” in Proc. OFC’06, 2006. [12] D. Epstein, “Finding the k shortest paths,” SIAM Journal of Computing, vol. 28, no. 2, pp. 652–673, 1998.
