Adapting Spectrum Resources using Predicted IP Traffic in Optical
Networks
Constantine A. Kyriakopoulos
1 a
, Petros Nicopolitidis
1 b
, Georgios I. Papadimitriou
1 c
and Emmanouel Varvarigos
2 d
1
Dept. of Informatics, Aristotle University, Thessaloniki, Greece
2
School of Electrical and Computer Engineering, National Technical University of Athens, Greece
Keywords:
Optical Networks, Particle Swarm Optimisation, Linear Regression, Analytics, Load Balancing, Traffic
Prediction.
Abstract:
Elastic optical networks provide the advantage of elaborate resource utilisation for achieving a wide range
of performance goals. Cross-layer optimisation is feasible by exploiting high layer IP traffic prediction for
achieving efficient lightpath establishment at the lower layer. Swarm Intelligence can provide a tool to adap-
tively allocate spectrum resources according to traffic analytics from the IP layer. A new algorithm is designed
and evaluated that exploits these analytics using particle swarm optimisation to allocate spectrum.
1 INTRODUCTION
Optical backbone networks reliably cover a wide
range of current and possibly future connectivity
needs. Intelligent resource allocation in this field un-
dergoes ongoing research efforts from the community.
Performance increases when resources are allocated
in an on-line fashion while the network is operating
(Kyriakopoulos et al., 2018), considering as input the
current traffic demand and adapting to available spec-
trum resources facilitating it efficiently.
Various types of technology are enabled at the
optical layer. Elastic optical network (EON) (Jinno,
2016) platforms offer flexibility for configuring, since
the usage of variable-rate transponders can provide
the right amount of resources on demand. Orthogo-
nal frequency division multiplexing (OFDM) (Chat-
terjee et al., 2017), (Zhang et al., 2012a) is adequate
to provide support for variable-rate light connections.
This is achieved by utilising many subcarriers for data
transfers. The overlapping of spectra between these
subcarriers facilitates the compactness of available
resources due to their orthogonal modulation. This
design increases the overall efficiency. Bandwidth-
variable transponders (BVTs) (Moreolo et al., 2016)
a
https://orcid.org/0000-0001-7874-2205
b
https://orcid.org/0000-0002-5059-3145
c
https://orcid.org/0000-0001-9529-9380
d
https://orcid.org/0000-0002-4942-1362
embed the enabling technologies for achieving these
goals.
In a cross-layer network design (Sartzetakis et al.,
2018), a relation is formed between the physical and
network layers. This is a push-pull design where both
layers facilitate each other for achieving important
performance goals. Traffic conditions taking place
at the IP layer may be exploited for efficiently estab-
lishing lightpaths at the physical layer. As an exam-
ple, connectivity between data centres follows spe-
cific traffic patterns. Predicting the state and varia-
tion of these patterns, useful analytics can be provided
to the physical layer for establishing lightpaths hav-
ing the right amount of spectrum resources for im-
proving performance. In the opposite direction, im-
pairments in the physical layer are estimated (Bouda
et al., 2018), (Fludger and Kupfer, 2016), (Beletsioti
et al., 2018) and considered for creating connections
from the above layers.
An adaptive tool that solves many optimisation
problems and is capable of exploiting higher layer
analytics for improving its performance is particle
swarm optimisation (PSO) (Mohemmed et al., 2008),
(Zhang et al., 2015b). Its applications include neural
network training, pattern classification and function
optimisation, among others. The main focus is the
emulation of animals’ social behaviour including in-
sects or birds. The main trait is that the individuals in-
side a group cooperate to find food. Each member in
Kyriakopoulos, C., Nicopolitidis, P., Papadimitriou, G. and Varvarigos, E.
Adapting Spectrum Resources using Predicted IP Traffic in Optical Networks.
DOI: 10.5220/0009819500530058
In Proceedings of the 17th International Joint Conference on e-Business and Telecommunications (ICETE 2020) - DCNET, OPTICS, SIGMAP and WINSYS, pages 53-58
ISBN: 978-989-758-445-9
Copyright
c
 2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
53

the swarm repositions itself by keeping track of past
movements, either its own or from its neighbourhood.
This type of social behaviour is standardised and can
be used to solve difficult computational problems. For
example, in a network graph topology, a population is
able to find paths with specific traits like short dis-
tances, or any other metric that replaces the weight of
an edge in the graph. Complexity decreases when this
logic is applied to large topologies.
A popular tool for usage in network traffic predic-
tion is linear regression (Mata et al., 2018). It is used
to fit a predictive model to a set of values, e.g., in
a Cartesian system. This way, when more values are
collected in the x-axis, the corresponding values in the
y-axis are predicted. It is used to detect the strength
of the response and explanatory variables. Models of
the regression are fitted by exploiting the least squares
method.
In related research (Morales et al., 2017) of this
field, a virtual topology reconfiguration mechanism is
proposed based on data analytics for traffic prediction.
It uses a machine learning algorithm that relies on an
artificial neural network (ANN) for providing adap-
tive traffic models. Data from traffic prediction are
used in the reconfiguration model and the problem is
solved utilising a heuristic method.
In this work, higher layer analytics are collected
by applying a linear regression method to predict fu-
ture traffic demand. These analytics are provided to a
PSO method for allocating the appropriate amount of
spectrum resources in an online fashion. Performance
is improved since the found paths carry spare band-
width that can facilitate future traffic demand. At the
same time, highly congested paths are avoided since
the PSO core adapts to those providing more spare
bandwidth to accommodate future demand. Accord-
ing to results, there is reduction in transponder num-
bers which leads to less power consumption. Also,
The percentage of valid paths to allocate resources
to is high and optical grooming is dominant at lower
rates. At the same time, path elongation is minor.
The rest of the text is organised as follows. Sec-
tion 2 describes the proposed Metis algorithm and
Section 3 presents the network environment that is
used to evaluate performance. Finally, in Section 4
simulation results are presented.
2 METIS ALGORITHM
Metis (a mythical titaness - mother of wisdom and
deep thought) relies on analytics from the IP layer to
establish lightpaths. Traffic requests between node
pairs arrive sequentially. A history log of previous
Figure 1: Ring particle neighbourhood.
values is maintained and a window of these is used to
predict the next arrival rate for every node pair. The
predicted value comprises a parameter for the edge
weight that the PSO core will use to calculate the ap-
propriate path when establishing lightpaths.
The PSO core uses the physical topology connec-
tions with modified weights. This way, it tries to find
paths avoiding routes where the load is predicted to
increase. The resulting resource allocation is more
balanced with low blocking probability, in compari-
son to the corresponding shortest path replacement.
Adaptivity to future load is a trait that renders Metis
suitable for online execution while network runtime
conditions vary.
PSO is a stochastic optimisation method (Zhang
et al., 2015b) that mimics the social behaviour of a
bird flock, etc. The algorithmic flow initiates with a
set of particles whose positions represent possible so-
lutions in the search space. The search for optimal po-
sitions leads to a solution by updating their velocities
in an iterative fashion. The fitness of each particle is
calculated, and the one having the highest value pro-
vides the best solution in the search space. Each parti-
cle’s velocity depends on the current and best position
it had so far. Also, it depends on the best neighbour
position. After a number of iterations, the solution is
provided by the particle that converged faster.
Particle population is organised in a ring topology
(Figure 1). The number of particles affects the per-
centage of valid paths to find. A larger number of
particles, in relation to large iteration number, pro-
vides more accurate results (using the appropriate fit-
ness function) but increases the computational com-
plexity. This is due to more fitness function evalu-
ations and candidate path constructions per particle.
Every particle alters its velocity and hence its posi-
tion according to the state of its neighbours.
Every incoming request is served by Metis which
is described in Algorithm 1. The first preprocess-
DCNET 2020 - 11th International Conference on Data Communication Networking
54

Algorithm 1: Metis Abstract Pseudo Code.
New incoming spectrum allocation request
For every topology edge, use analytics as weight
Feed PSO core with the updated topology
direct ← f alse
path = psoRouter(src, dest)
if path.size() > 0 then
for all edge ∈ pathEdges do
optical ← f alse
if optical = optGrooming(edge) 6= 0 then
continue
else
direct = directLP(edge)
end if
end for
else
// Enter failsafe mode
Get shortest path from ’src’ to ’dest’
for all edge ∈ pathEdges do
direct = directLP(edge)
end for
end if
ing event relates to the use of Equation 1’s result as
weight for every topology’s edge (connection). Mul-
tiple data transfers exist between node-pairs. For ex-
ample, if the third request is to be served, the previous
two comprise the prediction window. If the first two
are 10 and 20 Gbps, the predicted value for the next is
30 Gbps. The actual weight value to provide to PSO
core from Formula 1 is 35.
Next, the updated topology becomes PSO core’s
input for finding the appropriate path from initial node
to destination. If it succeeds, for every edge of the
path, optical grooming is attempted (Zhang et al.,
2012b). In this case, available transponder slices at
both ends of the edge comprise a newly established
lightpath (Kyriakopoulos et al., 2019). If no avail-
able slice(s) exist, one (or two) new transponder(s)
are used for creating the light connection.
It is possible for the PSO core not to find a valid
path. This happens if the number of iterations or
the population size is not large enough. The failsafe
mode initiates, where direct lightpaths (using two new
transponders) are established upon every edge of the
shortest path between end nodes.
Weight =
(
rate
n
+
ˆ
rate
n+1
2
,
ˆ
rate
n+1
> rate
n
rate
n
, otherwise
(1)
y = a
0
+ a
1
x + a
2
x
2
+ a
3
x
3
+ ·· · + a
n
x
n
+ ε (2)






y
1
y
2
y
3
.
.
.
y
n






=







1 x
1
x
2
1
··· x
m
1
1 x
2
x
2
2
··· x
m
2
1 x
3
x
2
3
··· x
m
3
.
.
.
.
.
.
.
.
.
.
.
.
1 x
n
x
2
n
··· x
m
n













a
0
a
1
a
2
.
.
.
a
m






+






ε
1
ε
2
ε
3
.
.
.
ε
n






(3)
−→
y = X
−→
a +
−→
ε (4)
−→
a =

X
T
X

−1
X
T
−→
y (5)
In Equation 2, y represents
ˆ
rate
n+1
which is the
predicted edge bandwidth. The previous values from
rate
1
to rate
n
comprise the prediction window. Pro-
viding to x the slot n + 1, y is calculated (Seber and
Lee, 2012). Values between y
1
. . . y
n
comprise the
window of previous rates. Values between x
1
. . . x
n
comprise the window of previous slots. A new repre-
sentation for the equation is in Formula 3 with the
purpose of finding the array of coefficients
−→
a . In
vector form is the Equation 4. Coefficients are cal-
culated from Equation 5 by using ordinary least mean
square estimation. The result is used in Equation 2
to calculate the next rate value which is the predicted
value. The ε values represent possible minor errors
which are ignored and the result is named an esti-
mated value.
3 NETWORK ENVIRONMENT
The purpose is to allocate spectrum resources while
incoming requests arrive one-by-one. The problem is
formally described and follows next.
• A directed graph is used to describe the topology
of the elastic network, i.e., G(V, E). Specifically,
V is the set of nodes and E is the set of links.
• A set of frequency slices F are used for transpon-
der end-point connections for each link ∈ E. F =
{ f
1
, f
2
, ·· · , f
n
}, where n is the ceiling of connec-
tions per fibre.
• A set of available modulation formats M =
{m
1
, m
2
, ·· · , m
n
}, where n is their maximum
number. Each format is described by a pair m =
h
f , r
i
, where f is the lightpath spectrum and r is
the optical reach.
• A set of traffic demands D that reside in a matrix.
Each entry is described by a tuple d =
h
s, d, b
i
,
where s is the request’s source, d the destination
and b represents the bitrate.
20 interconnected nodes comprise a topology to
use for evaluating Metis’ performance. Figure 2 de-
picts the connections and the corresponding weights.
Modulation is based on the available choices of
Table 1. These choices are input to the modulation
policy when light connections are established using
available transponder slices. Distance is a factor that
limits the subcarrier capacity and is obeyed for ev-
ery new request’s bitrate. To choose the appropriate
Adapting Spectrum Resources using Predicted IP Traffic in Optical Networks
55

Figure 2: 20-node topology (Mohemmed et al., 2008).
Table 1: Modulation Formats.
Format Subcarrier Capacity (Gbps) Distance (km)
BPSK 12.5 4000
QPSK 25 2000
8QAM 37.5 1000
16QAM 50 500
32QAM 62.5 250
64QAM 75 125
format, all these are sorted in descending order. The
value that is ceiling to incoming request’s rate is kept
by the policy.
Figure 3 contains the low level details of estab-
lishing the new lightpath λ
3
when λ
1
and λ
2
are al-
ready established. A prerequisite is the existence of
available transponder slices at source (left) and desti-
nation (right) nodes. The intermediate node consists
of a transmitter and a receiver.
4 RESULTS
The simulating environment consists of specific pa-
rameters that follow next. Variable-rate transponders
utilise up to 10 lightpath connections. Two adja-
cent frequency slots comprise a guardband. Avail-
able modulation formats are in Table 1 and each of
Figure 3: Optical grooming.
Figure 4: Transponders according to increased traffic de-
mand.
them has its own spectrum range. A table of spectrum
values according to data rates is found in Reference
(Zhang et al., 2015a).
Traffic demand values are generated by a random
function in the range [40, 2X − 40] Gbps, using steps
of 40 Gbps. Variable X ∈ {40, 80, 120, 160, 200}. 500
requests are established between uniformly selected
node-pairs. Past values of a node-pair comprise the
prediction window.
The PSO core relies on 750 iterations and popu-
lation size of 40, unless otherwise noted. The linear
regression prediction method uses a window of 5 pre-
vious traffic values between each node-pair.
The simulating environment is designed and im-
plemented in Modern C++ with the Clang/LLVM 10
compiler, the aid of Boost graph library 1.67 and Ar-
madillo linear algebra library 9, on x64 Debian 10.
In Figure 4, the number of utilised transponders
increases according to traffic demand. When its aver-
age value reaches 200 Gbps, there are many requests
that exceed the upper transponder limit of 400 Gbps,
so optical grooming is not feasible in this case. This
is the reason for the high performance of Metis at
lower rates. Its slight path elongation results in more
transponder usage at higher bitrates, in comparison
to the direct lightpath establishment that relies on the
shortest path between end nodes. When the average
rate reaches 200 Gbps, some values are close to the
maximum supported transponder rate which is 400
Gbps. So, optical grooming on existing transponders
is not feasible and new ones must be utilised on paths
not being shorter.
In Figure 5, the percentage of valid paths found
by the embedded PSO mechanism in Metis, is de-
picted according to the particle population size in-
crease. When the population is low, particles may not
DCNET 2020 - 11th International Conference on Data Communication Networking
56

Figure 5: PSO valid path percentage according to popula-
tion size.
Figure 6: Optical grooming ratio according to traffic.
converge to a valid solution. At higher values (x axis),
almost every execution returns valid paths, so the per-
centage (y axis) reaches the value of 100%. The aver-
age request rate is 80 Gbps with 10 iterations between
particles. The percentage of failures for PSO to find
paths can be considered as blocking probability for
the PSO core, since it then enters the failsafe mode
(Algorithm 1).
In Figure 6, the percentage of optically groomed
lightpaths is depicted according to the increasing av-
erage traffic demand value. Since the availability of
optical grooming at values above 400 Gbps is non-
existent, the percentage keeps decreasing. Metis’ per-
formance is high between low and mid-range traffic
values. From the grooming perspective, the percent-
age of direct lightpaths can be considered as blocking
probability.
In Figure 7, the effect of not utilising the short-
Figure 7: Hop-count according to increased traffic demand.
est path between request end nodes is depicted. This
compensates due to the higher performance as de-
scribed in the previous graphs. Also, longer paths are
established for accommodating the prediction of IP
layer’s future traffic demand.
5 CONCLUSIONS
Higher layer analytics are exploited for improving the
efficiency of the lightpath establishment procedures at
the lower layer. The particle swarm optimisation core
exploits the predicted future traffic demand and finds
paths with higher available spectrum resources. The
adaptivity to future IP traffic leads to higher overall
performance, in relation intelligent spectrum alloca-
tion.
ACKNOWLEDGEMENTS
This research was co-financed by the European Union
and Greek National Funds through the Operational
Program Competitiveness, Entrepreneurship and In-
novation, under the call RESEARCH-CREATE-
INNOVATE (project code: T1EDK-05061).
REFERENCES
Abkenar, F. S. and Rahbar, A. G. (2017). Study and analysis
of routing and spectrum allocation (rsa) and routing,
modulation and spectrum allocation (rmsa) algorithms
in elastic optical networks (eons). Optical Switching
and Networking, 23:5–39.
Adapting Spectrum Resources using Predicted IP Traffic in Optical Networks
57

Beletsioti, G. A., Papadimitriou, G. I., Nicopolitidis, P.,
and Miliou, A. N. (2018). Earthquake tolerant en-
ergy aware algorithms: A new approach to the design
of wdm backbone networks. IEEE Transactions on
Green Communications and Networking, 2(4):1164–
1173.
Bouda, M., Oda, S., Vassilieva, O., Miyabe, M., Yoshida,
S., Katagiri, T., Aoki, Y., Hoshida, T., and Ikeuchi, T.
(2018). Accurate prediction of quality of transmission
based on a dynamically configurable optical impair-
ment model. Journal of Optical Communications and
Networking, 10(1):A102–A109.
Chatterjee, B. C., Ba, S., and Oki, E. (2017). Fragmenta-
tion problems and management approaches in elastic
optical networks: A survey. IEEE Communications
Surveys & Tutorials, 20(1):183–210.
Deepa, O. and Senthilkumar, A. (2016). Swarm intelli-
gence from natural to artificial systems: Ant colony
optimization. Networks (Graph-Hoc), 8(1):9–17.
Fludger, C. R. and Kupfer, T. (2016). Transmitter im-
pairment mitigation and monitoring for high baud-
rate, high order modulation systems. In ECOC 2016;
42nd European Conference on Optical Communica-
tion, pages 1–3. VDE.
Jinno, M. (2016). Elastic optical networking: Roles and
benefits in beyond 100-gb/s era. Journal of Lightwave
Technology, 35(5):1116–1124.
Kyriakopoulos, C. A., Nicopolitidis, P., Papadimitriou,
G. I., and Varvarigos, E. (2019). Fast energy-efficient
design in elastic optical networks based on signal
overlap. IEEE Access, 7:113931–113941.
Kyriakopoulos, C. A., Papadimitriou, G. I., and Nicopoli-
tidis, P. (2018). Towards energy efficiency in vir-
tual topology design of elastic optical networks.
International Journal of Communication Systems,
31(13):e3727.
Mata, J., De Miguel, I., Duran, R. J., Merayo, N., Singh,
S. K., Jukan, A., and Chamania, M. (2018). Artificial
intelligence (ai) methods in optical networks: A com-
prehensive survey. Optical Switching and Networking,
28:43–57.
Mohemmed, A. W., Sahoo, N. C., and Geok, T. K. (2008).
Solving shortest path problem using particle swarm
optimization. Applied Soft Computing, 8(4):1643–
1653.
Morales, F., Ruiz, M., Gifre, L., Contreras, L. M., L
´
opez,
V., and Velasco, L. (2017). Virtual network topology
adaptability based on data analytics for traffic predic-
tion. IEEE/OSA Journal of Optical Communications
and Networking, 9(1):A35–A45.
Moreolo, M. S., Fabrega, J. M., Nadal, L., V
´
ılchez, F. J.,
Mayoral, A., Vilalta, R., Mu
˜
noz, R., Casellas, R.,
Mart
´
ınez, R., Nishihara, M., et al. (2016). Sdn-
enabled sliceable bvt based on multicarrier technology
for multiflow rate/distance and grid adaptation. Jour-
nal of Lightwave Technology, 34(6):1516–1522.
Sartzetakis, I., Christodoulopoulos, K., and Varvarigos, E.
(2018). Cross-layer adaptive elastic optical networks.
IEEE/OSA Journal of Optical Communications and
Networking, 10(2):A154–A164.
Seber, G. A. and Lee, A. J. (2012). Linear regression anal-
ysis, volume 329. John Wiley & Sons.
Zhang, G., De Leenheer, M., Morea, A., and Mukherjee, B.
(2012a). A survey on ofdm-based elastic core optical
networking. IEEE Communications Surveys & Tuto-
rials, 15(1):65–87.
Zhang, G., De Leenheer, M., and Mukherjee, B. (2012b).
Optical traffic grooming in ofdm-based elastic optical
networks. Journal of Optical Communications and
Networking, 4(11):B17–B25.
Zhang, J., Zhao, Y., Yu, X., Zhang, J., Song, M., Ji, Y.,
and Mukherjee, B. (2015a). Energy-efficient traffic
grooming in sliceable-transponder-equipped ip-over-
elastic optical networks [invited]. Journal of Optical
Communications and Networking, 7(1):A142–A152.
Zhang, Y., Wang, S., and Ji, G. (2015b). A comprehensive
survey on particle swarm optimization algorithm and
its applications. Mathematical Problems in Engineer-
ing, 2015.
DCNET 2020 - 11th International Conference on Data Communication Networking
58