Photonic Network Communications manuscript No.

(will be inserted by the editor)

Spectrum fragmentation issue in flexible optical networks

Analysis and good practices

Djamel Amar · Esther Le Rouzic · Nicolas Brochier · Jean-Luc Auge ·

Catherine Lepers · Nancy Perrot

Received: date / Accepted: date

Abstract Flexible grid optical networks allow an effi- tighter which may reduce transmission performance due
cient utilization of spectrum resources using 12.5 GHz to filtering limits.
frequency slot multiples instead of fixed spacing intro- Superchannels (SCs) [11] are an interesting solution
ducing however Spectrum Fragmentation (SF). In this (defined as a single routing entity compound of several
work we offer an exhaustive analysis and an accurate sub-channels each being generated by one transpon-
evaluation for SF issue in flexible optical networks. We der) to save spectrum resources since guard bands are
also propose new metrics for fragmentation measure- only required between SCs and not between each sub-
ments and some approaches to face such problem. channel composing the SC. SCs offer an ideal trade-
off between spectrum saving and transmission perfor-
Keywords Spectrum fragmentation · Flexible optical
mance.
networks · Optical network dimensioning
However, by considering both SCs and gridless tech-
nology, the efficient grid management becomes more
complex since different channel spectrum granularities
1 Introduction
have to be considered during the network dimensioning.
Networks operated with this flexible grid generate
Gridless technology is known to improve the spectral
SF when used with multiple types of transponders. The
efficiency of optical networks by attributing a certain
optical channels may then have different spectrum widths
12.5 GHz frequency slot number to each optical channel
(granularities) in terms of number of slots. This frag-
instead of using the traditional 50 GHz grid spacing [1].
mentation partitions the spectrum and creates isolated
This fine slot width and the grid flexibility allow an
slots or bands of slots which may not be used by some
adjusted tailoring of the spectrum to the real channel
types of transponders.
width to save spectrum. Nevertheless, channels become
In the literature, dynamic traffic scenarios have been
D. mar · E. Le Rouzic · N. Brochier · J.L. Auge often considered to evaluate the SF. However, in core
Orange Labs networks, traffic behavior is instead incremental.
2 Avenue Peirre Marzin
22300 Lannion, France
In this paper, we aim to evaluate to what extent
E-mail: [email protected] SF is a problem in the context of an operator network,
E-mail: [email protected] to identify parameters that influence SF and to derive
N. Perrot good practices in this incremental context.
Orange Labs The paper is organized as follows. In Section 2, we
38-40 rue du Gnral Leclerc give a brief overview on SF in optical networks. Some
92794 Issy-Les-Moulineaux, France
E-mail: [email protected]
existing metrics for SF measurements are presented and
compared with a new metrics we propose in this work
C. Lepers
Mines-Telecom/Telecom SudParis
using a Markov model in Section 3. In Section 4 net-
9 Rue Charles Fourier work model is defined in the context of an operator
91000 vry, France network. Numerical results are discussed in Section 5
E-mail: [email protected] before concluding in Section 6
2 Djamel Amar et al.

of networks, new demands are thus more likely to be

blocked due to fragmentation.
Several methods have been proposed to solve SF
issue. The simplest method is a fragmentation aware
Routing and Spectrum Assignment (RSA) algorithm
that specifies how traffic demands are routed in order
to minimize fragmentation [16,10]. Approaches split-
ting the blocked demand into sub-demands in order
to use finer granularities have also been proposed [6,
7]. Another intuitive method consists of defragmenting
the spectrum in such a way that free blocks are contigu-
ous [12]. These methods are either complex, inefficient,
or difficult to implement. In this work, we identify other
traffic engineering approaches that can be used to face
such a problem.
Each method requires a fragmentation metrics to
judge its robustness. Different metrics have been pro-
posed to quantify the fragmentation for different pur-
poses. These metrics are used to assess the level of frag-
mentation of a given network state [13]. They are also
Fig. 1 Horizontal and vertical fragmentation of spectrum used to optimize the network dimensioning: to improve
resources the provisioning policy if integrated in a fragmentation
aware RSA algorithm [15].

2 State of the art

3 Fragmentation metrics
Fragmentation problem already exists in other domains
than optical networks. It has been addressed, for time There are two SF effects in optical networks. The abso-
slot management in SONET (Synchronous Optical Net- lute spectrum loss occurs when some free slots cannot
works) protocol [2], and CDMA (Code Division Multi- be used whatever the required granularity is. This kind
ple Access) systems in wireless networks [3]. Similarly, of loss can be modeled as a knapsack problem whose
in computer memory, data fragmentation has been stud- the objective is to optimally fill the fiber (knapsack)
ied for a while [5]. However, the optical network context with used granularities (objects): G= {G1 , G2 ...Gm }.
(routing, spectrum continuity, number of slots, network Free slots (places) are then completely lost and cannot
size...), makes the SF issue different. be used unless a spectrum defragmentation process is
performed.
The SF in optical networks can be viewed accord-
Model (1) allows quantifying these lost slots (LS =
ing to two dimensions (Fig. 1). The first one occurs
S − Z ) where S denotes the size of the block to be
when some non-continuous free slots are created (hori-
evaluated, Z is the maximum number of slots that can
zontal fragmentation). Such fragmentation may be min-
be used in the most favorable case, and Yi is the num-
imized by maximizing the spectrum reuse, for example,
ber of times the granularity Gi ∈ G must be used in
thanks to a graph coloring method, a heuristic [16] or
optimal solution.
in the worst case by installing new wavelength convert-
ers. The second one occurs when the free spectrum is
divided into a set of blocks smaller than the demand
X
M ax Z = Yi Gi (1a)
width (vertical fragmentation). Therefore, despite the i
availability of resources, some demands can be blocked Subject to Z ≤ S (1b)
due to non-contiguous free slots. We consider only this
Yi ∈ Z+
0 (1c)
kind of fragmentation because it is specific to gridless
technology and does not concern traditional fixed grid The second kind of spectrum loss is relative as large
optical networks. granularities are more likely to be blocked than minor
The problem of fragmentation is usually considered ones. Thus, for a given state of the fiber a relative value
as serious in dynamic flexible optical networks [12] in is needed to assess the level of fragmentation. This is
which demands arrive and vanish randomly. In this type what some existing metrics are aimed at.
Spectrum fragmentation issue in flexible optical networks 3

The most used criterion to demonstrate that a given Table 1 Characteristics of metrics
metrics is minimizing the fragmentation is the traffic
blocking ratio between blocked demands and served de- Characteristics
metrics
mands. However, the information given by this crite- Complexity (na , mb ) Output Value
rion is insufficient, as traffic blocking ratio can reflect interval meaning
other blocking origins such as the lack of resources or EF n+1 [0, 1[ ***
the spectrum continuity constraint in transparent net-
SE 4n [0, +∞[ ***
works. For this reason, it is relevant to identify another
1
comparison criterion to quantify blocking caused by SF. ABP 4n(m − 4
) [0, 1] ****

To compare different metrics, mathematical model- a

number of fragments
ing methods are more efficient and more accurate com- b
number of granularities
pared with simulations ones. A Markov chain is pro-
posed in [9] to compare usual spectrum allocation poli-
cies in gridless network. To evaluate metrics perfor- 3.1 External Fragmentation metrics (EF)
mance, we adapt this model to identify the metrics
which allows the best spectrum use and minimizes the Fragmentation in computer memory has been well stud-
fragmentation level. To that aim, we use the blocking ied in the literature. Although, it is different from SF
probability due to fragmentation as an efficient criterion problem, external fragmentation metrics are still ap-
for the comparison (referred as fragmentation blocking plicable in optical network context. The most common
probability). metrics consists to calculate the ratio of the largest free
It is difficult to conceive an exhaustive fragmenta- contiguous fragment of memory to free memory [5] as
tion metrics that is able to reflect all fragmentation in 2.
damages while keeping a low complexity compatible Largest F ree F ragment
with a possible implementation in a control plane. Such EF = 1 − (2)
All F ree M emory
a metrics could be used to quantify the SF level to,
eventually, reduce its impact. Nevertheless, it does not The main limit of EF metrics is that it does not care
exist many metrics to measure the fragmentation is- about the small fragments if the size of the largest block
sue, and the existing ones suffer from some limits we does not change. Thus it does not prevent from creating
detail thereafter. To compare the relevance of the dif- small fragments. Furthermore, EF returned value is not
ferent metrics applied to optical networks, we compute always meaningful because this metrics is not able to
their complexity defined by the number of operations, determine if memory resources (e.g. spectrum by anal-
and determine a relative ”value meaning” mark with a ogy with optical networks) are completely fragmented
number of stars from 1 to 7. This secondary criterion or to identify the fragmentation that does not impact
shows to what extent a metrics returned value is mean- the fiber capacity.
ingful (Table 1). Mark attribution is done according to To evaluate EF ”value meaning”, we note that: it
whether the metrics is able to identify the following link is less complex than other metrics (g); when all free
cases: resources are contiguous EF equals 0(c), that clearly
identifies non fragmented states; the more the resources
are fragmented the more EF is close to 1 (f), but it
never reaches the value 1, that poorly identifies fully
(a) all slots are free (spectrum is not fragmented) fragmented states. Based on these considerations, we
(b) all slots are used (spectrum is not fragmented: block- attribute three stars to the ”value meaning” mark of
ing caused by lack of resources) EF (Table 1).
(c) all free slots are contiguous (spectrum is not frag-
mented)
(d) fragmented cases that do not reduce fiber capacity 3.2 Shannon Entropy fragmentation metrics (SE)
(spectrum is not fragmented)
(e) all free slots are lost (spectrum is completely frag- In information theory, the entropy estimates the infor-
mented: absolute loss) mation amount contained in a message. Based on this
(f) the more the spectrum is fragmented the greater is concept Shannon entropy as a quantitative metrics for
the returned value (relative fragmentation) SF measurements has been used in [15].
(g) non-complexity For every free fragment f if Sf denotes the number
of slots in f and S denotes the total number of slots
4 Djamel Amar et al.

in a fiber, SE estimates the fragmentation value as in

(3a).
Since the number of free slots is constant in all pro-
visioning possibilities to be compared and seeing that
S is a constant term, equation (3a) can be simplified to
(3b) which returns the same logical value when different
cases are compared, and allows reducing SE complexity
by half.
Like EF metrics, SE cannot conclude whether the Fig. 2 Example of fragmentation measured value by each
metrics for four different links (used slots in gray and free
spectrum is completely fragmented or not. The only slots in white) with 3 slots, 4 slots as possible granularities.
case in which SE value makes sense (a), is when all a) EF=0.00, SE=0.30, ABP=0.00
resources are free (SE equals 0). For instance when all b) EF=0.42, SE=0.72, ABP=0.00
used frequency slots are contiguous (spectrum is not c) EF=0.28, SE=0.66, ABP=0.33
d) EF=0.50, SE=0.64, ABP=1.00
fragmented) SE value is different from 0. However, SE
can identify relative fragmentation (f) and it is almost
as complex as EF (g). Therefore, we mark SE ”value may favor situations where some granularities cannot
meaning” with three stars (Table 1). be used. For instance, according to these metrics in
Fig. 2 c) spectrum is less fragmented than in Fig. 2 b)
whereas it is just the opposite, because in the former the
X Sf  
S
SE = ln (3a)
S Sf two-slot block can no longer be used, since it is smaller
f
  than the smallest granularity (3 in this example). ABP
X 1 metrics has thus the advantage to deal with available
SE ′ = Sf ln (3b)
Sf granularities. As we can see in Fig. 2 b) the fragmen-
f
tation value is zero because this kind of fragmentation
does not affect the fiber capacity. In Fig. 2 d) all free re-
3.3 Access Blocking Probability metrics (ABP) sources are absolutely lost, this information can simply
be read from ABP returned value.
We propose ABP to quantify the fragmentation. The However, ABP formula is more complex in terms
concept of ABP is as follows: based on the fact that of the number of operations compared with others as
blocking is strongly dependent on the transponders gran- shown in Table 1.
ularities being used, ABP calculates the ratio between
Now that we introduced fragmentation metrics and
the sum of cardinalities of the granularities set that can
made a global comparison we move to an accurate com-
be used in every fragment f i and the cardinality of the
parison in terms of fragmentation blocking probability
granularities set that can be used if all free slots are
and spectrum use in dynamic traffic scenario.
contiguous (4).
Pn Pm
(fi DIV Gk )
ABP = 1 − Pi=1 m Pk=1
n (4)
j=1 ( l=1 fl ) DIV Gj 3.4 System modeling
where DIV denotes the integer division and G is the
set of available granularities. In order to properly model the system, the designed
ABP metrics is able to determine meaningfully the model has to be independent from all metrics. It has
level of fragmentation: (e) ABP=1 when the spectrum also to consider only the blocking caused by SF, when it
is completely fragmented; (a)(c)(d) ABP=0 when the calculates the blocking probability as mentioned above.
spectrum is not fragmented, and (f) 0 <ABP< 1 when We adapt the model proposed in [9] to compare
spectrum is relatively fragmented. However, like EF the different metrics using a Continuous-Time Markov
metrics, ABP is not able to differentiate the link case Chain (CTMC). CTMC is a stochastic memoryless model
(a) from the link case (c). Based on these considera- (the transition from one state to another only depends
tions, ABP appears to be more relevant than the other on the current state and not on the previous states)
metrics and we attribute four stars to its ”value mean- with a limited state sequence, in which the holding time
ing” (Table 1). follows an exponential distribution.
Fig. 2 shows different values measured by each met- As said earlier, we are only interested in the frag-
rics for four ten-frequency slot links. Because EF and mentation caused by the gridless technology. Thus, the
SE do not take into account possible granularities, they system can well and truly be represented by one fiber
Spectrum fragmentation issue in flexible optical networks 5

Table 2 Demand types Table 3 Transition rate matrix

Statesa 0 1 2 3 4 5 6 7
Granularity (slotsa ) Bitrate (Gb/s)
0 −λ 0 0 0 λ 0 0 0
1 10 1 µ −λ − µ 0 0 0 λ 0 0
2 µ 0 −λ − µ 0 0 0 λ 0
2 40
3 0 µ µ −λ − 2µ 0 0 0 λ
3 100 4 µ 0 0 0 −λ − µ 0 λ 0

4 200 5 0 µ 0 0 µ −λ − 2µ 0 λ
6 0 0 µ 0 µ 0 −λ − 2µ λ
a 7 0 0 0 µ 0 µ µ −3µ
These demands types are only used with the dynamic
provisioning scenario and not in incremental scenario. a
For simplicity the state identifier in fig. 3 is converted from binary to decimal,
for example, the state in which all slots are used (111) is called here 7.

and its spectrum slots [17]1 . All possible instantaneous

Where π(π0 , π1 ...π2SL−1 ) is the steady-state vector for
fiber states (instantaneous spectrum occupancy) repre-
the matrix Q (πi is the probability that the system is
sent the CTMC state space (S).
in state i ∈ S).
Considering a dynamic provisioning scenario, de- To distinguish the blocking states that are due to
mand types in Table 2, are following a Poisson distri-
resource lack from the ones due to fragmentation, the
bution with a homogeneous arrival rate λ and an ex-
blocking probability Pgi for demand type Gg ∈ G in the
ponential holding time 1/µ. The load of the system is state i ∈ S can be calculated as in (6). This equation
thus λ/µ. counts the blocking only if there are free resources. For
Every metrics has its own CTMC depending on instance, in Fig. 3 there is no blocking caused by SF for
whether the transition between states is possible or not, the granularity (Gg = 1). The only blocking is caused
and therefore, there is a different steady-state vector (π) by resource lack when all slots are used.
for every metrics.

In order to illustrate the model, we show the state
X
 1 If ((Gg ≤
 fl ) and (Gg > largest(fl ))
diagram for a 3-slot fiber (SL=3) and only one possible Pgi = l
granularity of 1 slot (fig. 3). In this particular case the 
 0 Otherwise
state diagram is the same for all metrics.
(6)
Since demand departure is independent from the met-
rics, the dynamicity is considered as the most influential The total blocking probability in the system for a
parameter leading to SF in the investigated system. If demand type Gg ∈ G equals the sum of the probabili-
the dynamicity was not allowed, most of the transition
to fragmented states would have been canceled.
From the state diagram the so-called transition rate
matrix (Q) can be easily found (Table 3). The maxi-
mum number of states in the diagram is 2SL−1 . The
diagonal of the matrix Q must be filled in such a way
that the sum in each row is equal to zero.
Once the Q matrix is constructed, the linear sys-
tem (5) has to be solved. Equation (5a) is obtained
based on the stationary distribution property when time
tends to infinity (t → ∞). Equation (5b) constrains the
set of possible solutions for (5a), because the sum of all
steady-state probabilities must be equal to 1.

πQ = 0 (5a)
X
πi = 1 (5b)
Fig. 3 Example of diagram state for a 3-slot fiber and 1-
i∈S slot granularity. A used (free) slot is in gray (respectively
white). Transition from one state to another is represented by
1
In the case of a real network the metrics can be applied a directed arrow with an arrival or a service rate depending
on all the path calculating the sum of links corresponding on the event: arrival or departure. In this particular case, all
values. metrics have this diagram state in common.
6 Djamel Amar et al.

ties corresponding to states in which this demand type a more efficient provisioning strategy to reduce the SF,
is blocked because of SF (7). In other words, Pg is the instead of the ones mostly used in the traditional trans-
time fraction the system is in states where the gran- port networks. This is especially true when the network
ularity Gg cannot be used because of SF. Note that is highly loaded (the blocking probability of FF exceeds
Pg cannot be equal to 1, even if the granularity Gg is 30%), and thus when fragmentation is the most proba-
greater than the maximum number of possible slots in ble.
the fiber, simply because Pg concerns only the blocking Fig. 5 shows the theoretical average throughput the
caused by fragmentation. system can reach. The throughput is a relevant perfor-
The sum of blocking probabilities for every demand mance criterion because it is strongly affected by SF:
type (Gg ∈ G) is denoted P . This sum is used in the the more free frequency slots are contiguous in the sys-
following as an efficient criterion correlated with the tem, the more important is its theoretical capacity. In
blocking in the whole system (8): the fragmentation Fig. 5 the average throughput is decreasing with the
blocking probability. load increase. This is because when the load increases
X the system stays longer in states where the free spec-
Pg = Pgi πi (7) trum is either fragmented or null. Reversely, when there
i∈S
is no demand in the system, the theoretical average
X throughput is maximum: all resources can theoretically
P = Pg (8) be used.
Gg ∈G
ABP metrics is, again, the most efficient in terms of
Based on the same steady-state vector, equation (9) average throughput, but the curves in Fig. 4 and Fig. 5
calculates the average throughput (T h) in the whole have rather different behaviors. As described above, the
system for a given metrics, where nig , bg , µg denote average throughput is directly derived from the steady-
respectively the number of current demands in state state vector (8). It is then influenced at the same time
i ∈ S, the bitrate and the service rate corresponding by the SF and the lack of resources, which explain the
to the demand granularity Gg ∈ G. nig is calculated difference between curves in Fig. 4 and Fig. 5. Indeed
based on the maximum capacity the current state can in fig. 5 the gap between FF and the other curves is
sustain. smaller than in Fig. 4 and the gap between EF and SE
X X is wider. As FF policy is more fragmenting the spec-
Th = nig bg µg πi (9) trum (Fig. 4), it rejects more demands than the others
Gg ∈G i∈S
do, and consequently it is less blocked due to the lack
of resources. This gives to FF the opportunity to im-
3.5 Comparison results prove a little bit its behavior compared to Fig. 4. The
difference, between SE and EF is clearer for the same
Since the states space is exponentially proportional to reason for which FF policy curve has slightly changed
the number of slots by fiber (SL), numerical results have its behavior.
been obtained for a ten-slot fiber. However, the math- Globally, the gaps between the curves of Fig. 5 are
ematical aspect behind these results can, obviously, be explained by the ability of the schemes (FF or met-
generalized for more important instances. rics) to stay more or less time in the states where the
To show the interest of using a sophisticated metrics spectrum is more fragmented.
instead of a traditional allocation policy for spectrum In Fig. 6 metrics are compared in terms of the prob-
provisioning, we compare the widely used First Fit (FF) ability to be in the state where there is no free slot (all
policy with three metrics (EF, SE, and ABP). FF al- resources are used). It is an interesting parameter be-
ways allocates the first free slots to demands upon their cause it reflects the blocking ratio: the more a metrics
arrival. rejects demands (because of fragmentation), the less it
In Fig. 4 we compute the blocking probability in the is in this state. Fig. 6 shows that all curves are growing
whole system for FF, EF, SE and ABP versus the sys- with the load increase, as expected, with an advantage
tem load. ABP metrics is the best one in terms of block- of ABP to be more efficient in terms of spectrum uti-
ing probability whatever the load is. This is mainly lization.
due to the fact that it deals with available granulari- Please note that fragmentation impact will be more
ties and tries to avoid situations where these granulari- important for real instances, because the number of
ties are blocked. As expected, compared with FF policy, cases where blocking is related to fragmentation (not to
the three metrics are significantly reducing the block- resources lack) increases with the increase of the net-
ing probability, which confirms the necessity to apply work size.
Spectrum fragmentation issue in flexible optical networks 7

0.28 1
ABP ABP
0.26 SE SE
EF 0.8 EF
Blocking probability

Probability (10−3 )
FF FF
0.24
0.6
0.22
0.4
0.2

0.18 0.2

0.16 0
0.7 0.75 0.8 0.85 0.9 0.95 1 0.7 0.75 0.8 0.85 0.9 0.95 1
Load Load

Fig. 4 System blocking probability versus system load for Fig. 6 The probability to be in the state where all slots
ABP (Access Blocking Probability), SE (Shannon Entropy are used for different metrics: ABP (Access Blocking Proba-
Fragmentation), EF (External Fragmentation) metrics and bility), SE (Shannon Entropy Fragmentation), EF (External
FF (First Fit) provisioning strategy. Fragmentation) and FF (First Fit) provisioning policy versus
the load.

4 Operator network context

72
ABP
SE In operator core network, the traffic behavior is less
70
Average throughput (%)

EF fluctuating and demands do not vanish unlike in dy-

68
FF namic traffic scenario. Besides that, the spectrum re-
sources assignment follows a first-fit provisioning policy
66 and there are some operational constraints that might
make SF behavior different.
64 The traffic follows an incremental growth to repro-
duce successive discrete intervals of time (period) set
62 at 12 months. The maximum number of periods the
network can sustain before first blocking occurs is con-
60 sidered using several grid scenarios and traffic strategies
0.7 0.75 0.8 0.85 0.9 0.95 1 as detailed below. Traffic growth is assumed to be ex-
Load ponential with an annual growth rate α = 35%. Equa-
tion 10 calculates the traffic demand DPi at the period
Fig. 5 Average throughput percentage in the maximum
achievable throughput for different metrics: ABP (Access
i + 1 as a function of the same initial traffic demand
Blocking Probability), SE (Shannon Entropy Fragmentation), (DP0 ) in the first period.
EF (External Fragmentation) and FF (First Fit) provisioning
policy versus the load. DPi = α × (1 + α)i−1 × DP0 (10)

The network dimensioning is accomplished respecting

the following operational constraints:
So far, fragmentation metrics are compared with a – Demand blocking is not allowed. When demands can
FF provisioning policy in dynamic traffic scenario. Re- no longer be served the current period is not taken
sults show the interest of using a sophisticated metrics into account.
instead of a simple provisioning. Compared to other – Traffic demand at the period i + 1 can share re-
metrics, ABP reduces the total blocking probability in sources (transponders) with another one from the
the system, improves the average throughput and leads period i provided that they have the same source
to a good spectrum use. and the same destination.
In the following, ABP is used to evaluate SF and not – When several transponders are selected to serve the
to optimize spectrum provisioning process. In this way, same traffic demand, the corresponding physical path
SF is evaluated in real circumstances. must be the same.
8 Djamel Amar et al.

– Same logical topology is used for each period. 4.2 Traffic strategies
– Assigned demands of previous periods are kept un-
changed. To highlight traffic growth anticipation impact on both
gridless technology benefits and SF issue, following traf-
fic strategies are compared:
– Best Fit strategy (BF): This strategy chooses the set
of transponders that exactly fit the required bitrate.
For instance, this strategy selects either three 100
Gb/s transponders or one 100 Gb/s and one 200
4.1 Transponder model
Gb/s to serve a 300 Gb/s traffic demand. The choice
is depending on the objective function and available
An uncompensated Orange core network scenario de- paths.
ployed with 100 Gb/s and beyond is considered. More – Over-Dimensioning strategy(OD): This strategy over-
specifically, several possible transponders and SC types sizes the required capacity when it does not fit the
to allow different bitrate generation: 100, 200, 400 and 100 Gb/s, 200 Gb/s, 400 Gb/s, nor 1 Tb/s solu-
1Tb/s are used (Table 4). tion. For instance this strategy selects a 400 Gb/s
The granularity represents the 12.5 GHz frequency solution for a 300 Gb/s traffic demand. In this way,
slot number they occupy. The maximum transmission SCs interest and potential traffic anticipation bene-
reach is determined for each channel or SC using the fits are combined. Future traffic demands can then
Gaussian Noise model proposed in [8]. The optical trans- be assigned the additional unused capacity respect-
mission link is composed of 100 km spans length of SMF ing the rules detailed above.
fiber (span loss = 22dB) with Erbium Doped Fiber Note that unused transponders are not considered in
Amplifier (EDFA) with a Noise Figure of 6 dB. All the cost calculation even if the spectrum is reserved
transponders and SCs use a Soft Decision Forward Er- for them.
ror Correction (SDFEC).
The filtering penalty from pass through Reconfig- 4.3 Grid scenarios
urable Optical Add-Drop Multiplexers (ROADMs) is
modeled by an additional OSNR penalty, considering Combined with previous traffic strategies (BF and OD),
the presence of a ROADM every 3 spans. This penalty is we compare the following grid scenarios to identify the
calculated using the following assumptions: a ROADM impact of channel spacing and grid flexibility on both
penalty of 0.05 dB is allocated for 32 Gbaud (Gbd) SF level and gridless technology gained capacity:
transponders occupying 4 frequency slots, 0.64 dB [4]
– Fixed Grid without superchannels (FG): This sce-
for 32 Gbd transponders occupying 3 slots and 0.05 dB
nario allows only the 4-slot based transponders and
for the 16 Gbd transponder. The SCs take benefit from
will be used as a reference for other scenarios. When
reduced filtering penalties since the demultiplexing of
the FF provisioning policy is used, FG scenario does
SCs into individual sub-channels is not possible in in-
not create unsable blocks.
termediate nodes. Filtering penalties are thus limited
– Fixed Grid with SuperChannels (FGSC): To be com-
to the edges of SCs. By considering a spectrum guard-
pliant with the traditional fixed grid and to take ad-
band on each SCs sides equivalent to the one we have
vantage of SCs solutions, in FGSC scenario all SCs
for 32Gbd transponders occupying 4 frequency slots, we
that occupy a multiple of four slots are allowed.
have considered 0.05 dB OSNR penalty per ROADMs
– GridLess (GL): GL scenario allows using all type
for the SCs.
of transponders granularities without any restric-
The cost model is based on the reference cost fixed tion. In this way, the spectrum efficiency is improved
at 1 for the 100 Gb/s QPSK regardless of the allocated compared to previous scenarios (FG and FGSC).
slot number. The cost of SC is calculated by adding the However, GL scenario is more subjected to SF.
cost of each individual sub-channel. As a consequence
the cost of a partially equipped SC is calculated by
adding the cost of the sub-channels number really used
(even if the total number of slots of the entire SC is pre-
reserved to a single source destination). SCs based on
Photonic Integration Components supposed to reduce
SC overall cost are not considered in this study.
Spectrum fragmentation issue in flexible optical networks 9

Table 4 Characteristics of transponders

Bitrate(Gb/s) Modulation format Baudrate (Gbaud) Granularity (slot) Reach (km) Cost (a. u.)
100 16QAM SDFEC 16 3 400 0.7
Single channel

100 16QAM SDFEC 16 4 400 0.7

100 QPSK SDFEC 32 3 1400 1
100 QPSK SDFEC 32 4 2100 1
200 16QAM SDFEC 32 3 300 1
200 16QAM SDFEC 32 4 400 1

200 2×QPSK SDFEC 2 × 32 7 1900 2

Superchannel

400 2×16QAM SDFEC 2 × 32 7 400 2

400 4×QPSK SDFEC 4 × 32 13 1900 4
1000 (1 Tb/s) 5×16QAM SDFEC 5 × 32 16 400 5
1000 (1 Tb/s) 10×QPSK SDFEC 10 × 32 31 1900 10
1000 (1 Tb/s) 10×QPSK SDFEC 2 × 32 32 1900 10

4.4 RSA algorithm Algorithm 1 RSA algorithm

Require: number path > 0, number slots > 0
sort demands decreasing way
dimensioning ← true
period ← 0
while (dimensioning = true) do
The RSA np-hardness has been proven whatever the demands ← new demands
network topology is [14]. To solve RSA problem, we for d in demands do
served ← f alse
use a k-shortest path routing algorithm that uses a FF k←0
provisioning policy and calls the model (11) to opti- while ((served = f alse) and (k < number path)) do
mally select transponders and place regenerators (Al- k ←k+1
gorithm 1). Each traffic demand D is then served on the sl ← 0
while ((served = f alse) and (sl < number slots))
first shortest path possible selecting the set of transpon- do
ders that optimize in a hierarchical way both cost and sl ← sl + 1
spectrum usage. Wavelength regeneration process must M odel 11 (P A(k), d)
respect the constraint that all regenerators can only be if ((served = f alse) and (sl =
number slots) and (k = number path)) then
placed in an existing node and not anywhere else on the stop dimensoning
link. end if
end while
end while
period ← period + 1
end for
Model (11): Equation (11a) is the objective func- end while
tion minimizing the cost. When several solutions have
the same cost; it minimizes the spectrum occupancy
(ǫ is a small coefficient that weights this second objec- XXX XX
tive). Equations (11b) ensure the total traffic of every M in Xipk Ci + ǫ Xik Si (11a)
link of the path PA is equal to D. Equations (11c) en- i p k i k
XXX
sure that all sub-paths which are used by a transpon- Subject to Xipk Upe Di = D ∀e ∈ E (11b)
der of type i have no shared link. Inequalities (11d) i p k
mean that the length of a sub-path p must be less than
X
Xipk Upe = Xik ∀i ∈ I, ∀k ∈ Ki , ∀e ∈ E
the reach of corresponding used transponders. Finally, p
equations (11e) specify variable types (variables and (11c)
data are explained in Table 5).
Lp Xipk ≤ Ri ∀i ∈ I, ∀p ∈ P, ∀k ∈ Ki (11d)
Xik , Xipk ∈ {0, 1} (11e)
10 Djamel Amar et al.

Table 5 Data and variables

Data
PA(E, P) A path PA where E is the set of links e and P is the set of all possible sub-paths p
I Set of transponder types (Table 4)
D Demand size (Gb/s)
Di Bitrate of transponder type i (Gb/s)
Si Number of slots that a transponder of type i occupies
j k
D
Ki = D indicates the maximum number of transponders of type i that can be used
i

Ki to serve D, {1, 2, ...Ki } is the set of what is called here lines

Lp Length of sub-path p (km)
Ri Reach of transponder of type i (km)
Ci Cost of transponder of type i (a.u.)
Upe Equal to 1 if the sub-path p passes through the link e ; 0 otherwise

Variables
Xik Equal to 1 if the transponder of type i is used on the line k, 0 otherwise
Xipk Equal to 1 if the sub-path p is used by the transponder of type i on the line k, 0 otherwise

Table 6 Network topology less than 1400 km2 (see fig. 8). This means that the
spectrum gained capacity in this interval is closer to
Number of links 42 theoretical gridless technology gain (25%). Similarly,
Number of nodes 32
the fragmented spectrum should be small because of
the 3-slot granularity dominance. Thus, we conclude
Min 10 that the diversity of spectrum slot occupancy that leads
Link length (km)
Mean 280 to SF is strongly depending on the logical topology.
Max 930 Since it is more likely to have high SF when different
granularities are used, we can estimate that favoring a
logical topology in this way should lead to the worst
case of SF. We, thus, choose a logical topology such a
5 Results and discussion way that path lengths are mostly outside that interval.

Simulations are performed on a 32-node and 42-link Simulation results are averaged over thirty initial
typical transcontinental European backbone topology traffic matrices (set of DP0 ) normalized to 6 Tb/s and
(Table 6). At most three 360-slots-fiber pairs by link randomly drawn according to the logical topology. For
can be deployed. The logical topology is chosen accord- each combination (strategy scenario) simulation tool
ing to the following assumptions: computes network cost, gained spectrum, relative spec-
Independently of the physical topology, fig. 7 shows trum loss (using ABP), and absolute spectrum loss (us-
the granularities of selected transponders for each cou- ing the model (1)).
ple (demand, distance) when FGSC scenario is enabled. Due to the used strategy to calculate the cost of OD,
Most importantly, it shows that when the traffic vol- there is no significant difference between different com-
ume is less than 1 Tb/s FGSC scenario is equivalent to binations (not illustrated in a figure).
FG scenario (no saved spectrum) whatever the logical Using a 95% confidence interval, fig. 9 shows the
topology is and then the spectrum will not be frag- percentage of saved spectrum for different combinations
mented. between scenarios and strategies with respect to the ref-
erence best fitted conventional fixed grid scenario (BF
For the GL scenario, the 3-slot granularity widely
dominates the optimal choice when the path length is 2
Maximum reach of 3-slot transponders
 Spectrum fragmentation issue in flexible optical networks 11

20
2000

Saved spectrum (%)

15
Distance (km)

1500

4
1000 16 10
32

500 5

0
200 400 600 800 1000
1 2 3 4 5 6 7 8 9
Demand (Gb/s)
Periods
Fig. 7 Variation of used granularities in terms of traffic vol-
ume and required reach for FGSC scenario. BF GL BF FGSC
OD GL OD FGSC

Fig. 9 Saved spectrum slots for different scenarios/strategies

2000 with respect to conventional FG scenario using a 95% confi-
dence interval.
Distance (km)

1500
is still about 13% for GL scenario when traffic growth
3
4 is not anticipated. We explain this by observing that
1000
7 4-slot based transponders are at times selected due to
13 the chosen logical topology (earlier discussed in fig. 8).
500 16 Fig. 10 shows the relative fragmentation measured
31 by ABP metrics in best fitted GL scenario versus FGSC
scenario when SCs are favored through overdimension-
200 400 600 800 1000 ing4 . Due to the flexibility introduced by SCs and due
Demand (Gb/s)
to the significant differences between possible granu-
larities (4, 16, 32), the relative fragmentation exists in
Fig. 8 Variation of used granularities in terms of traffic vol- FGSC scenario and it is comparable to the one in GL.
ume and required reach for GL scenario.
However, a part of the saved spectrum (fig. 9) may
not be usable because of heterogeneous created blocks
FG). As expected, both GL and FGSC scenarios al- of slots. In order to evaluate this loss, we compute the
low saving the spectrum with more advantage for the number of unusable slots as defined in Section 3.
former. However, the traffic growth anticipation strat- Fig. 11 illustrates the percentage of unusable slots in
egy OD is more efficient with FGSC scenario3. This is the saved spectrum for different combinations between
mainly due to the fact that there are more possibilities scenarios and strategies. As expected, FGSC scenario
(in terms of transponders diversity) when GL scenario does not lose the spectrum since at the worst case 4-
is used that relatively reduces traffic growth anticipa- slot based transponders can unblock the situation. In
tion benefits. the GL scenario, traffic growth anticipation through
Remarkably, BF FGSC had to wait until the sixth pe- overdimensioning reduces this loss by half in the last
riod before saving the spectrum. This is because the period.
reference FG is equivalent to FGSC when traffic volume Note that the only way to recuperate this loss is
is less than 1 Tb/s (shown in fig. 7). In other words, in by spectrum defragmentation. However, this may not
FGSC scenario SCs are only used when traffic volume appear relevant, because the loss is very small: in BF
attains 1 Tb/s which impacts the saved spectrum. GL case this loss is still less than 6% out of 13% (i.e.
Even with a k-shortest path routing that is supposed < 1% of saved spectrum). Furthermore, this small loss
to most penalize the FG reference, the gained capacity means when some demands are blocked, switching to
3 4
The gap between OD FGSC and BF FGSC is greater than OD FGSC and BF GL are compared in terms of SF since
the one between BF GL and OD GL they are comparable in terms of saved spectrum.
12 Djamel Amar et al.

100 100
BF GL BF GL
OD FGSC OD GL
80 80

60 60
ABP (%)

ABP (%)
40 40

20 20

0 0
40 50 60 70 80 90 100 110 120 40 50 60 70 80 90 100 110 120
Fiber ID Fiber ID

Fig. 10 Relative fragmentation in FGSC when traffic growth Fig. 12 Traffic anticipation impact on relative fragmenta-
is anticipated versus BF GL (last period). Fibers are ordered tion in GL scenario (last period). Fibers are ordered according
according to FGSC values. to BF values.

6 100

5
80
Absolute loss (%)

4
60
ABP (%)

3
40
2

1 20

0
0
1 2 3 4 5 6 7 8 9
0 20 40 60 80 100
Periods
Fiber occupancy
BF GL BF FGSC Fig. 13 Correlation between relative fragmentation and link
OD GL OD FGSC occupancy in GL scenario (last period).

Fig. 11 Percentage of unusable saved spectrum in the saved

spectrum shown in fig. 9 for different scenarios/strategies us-
ing a 95% confidence interval. cupancy at the last period represented by a high cor-
relation coefficient equal to 0.93. Since OD strategy re-
duces links occupancy (because it saves the spectrum),
less large granularities would often unblock the blocked this strong relationship is then another reason for which
bandwidth. traffic growth anticipation reduces SF. In addition, this
Another advantage of traffic growth anticipation rep- strong relationship means that load balancing between
resented by OD strategy is its impact on relative frag- links will make SF lower.
mentation (Fig. 12). Relative fragmentation level is re- Traffic growth anticipation is then a successful strat-
duced on every fiber in the network. This is due to egy to save the spectrum and reduce both absolute and
SCs that are more likely to be chosen when traffic is relative losses caused by SF. However, when significant
anticipated. Indeed, SCs impose the contiguity of its errors are considered in the growth anticipation ap-
subchannels that leads to free spectrum contiguity and proach these benefits could be reduced. Similarly, some
then reduces SF. simple approaches like load balancing and the use of
For best fitted GL scenario, fig. 13 shows the strong minor granularities are expected to reduce the residual
relationship between relative fragmentation and link oc- fragmentation.
Spectrum fragmentation issue in flexible optical networks 13

6 Conclusions References

1. Spectral grids for WDM applications: DWDM frequency

In this paper, we evaluated spectrum fragmentation
grid, ITU-T G.694.1
(SF) problem in both dynamic and incremental traf- 2. Acharya, S., Gupta, B., Risbood, P., Srivastava, A.: Mo-
fic scenarios and proposed some simple approaches to bipack: optimal hitless sonet defragmentation in near-
face this problem in the context of operator network. optimal cost. In: INFOCOM 2004. Twenty-third Annu-
alJoint Conference of the IEEE Computer and Commu-
In dynamic traffic scenario, several fragmentation nications Societies, vol. 3, pp. 1819–1829 vol.3 (2004)
metrics are compared using a Markov modeling ap- 3. Chao, C.M., Wang, S.H.: A joint code/time assignment
proach. Numerical results show the efficiency of using strategy with minimal fragmentations for cdma systems.
a fragmentation metrics instead of a traditional alloca- In: Wireless Communications and Networking Confer-
ence, 2006. WCNC 2006. IEEE, vol. 3, pp. 1620–1625
tion policy like the first-fit provisioning. (2006)
Shannon Entropy based metrics (SE) and External 4. Eira, A., Pedro, J., Pires, J.: On the impact of optimized
Fragmentation metrics (EF) already offer good perfor- guard-band assignment for superchannels in flexible-grid
optical networks. In: Optical Fiber Communication Con-
mances. However none of them consider transponder ference and Exposition and the National Fiber Optic En-
granularity and both of them are more interested in gineers Conference (OFC/NFOEC), 2013, pp. 1–3 (2013)
leaving large free blocks even if inconvenient blocks 5. Johnstone, M.S., Wilson, P.R.: The memory fragmenta-
of slots are simultaneously created. We have proposed tion problem: solved? (1998)
6. Pages, A., Perello, J., Spadaro, S.: Lightpath fragmenta-
Access Blocking Probability (ABP) and demonstrated tion for efficient spectrum utilization in dynamic elastic
that it performs well compared to other metrics. This optical networks. In: Optical Network Design and Mod-
is because it takes into account available transponders eling (ONDM), 2012 16th International Conference on,
and avoids creating blocks that do not match with their pp. 1–6 (2012)
7. Pages, A., Perello, J., Spadaro, S., Comellas, J.: Opti-
granularities. mal route, spectrum, and modulation level assignment in
Despite the advantages of these metrics to evaluate split-spectrum-enabled dynamic elastic optical networks.
and/or reduce fragmentation, they add more complex- Optical Communications and Networking, IEEE/OSA
Journal of 6(2), 114–126 (2014)
ity to the control plane and may be difficult to apply 8. Poggiolini, P.: The gn model of non-linear propagation
in highly dynamic large networks due to large compu- in uncompensated coherent optical systems. Lightwave
tation time. Technology, Journal of 30(24), 3857–3879 (2012)
9. Rosa, A., Cavdar, C., Carvalho, S., Costa, J., Wosinska,
In incremental traffic scenario we evaluated spec-
L.: Spectrum allocation policy modeling for elastic op-
trum fragmentation (SF) in the context of operator tical networks. In: High Capacity Optical Networks and
network. We have shown that fragmentation is not only Enabling Technologies (HONET), 2012 9th International
due to the traffic behavior, but also strongly dependent Conference on, pp. 242–246 (2012)
10. Sone, Y., Hirano, A., Kadohata, A., Jinno, M., Ishida,
on the transponders granularity diversity, on the logical
O.: Routing and spectrum assignment algorithm maxi-
topology and on the network load. mizes spectrum utilization in optical networks. In: Opti-
Numerical results indicate that SF is a real effect cal Communication (ECOC), 2011 37th European Con-
ference and Exhibition on, pp. 1–3 (2011)
but does not compromise gridless technology benefits
11. Velasco, L., Ruiz, M., Castro, A., Pedrola, O.,
in incremental traffic scenarios. Klinkowski, M., Careglio, D., Comellas, J.: On the perfor-
Specifically, as the absolute loss is negligible and see- mance of flexgrid-based optical networks. In: Transpar-
ent Optical Networks (ICTON), 2012 14th International
ing that the higher values of relative fragmentation cor-
Conference on, pp. 1–4 (2012)
respond to links close to full capacity, we conclude that 12. Wang, R., Mukherjee, B.: Provisioning in elastic opti-
periodical defragmentation processes are not needed in cal networks with non-disruptive defragmentation. Light-
the context of operator network. wave Technology, Journal of 31(15), 2491–2500 (2013)
13. Wang, R., Mukherjee, B.: Spectrum management in het-
Instead, spectrum fragmentation (SF) can be better erogeneous bandwidth optical networks. Optical Switch-
addressed during network planning phase. For example, ing and Networking 11, Part A(0), 83 – 91 (2014)
using a specific fragmentation metrics for resources pro- 14. Wang, Y., Cao, X., Pan, Y.: A study of the routing
visioning, anticipating traffic growth, balancing traffic and spectrum allocation in spectrum-sliced elastic optical
path networks. In: INFOCOM, 2011 Proceedings IEEE,
between links, and/or switching to minor granularities pp. 1503–1511 (2011)
when the blocking occurs would be good strategies to 15. Wright, P., Parker, M., Lord, A.: Simulation results of
face such problem. shannon entropy based flexgrid routing and spectrum as-
signment on a real network topology. In: Optical Commu-
nication (ECOC 2013), 39th European Conference and
Acknowledgements This work was partly supported by Exhibition on, pp. 1–3 (2013)
the DGCIS, in the frame of the CELTIC-Plus project SASER- 16. Yin, Y., Zhang, M., Zhu, Z., Y, S.: Fragmentation-aware
SIEGFRIED. routing, modulation and spectrum assignment algorithms
14 Djamel Amar et al.

in elastic optical networks. In: Optical Fiber Communica-

tion Conference and Exposition and the National Fiber
Optic Engineers Conference (OFC/NFOEC), 2013, pp.
1–3 (2013)
17. Yu, Y., Zhang, J., Zhao, Y., Cao, X., Lin, X., Gu, W.:
The first single-link exact model for performance analysis
of flexible grid wdm networks. In: Optical Fiber Com-
munication Conference and Exposition and the National
Fiber Optic Engineers Conference (OFC/NFOEC), 2013,
pp. 1–3 (2013)