Planning Optical Networks for Unexpected Traffic Growth

Sai Kireet Patri(1,2) , Achim Autenrieth(1) , Jörg-Peter Elbers(1) , Carmen Mas Machuca(2)

(1)
ADVA, Fraunhoferstr. 9A, 82152 Martinsried, Germany SPatri@adva.com
(2)
Chair of Communication Networks, TU Munich, Arcisstr. 21, 80331 Munich, Germany

Abstract A lightpath configuration algorithm considering a multi-period traffic model is presented and
evaluated to a realistic German core network study. Strategically exploiting BVT excess capacity, we
show that offered traffic can be carried up to an additional five years in all traffic growth scenarios.

Introduction Traffic Generator and Growth Model

With 5G deployment becoming a reality across As network operators do not share their offered
countries, core network operators are realising traffic data openly due to confidentiality, network
that deployed optical networks need to be up- planners have relied on traditional empirical traffic
arXiv:2012.04360v1 [cs.NI] 8 Dec 2020

graded to quickly capitalize gains. Due to the models like the gravity model[6] , which assumes
recent pandemic, many network operators (e.g., traffic to be proportional to the human popula-
BT and Telefónica), have reported a 30-50% in- tion of the source and destination. However, in
crease in broadband traffic in Q1 2020, compared present times, most of the traffic between two lo-
to Q4 2019[1] . Given the worldwide impact on lo- cations are exchanged between either Data Cen-
gistics, operators must invest in software config- ters (DCs) or Internet Exchange Points (IXPs) and
urable optical terminals[2] and Bandwidth Variable there exists a higher correlation of offered traf-
Transponders (BVTs), which offer opportunities to fic to node degree as compared to human pop-
increase network throughput in order to carry the ulation. For e.g., Frankfurt and Duesseldorf are
growth in offered traffic. We define offered traffic less populated cities as compared to Hamburg
as the total throughput of the network if all de- and Munich, but exchange traffic at higher rates[7] .
mands are met by adding sufficient BVTs in the Based on these assumptions, we define initial
network. traffic (time t = 0), between source i and desti-
Traditionally, lightpath configuration in network nation j as:
planning consists of algorithms exploiting Inte- (
N


ger Linear Programming (ILP) as well as heuris- 2· 2 · ∆i · ∆j if N > 2 · N̄
τ (i, j, 0)[Gbps] =
tics. A link state heuristic for optical networks[3] N · ∆ i · ∆j otherwise
is of particular interest; however, no insights on (1)
its usability in a multi-period scenario are pro- where N̄ is the average node degree of the topol-
vided. The SNAP Algorithm[4] offers progressive
P
ogy N = κ∈{i,j} Nκ is the combined node de-
traffic loading, randomly allocating traffic between gree, and ∆κ is the absolute difference between
nodes without considering realistic growth. Both the number of co-located DCs and IXPs at a given
works also assume homogeneous spans across node κ[8] .
their physical topology, limitations of which have For t > 0, the offered traffic between nodes is
been discussed in our previous work[5] . given as τ (i, j, t) = γt ·τ (i, j, 0), where γt is the ex-
In this work, we first present a traffic gener- pected growth in offered traffic[9] . For unexpected
ator and discuss two traffic growth models viz., traffic growth, we modify γt with an increase of
Expected and Unexpected. Using our developed 30-90% between 2023 and 2026, as compared to
solution (Scheme 1) on a German core topol- its original value. As seen in Fig. 2a, this brings
ogy, we show that the offered traffic can be about an overall realistic increase of 40% in the
met by strategically leveraging the configured aggregate offered traffic growth for the given net-
BVTs’ capacity, hence saving bandwidth for future work topology.
growth. Scheme 1 is evaluated against a simple
baseline (Scheme 2), which optimally adds new Algorithm and Multi-period Planning Study
BVTs, instead of upgrading the deployed BVTs to The multi-period planning algorithm is shown in
a higher datarate when possible. Finally, results Fig. 1. It takes as input the topology, along with
in terms of aggregate network throughput (carried information on the number of DCs and IXPs at
traffic) and provisioned BVTs are discussed. each source destination pair (i, j). We then find
 nel configurations ρi,j,t ⊂ Ci,j,t is found, hav-
ing their channel bandwidth BWρ less than or
equal to the provisioned LP’s channel band-
width BWλ . This ensures upgrade of provi-
sioned LPs at the same central channel fre-
quency, without additional bandwidth usage. We
then iterate over ρi,j,t sorted according to high-
est datarate (DR) to find the first ρ, which has a
higher DR than λ. If found, λ is updated with ρ
and the spare capacity generated can be used to
satisfy some of the additional traffic, thereby re-
ducing θi,j,t .
P
minimize c ∈ Ci,j,t nc
P
subject to : θi,j,t ≤ nc · DRc < θi,j,t + δ
P P
0< nc · ηN LIc ≤ ∀λ∈Li,j,t ηN LIλ
(2)
After the provisioned LPs are upgraded, we find
the number of additional LPs nc of each valid
channel configuration c, which are needed to
carry the remaining θi,j,t using the ILP shown in
Eq. 2. This ILP minimizes integer nc , such that
Fig. 1: Proposed solution flow in a multi-period scenario with the configured total datarate can only be over-
two schemes and optional physical network upgrade provisioned by δ Gbps. For our network study,
we fix δ as 100 Gbps. We also use a power in-
the initial offered traffic using Eq. 1 and also
dependent NLI co-efficient constraint, ηN LI , fol-
do a Routing Wavelength and Spectrum Assign-
lowing definitions and calculations of ACF-EGN
ment (RWSA), using a weighted probabilistic rout-
model[10] . This NLI constraint restricts the to-
ing based on Yen’s k-Shortest Path Algorithm and
tal ηN LI being added every t to the sum of pro-
the number of continuous empty frequency slots
visioned LPs’ ηN LI . The LPs are then added
in each of the paths. This heuristic for routing
to Li,j,t+1 , defined as a set of configured LPs be-
adds a randomness in choosing the candidate
tween nodes (i, j) for the next planning period.
path list for each (i, j). We use first-fit channel al-
After additions are completed, we also check
location strategy for spectrum assignment.
on the number of empty frequency slots in the
The objective of the algorithm is to minimize the
link. If more than 75% of the frequency slots have
number of lightpaths (LPs) added to the network
been filled up in any of the given links, it warns
for planning period t, while trying to meet the of-
operators about upcoming frequency slot satura-
fered aggregate traffic. It is pertinent to note that
tion in the configured fiber pair so that they plan
one LP is associated to one BVT. To begin, we
a localized physical upgrade by exploring either
take as input the candidate path list of all de-
additional bands or utilizing available dark fiber
mands (Dt ), all currently provisioned LPs in the
pairs[4] .
network (Lt ) and the traffic matrix (Tt ) calculated
using Eq. 1 and Fig. 2a. For each candidate path Nodes Span Noise Node
Network Links variation Figure Deg.(min/
di,j,t ∈ Dt , we compute a list of LPs Li,j,t ⊂ Lt Demands (km) (dB)[4] avg/max)
already present between nodes (i, j) and addi- Germany 17 17/26/272 30-120 4.3 2/3.05/6
US Abilene 12/15/132 20-100 4.3 1/2.5/4
tional offered traffic θi,j,t , defined as τ (i, j, t) − Tab. 1: Reference topology information[11]
τ (i, j, 0). Using a datarate granularity of 50 Gbps
between 100-600 Gbps in steps of 50 Gbps with Using the discussed flow, we setup a multi-
QAM values of QPSK, 8, 16, 32 and 64 QAM, period network study for a 17 node German back-
we generate more than 60 different channel con- bone network with single mode fibers having het-
figurations (combination of {Datarate, QAM }) for erogeneous span lengths[11] . We assume single
each BVT. The list of valid channel configura- fiber pair C-Band operation with variable gain ED-
tions Ci,j,t is then filtered using HeCSON[5] . For FAs having constant noise figure as shown in
every provisioned LP (λ ∈ Li,j,t ), a set of chan- Tab. 1. The topological details, initial offered traf-
 225
300 Expected Growth Expected Total Unexpected Total
Unexpected Growth 200 Offered Traffic 300 Offered Traffic
Scheme 1 Scheme 1
250 41.67% 175 Scheme 2 250 Scheme 2
Scheme 1
& Network Upgrade

Throughput (Tbps)
Throughput (Tbps)
150
Demand (Tbps)

200 200
125
150 150
100
100 75 100

50 50
50
25
2020 2022 2024 2026 2028 2030 2020 2022 2024 2026 2028 2030 2020 2022 2024 2026 2028 2030
Planning Period Planning Period Planning Period
(a) Aggregate Offered Traffic Growth models (b) Expected Traffic Throughput (c) Unexpected Traffic Throughput
Fig. 2: Yearly aggregate throughput of Germany17 topology from 2020-2030

fic and results for other topologies are also made arises as to when must the operator upgrade their
publicly available[11] . network. In the evaluated topology, we observe
that three links cross the occupied slot threshold
Results and Discussion
of 75% in the year 2027, which is when additional
We evaluate Scheme 1 and Scheme 2 in afore-
capacity must be planned for. Post upgrade, the
mentioned traffic scenarios. It must be noted that
offered traffic can be met by the algorithm.
an additional scheme, where all BVTs are placed
In Fig. 3 we see that Scheme 1 achieves a
according to their maximum possible datarate has
40% increase in throughput for the same num-
been evaluated in previous work[5] , compared to
ber of BVTs in Germany 17 topology. Conversely,
which, both Scheme 1 and Scheme 2 provide
similar throughput can be achieved by Scheme 1
a higher network throughput. For expected traf-
utilizing 25% lesser BVTs. For the 12 node US
fic growth shown in Fig. 2b, we see that of-
Abilene topology (US12), a 30% increase in
fered traffic can be satisfied by both schemes till
throughput by using 18% lesser BVTs is ob-
2025. After which, we observe that the overall
served. The reduction may be because of long
expected throughput for Scheme 2 (dashed line
distance demands in the US12 topology, due
with squares) cannot meet the offered traffic. The
to which offered traffic can only be met using
cause may be attributed to the lack of additional
more BVTs at lower modulation formats. How-
frequency slots on some links, which cause ad-
ever, Scheme 1 still provides savings on the num-
ditional LPs (nc ) found in Eq. 2 to remain un-
ber of BVTs.
provisioned. This translates to a loss of approxi-
mately 40 Tbps expected throughput at 2030. In Conclusions
the same scenario, we see that Scheme 1 allows With increased global travel restrictions, core net-
us to meet offered traffic without the need of a net- work traffic is bound to grow at a rate higher than
work upgrade for the entire t. Similarly, for unex- initial forecasts. Pragmatically, operators may use
1000 BVT over-provisioning to meet such surges in traf-
fic. However, our proposed solution, evaluated
500 with two network planning schemes, shows that
# BVTs in network

a 40% overall increase in aggregate offered traf-

fic for a realistic German core network can be
100
carried up to five years, by efficiently planning
Scheme 1 Germany17 the datarate and bandwidth usage of configured
Scheme 2 Germany17
Scheme 1 US12
Scheme 2 US12 BVTs. For higher traffic growth, an estimate on
50 100 150 200 250 300 when to plan a physical network upgrade is also
Offered Throughput(Tbps)
Fig. 3: BVTs vs Throughput for Germany 17 and US 12 provided.
pected traffic growth (shown in in Fig. 2c), upgrad-
Acknowledgements
ing provisioned LPs using Scheme 1 enables the
This work is partially funded by Germany’s Fed-
operators to utilize the C-Band for five additional
eral Ministry of Education and Research under
years. However, Scheme 1 also shows a down-
project OptiCON (grant IDs #16KIS0989K and
ward trend in provisioned LPs, beginning from
#16KIS0991). The authors would also like to
2029. We infer that in order to meet the offered
thank Dr. Jose-Juan Pedreno-Manresa for fruitful
traffic, network operators would have to upgrade
discussions.
their network, using new equipment and fibers or
by exploring additional bands. The question then
References
[1] Keeping the Internet up and running in times of crisis,
Accessed: Aug 2020. [Online]. Available: www . oecd .
org/coronavirus/policy-responses/keeping-the-
internet- up- and- running- in- times- of- crisis-
4017c4c9.
[2] ADVA Teraflex, Accessed: Aug 2020. [Online]. Avail-
able: www . adva . com / en / products / open - optical -
transport/fsp-3000-open-terminals/teraflex.
[3] R. Wang et al., “Load-aware Nonlinearity Estimation
for Elastic Optical Network Resource Optimization and
Management”, IEEE/OSA JOCN, vol. 11, no. 5, 2019.
[4] E. Virgillito et al., “Statistical Assessment of Open Opti-
cal Networks”, MDPI Photonics, vol. 6, no. 2, 2019.
[5] S. K. Patri et al., “HeCSON: Heuristic for Configuration
Selection in Optical Network Planning”, in Proc. OFC,
2020.
[6] A. Dwivedi et al., “Traffic model for USA long-distance
optical network”, in Proc. OFC, 2000.
[7] Germany DE-CIX, Accessed: Aug 2020. [Online]. Avail-
able: www.de-cix.net/en/locations/germany.
[8] Data Center Map, Accessed: Aug 2020. [Online]. Avail-
able: www.datacentermap.com.
[9] O. Turkcu et al., “Robust Optical Networks with Emerg-
ing Coherent Technologies and Traffic Trends”, in Ad-
vanced Photonics, 2018.
[10] P. Poggiolini et al., “Accurate Non-Linearity Fully-
Closed-Form Formula based on the GN/EGN Model
and Large-Data-Set Fitting”, in Proc. OFC, 2019.
[11] Physical Network Information, Accessed: Aug 2020.
[Online]. Available: www . github . com / SaiPatri /
PhyNWInfo.