Professional Documents
Culture Documents
jocn-10-3-195
jocn-10-3-195
jocn-10-3-195
net/publication/323319404
CITATIONS READS
88 852
3 authors:
Qian Wu
Hosei University
13 PUBLICATIONS 165 CITATIONS
SEE PROFILE
All content following this page was uploaded by Mingcong Yang on 24 October 2018.
Abstract—In this paper, we focus on the static routing, recently been proposed, and it is considered to be a prom-
spectrum, and core assignment (RSCA) problem in space-
ising technology for overcoming the physical limitations
division multiplexing (SDM)-based elastic optical networks
(EONs) with multi-core fiber (MCF). In RSCA problems, it is of SMF by using multi-core fiber (MCF), multi-element fi-
a challenging task to control the inter-core interference, ber (MEF), multi-/few-mode fiber (MMF/FMF), and their
called inter-core crosstalk (XT), within an acceptable level combinations [4]. However, some new challenges, such as
and simultaneously maximize the spectrum utilization. We network planning and resource optimization, should be ad-
first consider XT in a worst interference scenario (i.e., XT-
unaware), which can simplify the RSCA problem. In this dressed. A typical problem of resource assignment is the
scenario, we formulate the RSCA problem using a node- routing, spectrum, and core assignment (RSCA) problem
arc-based integer linear programming (ILP) method in in space-division multiplexing elastic optical networks
which the numbers of both variables and constraints are (SDM-EONs) with MCF. In contrast to the routing and
greatly reduced compared with previous ILP methods,
thereby leading to a significant improvement in conver- spectrum assignment (RSA) problem in EONs with SMF,
gence efficiency. Then, we consider the XT strictly (i.e., SDM-EONs with MCF have introduced some new features
XT-aware) and formulate the problem using a mixed integer that make the RSCA problem more challenging.
linear programming (MILP) method, which is an extension
of the above node-arc-based ILP method. It is more suitable In the traditional WDM systems with SMF, several paral-
for different XT thresholds and/or geographically large net- lel WDM systems can increase system capacity, but they also
works, in that it has a higher degree of generalizability. increase system cost and energy consumption, leading to the
Finally, we propose an XT-aware-based heuristic algorithm.
The simulation results demonstrate that our heuristic algo-
result that the cost and energy consumption remain the
rithm achieves higher spectrum efficiency, higher degree of same per bit [5]. However, to satisfy the potential increasing
generalizability, and higher computational efficiency than demands for communications capacity, transmission equip-
the existing heuristic algorithm(s). ment is asked to reduce both energy and cost per bit.
Compared with the parallel WDM systems, component inte-
Index Terms—ILP/MILP optimization model; Multi-core
fiber; Routing, spectrum and core assignment; Space gration is the crucial feature of SDM systems, which can
division multiplexing; XT-aware. significantly reduce system cost and energy consumption.
In SDM systems, integrated components [5,6] in the spatial
domain, including transponders, optical amplifiers, and net-
I. INTRODUCTION working elements such as optical switches, can share system
components among spatial channels to reduce both capital
(CAPEX) and operational (OPEX) expenditures.
W ith the development and increasing popularity of
cloud computing, video-on-demand (VoD), internet
of things (IoT), and other emerging Internet services, net-
Another feature of SDM-EONs with MCF is core switch-
ing. Three different SDM switching strategies have been
work traffic is growing at an extremely rapid rate [1]. proposed [7–9]: 1) independent switching (InS), where all
However, the growth in the transmission capacity of stan- frequency slots (FSs) and spatial modes can be independ-
dard single-mode fiber (SMF) has dramatically slowed ently directed to any output port; 2) joint switching (JoS),
because the transmission capacity per fiber is close to where all spatial modes/cores are treated as a single entity,
the nonlinear Shannon limit of the existing SMF [2,3]. while FSs can be freely switched by the wavelength-
To satisfy the strongly increasing future traffic demand,
selective switch; and 3) fractional JoS (FJoS), where a few
the space-division multiplexing (SDM) technique has
groups of spatial modes and all FSs can be independently
switched to all output ports. The core switching can miti-
Manuscript received November 13, 2017; revised December 23, 2017; gate the constraint of spectrum continuity compared to
accepted December 28, 2017; published February 21, 2018 (Doc. ID 313350). RSA since there are multiple cores that can be used on a
The authors are with the Graduate School of Systems and Information
Engineering, University of Tsukuba, Tsukuba, Japan (e-mail:
fiber link. In other words, a lightpath from a source to
ymc6642664@gmail.com). its destination can have a larger routing space. In this pa-
https://doi.org/10.1364/JOCN.10.000195 per, the InS strategy is considered, because it is the most
6th FS has been assigned to R:4 and their cores are adja-
In Eq. (2), the XT between two adjacent cores XTcc0 is cal-
cent, there is an occurrence of weak XT for R:2 and R:6 on culated using Eq. (1), where Ac represents the set of active
the 6th FS. Moreover, because the XTs are affecting each adjacent cores of core c; in other words, for each core c0 ∈ Ac ,
other on the adjacent cores, there is an occurrence of strong the same f th FS on core c0 has been assigned to serve the
XT for R:4 on the 6th FS because it is affected by R:2 and other transmission requests. Generally, the parameters h,
R:6 at the same time. For a transmission request, only m, and L are fixed for a given MCF. Thus, the XT of core c on
when the XT on each assigned FS for the transmission re- the f th FS depends only on Ac . Therefore, in addition to the
quest is less than an acceptable XT threshold, the signals development of MCF on the physical layer, we can also
can be successfully transmitted on these assigned FSs. reduce the XT by avoiding spectrum overlaps between ad-
To minimize the XT and achieve a dense core arrange- jacent cores, that is, reducing the number of elements in Ac .
ment to enable high-capacity and long-distance transmis-
sions, how to suppress the XT has become a primary focus
in MCF research [13–19]. Takenaga et al. developed a A. Related Works
trench-assisted multi-core fiber (TA-MCF) in [19]. In TA-
MCF, XT is directly related to the mode-coupling In contrast to the RSA problem with SMF, which should
coefficient, which can be estimated without the need for satisfy the constraints of spectrum continuity, contiguity,
numerical simulations. and non-overlapping, the XT constraint should be taken
According to the equations proposed in Ref. [20], the XT into account in the RSCA problem with MCF when as-
between two adjacent cores c and c0 can be obtained by the signed the spectral resources, i.e., FSs. There are mainly
following equation: three approaches, as shown in Fig. 2, when considering
the XT constraint.
The first approach, called XT-avoid, is to avoid spectrum
overlaps between adjacent cores in the fiber links while al-
locating FSs and cores to different transmission requests
[23]. As shown in Fig. 2(a), the same FSs cannot be as-
signed to different transmission requests that traverse
through adjacent cores in the common links. The second
approach, called XT-WC, is to consider the XT in a worst
interference scenario [18,21]. As shown in Fig. 2(b), the
blank FSs are the empty FSs (i.e., have not been assigned
to any transmission request), but for the XT-WC approach,
they are considered to be occupied to evaluate the XT. The
third approach, called XT-aware, is to compute the XT
strictly depending on the interference of a core with other
active adjacent cores that share the same FS and link
[3,10–12], as shown in Fig. 2(c).
Both the XT-avoid and the XT-WC approaches can
Fig. 1. Example of inter-core crosstalk in a 7-core MCF. simplify the RSCA problem and significantly reduce the
Yang et al. VOL. 10, NO. 3/MARCH 2018/J. OPT. COMMUN. NETW. 197
shorter execution time compared with the previous R: Transmission request matrix, r fsr; dr; λr g ∈ R,
arc-path-based ILP methods. where sr and dr are the starting and ending
Second, we introduce a MILP optimization model consid- nodes of r, respectively, and λr is the required
ering the XT strictly based on the XT-aware approach. In FSs of r.
contrast to the XT-WC-based ILP methods, XT is calculated Li : Set of links that start or end at node i.
depending on the actual interference between cores on a xtecc0 : The XT between core c and core c0 on link e based on
common link and therefore ensures a complete feasible Eq. (1) by setting L equal to the length of link e.
routing path space for each transmission request; conse- Θm : Inter-core crosstalk threshold of the given
quently, a strictly optimal solution can be obtained by modulation format m.
our MILP optimization model. F: Set ofPtotal required FSs (in the worst case,
jFj r∈R λr ).
Furthermore, the static RSCA problem is an extension of M: A large number.
the static RSA problem, which is known to be NP-hard [25].
This problem is difficult to solve in an acceptable execution 2) Variables:
time by the ILP/MILP methods for large-scale problem
instances. For this reason, we propose an XT-aware-based
r ∈ f0; 1g:
lec A binary variable that is equal to 1 if core c
heuristic algorithm for achieving scalability. In our heuristic on link e is selected to serve transmission
algorithm, a strict XT-affected FS check mechanism is pro- request r and is otherwise equal to 0.
posed to achieve the XT-aware approach. The simulation ex- nir ∈ f0; 1g: A binary variable that is equal to 1 if the
periments show that our proposed heuristic algorithm can lightpath that is used to serve transmission
obtain solutions that are quite close to the optimal solutions request r passes through node i and is
compared to the previous heuristic algorithms. To the best of otherwise equal to 0.
our knowledge, our work is the first work that consider the αr ∈ Z : An integer variable that denotes the starting
XT-aware approach for the static RSCA problem. FS index of the lightpath that is used to serve
transmission request r.
C. Organization of This Paper βr ∈ Z : An integer variable that denotes the ending
FS index of the lightpath that is used to serve
transmission request r.
The remainder of this paper is organized as follows. In 0
φrr ∈ f0; 1g: A binary variable that takes a value of 0 if
Section II, we present our node-arc-based ILP/MILP opti- the ending FS index βr0 of the lightpath that
mization models for the static RSCA problem based on the is used to serve transmission request r0 is
XT-WC and XT-aware approaches. In Section III, we smaller than the starting FS index αr of
propose a novel heuristic algorithm based on the XT-aware the lightpath that is used to serve transmis-
approach. In Section VI, we present the results and sion request r.
evaluate the performance. Section V concludes the paper. F max ∈ Z : An integer variable that indicates the
number of maximum required FSs for the
II. PROBLEM FORMULATION network.
3) Objective Function:
In this section, we first present a node-arc-based ILP
optimization model based on the XT-WC approach. Then,
Minimize F max : (3)
we present an extended (node-arc-based) MILP optimiza-
tion model based on the XT-aware approach. Finally, we 4) Constraints:
compare our optimization models with the existing arc-
a) Cost Constraint:
path-based ILP optimization models. The spectral super-
channel policy [27,28], which is created by allocating βr ≤ F max ∀ r ∈ R: (4)
sub-channels in contiguous FSs on just one of the available
fibers/cores, is considered in this paper. Additionally, we as- The constraint in Eq. (4) ensures that the maximum FS
sume that the modulation format m used for transmission index F max is not smaller than the ending FS index of the
is given beforehand. lightpath for any transmission request r.
b) Route Selection Constraints:
A. Node-Arc-Based ILP Optimization Model Based X
r ≤ 1 ∀ r ∈ R; e ∈ E;
lec (5)
on XT-WC Approach c∈C
X X
1) Parameters:
r 1 ∀ r ∈ R;
lec (6)
c∈C e∈Lsr
V: Set of network nodes.
E: Set of network links. X X
C: Set of cores on each MCF. r 1
lec ∀ r ∈ R; (7)
Nc: Set of adjacent cores of core c. c∈C e∈Ldr
Yang et al. VOL. 10, NO. 3/MARCH 2018/J. OPT. COMMUN. NETW. 199
from node sr. Thus, the sum must be equal to one. The FS of core c on link e is assigned to serve transmission
constraint in Eq. (7) ensures that for a transmission re- request r and is otherwise equal to 0.
quest r, the last link of a lightpath ends at node dr. tef
r : A continuous variable that indicates the XT on link e
Thus, the sum must be equal to one. The constraint in at the f th FS for transmission request r.
Eq. (8) ensures that for any intermediate node traversed
2) Constraints:
by a lightpath, there is one link that should end at the node
and one link that should start from the node. The con- X ecf
or − λr lec
r 0 ∀ r ∈ R; e ∈ E; c ∈ C; (14)
straint in Eq. (9) ensures that if the lightpath for a trans-
f ∈F
mission request r traverses a link e:i; j, it must also
traverse nodes i and j. X
c) Spectrum Assignment Constraints: jFj oecf
r − 1 ≤ f − αr ∀ r ∈ R; e ∈ E; f ∈ F; (15)
c∈C
0
φrr φrr0 1 ∀ r; r0 ∈ R (10) X
jFj oecf
r − 1 ≤ βr − f ∀ r ∈ R; e ∈ E; f ∈ F: (16)
c∈C
d) Crosstalk Constraints:
The constraint in Eq. (17) ensures that if the f th FS of
XX core c on link e is assigned to serve transmission request r
r ≤ Θm
xtecc0 · lec ∀ r ∈ R; c ∈ C: (13)
(oecf
r 1), then the XT of transmission request r on link e at
c0 ∈N c e∈E
the f th FS (tef r ) should be greater than the sum of the
XT due to the active adjacent cores on the common link
The constraint in Eq. (13) ensures that for each trans- e that share the same f [based on Eq. (2)]. For example,
mission request r, the XT of the lightpath that is selected when
P oecf
Pr 1, the constraint in Eq. (17) becomes
0
e ec f ef 0
by the route selection constraints in Eqs. (5)–(9) should be c0 ∈N c r0 ∈R xtcc0 or0 ≤ tr . If an active adjacent core c of
less than the XT threshold where, based on the XT-WC core c (in N c ) on the common link e is excited to serve an-
0
approach, the worst interference scenario is considered; other transmission request r0 (orec0 f 1), then the XT be-
that is, we set Ac [see Eq. (2)] equal to N c . tween these two cores (xtcc0 ) should be summed. Then, tef
e
r
200 J. OPT. COMMUN. NETW./VOL. 10, NO. 3/MARCH 2018 Yang et al.
should be greater than the sum. In another case, when For our ILP optimization model (Section II.A), the dom-
oecf
r 0, the constraint in Eq. (17) is deactivated due to inant numbers of variables and constraints are OjRj2
0
the larger value jFj on the right-hand side of the constraint. (according to φrr ) and OjRj2 · jEj [according to the con-
In other words, tef
r can be 0 in this case. The constraint in straint in Eq. (11)], respectively, where jRj and jEj re-
Eq. (18) ensures that for each transmission request r, the present the total numbers of transmission requests and
XT of the selected lightpath should be less than the XT links in the network, respectively.
threshold on each assigned FS. Overall, according to the For the ILP optimization model in Ref. [21], the number
constraints above, we have incorporated the XT-limit con- of variables is bounded by OjRj2 · jEj, while the number of
straint into the MILP formulation based on the strict constraints is bounded by the larger of OjRj2 · jEj and
XT-aware approach. OjRj · jPj, where jPj represents the number of pre-
calculated candidate routing paths per transmission
request. When all the possible routing paths are consid-
C. Analysis of our ILP/MILP Models
ered, the ILP method in Ref. [21] can safely ensure finding
the optimal solution, but an upper bound for jPj can be up
In this subsection, we summarize our proposed node- to O2jNj [24]. Meanwhile, when only the k-shortest paths
arc-based ILP/MILP optimization models and compare are considered, an optimal solution may not be found, but
them with the previous arc-path-based ILP optimization the jPj can be significantly reduced to k. For the ILP opti-
models [18,21]. The major differences are stated as follows. mization model in Ref. [18], the numbers of both variables
First, our proposed ILP/MILP methods, which belong to and constraints are bounded by OjRj2 · jPj2 .
the node-arc category, decide the optimal routes and allo- Finally, for our MILP optimization model (Section II.B),
cate FSs and cores for all the transmission requests simul- the dominant number of variables is the larger one be-
0
taneously, whereas the ILP methods in Refs. [18,21], which tween OjRj2 (according to φrr ) and OjRj · jEj · jFj (accord-
belong to the arc-path category, only allocate FSs for the ecf
ing to or ), and the dominant number of constraints is the
lightpath services but do not find routes for the lightpath larger one between OjRj2 · jEj [according to the constraint
services. As a result, several candidate paths between each in Eq. (11)] and OjRj · jEj · jFj [according to the constraint
source–destination pair should be pre-calculated as the in- in Eq. (17)], where jFj is the total number of required FSs,
put parameters of the ILP optimization models. If the XT which is the objective of the optimization model. Because it
constraint is not taken into account, the optimal solutions is difficult to find an exact number for jFj before the opti-
can be obtained using our node-arc-based ILP method. In mization model isP solved, we set it as its upper bound,
contrast, only when all the routing paths for each transmis- which is equal to r∈R λr , where λr is the traffic demand
sion request are considered is it guaranteed that the units in FS for transmission request r. Here, we set λ as
optimal solutions can be obtained using the existing the average number of traffic demand units in FS for
arc-path-based ILP methods in Refs. [18,21]. If only a each transmission request; thus, we can derive the upper
few (e.g., the k-shortest) pre-calculated candidate paths bound for the dominant numbers of both variables and
are considered for each transmission request, a very effi- constraints as OjRj2 · jEj · λ.
cient or even optimal solution may also be found, but we Overall, we can observe that our ILP optimization model
cannot claim that the current solutions are the optimal sol- has considerably smaller numbers of variables and
utions if all the routing paths for each transmission request constraints. However, in our MILP optimization model,
are not considered in the optimization process. to achieve the XT-aware approach, we need to mark the sta-
Second, the previous ILP methods [18,21] consider the tus of each FS on each core and link, which inevitably in-
XT in a worst interference scenario (XT-WC), and the fea- troduces more variables and constraints. We summarize
sible lightpath space may become smaller (see Section I.A) the convergence efficiencies of the four optimization models
for a strict XT threshold, particularly in geographically in Table I.
large networks. In an extreme case, some transmission re- We also present the total numbers of columns (variables)
quests may not be satisfied, since in a worst interference and rows (constraints) given by the Gurobi ILP solver [29]
scenario, the XT may be greater than the XT threshold, when solving the optimization models for a six-node nine-
even for the shortest path of these transmission requests. link simple network [21] [Fig. 4(a) in Section IV] in Table II,
In contrast, our proposed MILP method considers the XT where jRj is set to be 15 (a transmission request between
strictly (i.e., XT-aware approach). Therefore, the optimal each node pair), and λr is set as a random number from 1–3
solution can be obtained. (FSs). The number of cores of a multi-core fiber jCj is set to
Next, we examine the convergence efficiency of the optimi-
zation models by counting the numbers of variables and con- TABLE I
straints of the optimization models. Note that in some cases, COMPARISON AMONG FOUR OPTIMIZATION MODELS
adding constraints in an optimization model will not reduce Model # Variables # Constraints
but rather improve the convergence speed of the model.
However, the numbers of variables and constraints are still Our ILP model OjRj 2
OjRj2 · jEj
Our MILP model OjRj2 · jEj · λ OjRj2 · jEj · λ
a metric that is worth considering [25]. For all the optimiza-
ILP model in Ref. [21] OjRj2 · jEj OjRj2 · jEj or OjRj · jPj
tion models, we assume that the number of cores of a multi-
ILP model in Ref. [18] OjRj2 · jPj2 OjRj2 · jPj2
core fiber, denoted by jCj, is equal to an invariant constant.
Yang et al. VOL. 10, NO. 3/MARCH 2018/J. OPT. COMMUN. NETW. 201
8
TABLE II >
>
> ar
<
If FS f on core c of link e
COMPARISON AMONG FOUR OPTIMIZATION MODELS FOR THE N6S9
NETWORK (BY GUROBI SOLVER) Ue; c; f has been assigned to request ar :
>
>
>
: None Otherwise
Model Columns Rows Non-Zeros
Our ILP model (Section II.A) 736 6,285 31,185
Our MILP model (Section II.B) 35,836 33,823 449,578 • A table of the current XT of previously assigned trans-
ILP Model in Ref. [21] (All paths) 15,623 23,239 95,490 mission requests ar on their assigned FSs, which is
ILP Model in Ref. [21] (jPj 5) 14,671 23,205 94,026 denoted as follows:
ILP Model in Ref. [21] (jPj 3) 13,831 23,175 92,796 8
ILP Model in Ref. [18] (All paths) 44,674 99,560 375,536 >
> The current XT of ar on FS f
<
ILP Model in Ref. [18] (jPj 5) 21,586 47,850 180,240
XTar; f : If f is assigned to ar :
ILP Model in Ref. [18] (jPj 3) 8,026 17,550 65,820 >
>
:
None Otherwise
be 3. To decide the numbers of variables and constraints The pseudocode for our strict XT-affected FS check
of our MILP optimization model, we considered the mechanism is shown in Algorithm 1. The inputs of the al-
worst
P case for jFj; in other words, we set jFj equal to gorithm are the pre-assigned routing path p ∪︀e; c con-
λ
r∈R . Three different cases of pre-calculated candidate
r taining links and cores along the path and the pre-assigned
paths per node pair P are considered, which represent FSs, that is, the starting FS f s and the ending FS f e. The
all the routing paths (for each transmission request) in output of the algorithm is a Boolean number, where 1 rep-
the n6s9 network topology, 5-shortest paths, and 3-shortest resents that the transmission request can be transmitted
paths. successfully on these pre-assigned FSs and 0 otherwise.
As shown in Table II, our proposed ILP optimization
model (Section II.A) has significantly fewer rows and Algorithm 1: Strict XT-Affected Frequency Slot Check
columns than previous ILP optimization models, even Mechanism
though we have bound the number of the pre-calculated Input: p ∪︀e; c, f s , f e
candidate paths jPj equal to 3. These results are consistent Output: 1 or 0
with the analytical results shown in Table I. 1: for frequency slot f in range f s ; f e do
2: Set XT self
f equal to 0 as the XT of the new transmis-
sion request itself on f
III. HEURISTIC ALGORITHM FOR RSCA PROBLEM 3: for (link e, core c) in routing path p do
4: for each adjacent core c0 of core c do
In this section, we present our XT-aware heuristic algo- 5: Set ar equals to Ue; c0 ; f , that is, the previously
rithm, which contains two components: a strict XT-affected assigned transmission request that occupies the
FS check mechanism and a RSCA component. f on core c0 of link e
6: if ar ≠ None then
7: The current XT of previously assigned
A. Strict XT-Affected Frequency Slot Check transmission request
Mechanism XT ar;f Δxtc;c0 ;e
XT self 0
8: f Δxtc; c ; e
9: if XTar; f > XT -threshold then
Our strict XT-aware FS check mechanism works as fol-
10: return 0
lows. When a new transmission request arrives, we should
11: end if
check the increased XTs of the previously assigned
12: end if
transmission requests to ensure that they are still less
13: end for
than the XT threshold, because the XTs of previously as-
14: end for
signed transmission requests will increase if the same
15: if XT self
f > XT -threshold then
FS(s) in an adjacent core are assigned to the new transmis-
16: return 0
sion request. Moreover, since XT between two adjacent
17: end if
cores is mutually affected, we should also ensure that 18: end for
the XTs on each assigned FS of the new transmission 19: return 1
request itself are less than the XT threshold, which are
related to the number of active adjacent cores on the same We present a simple example in Fig. 3. For simplicity, we
FS (used to serve the previously assigned transmission consider a simple chain network, where each link has a
requests). 3-core MCF. We assume that there are some FSs already
We use the following two tables in the strict XT check successfully assigned to other transmission requests ar
mechanism. (shown in green), that is, for each ar, the XTs on their as-
signed FSs (i.e., XTar; f ) are less than the XT threshold.
• A usage table represents the FS usage per core and link, For instance, we can observe that ar2 shares the same
which is denoted as follows: FSs (i.e., f 1 and f 2 ) with ar1 and ar3 on a common link
202 J. OPT. COMMUN. NETW./VOL. 10, NO. 3/MARCH 2018 Yang et al.
e12 ; therefore, the XT of ar2 on f 1 (i.e., XTar2 ; f 1 ) can be successfully transmitted. As can be seen in Fig. 3, the
calculated by Eq. (2), which equals the sum of XT of the new transmission request itself on the f 3 (XTself f3 )
Δxtc1 ; c2 ; e12 [see Eq. (1), where L equals the length of link is the sum of Δxtc1 ; c2 ; e12 , Δxtc2 ; c3 ; e12 , Δxtc1 ; c2 ; e23 ,
e12 ] and Δxtc2 ; c3 ; e12 . In the same way, XTar2 ; f 2 can also and Δxtc2 ; c3 ; e23 . Using the same approach, the XTs on
be calculated as Δxtc1 ; c2 ; e12 Δxtc2 ; c3 ; e12 . As we have the f 4 , f 5 , and f 6 can be calculated to be Δxtc1 ; c2 ; e12 ,
mentioned above, XTar2 ; f 1 and XTar2 ; f 2 are assumed 0, and 0, respectively. Then, we should check these
to be less than the XT threshold to ensure a successful XTs, and if any of them is greater than the XT threshold,
transmission. the new transmission request cannot be successfully
We now show how our XT-aware mechanism works. transmitted.
Here, a new transmission request arrives with a require-
ment of four FSs from source node 1 to destination node 3. B. Routing, Spectrum, and Core Assignment
We assume that the transmission request is pre-assigned
to core 2 starting from the f 3 FS on each link. We need to Algorithm 2: Routing, Spectrum, and Core Assignment
check, for example, whether we can successfully assign the
Input: Transmission request matrix
FSs, shown in red, to the new transmission request.
R (r fsr; dr; λr g ∈ R)
First, we check the increased XTs of the previously as- 1: Sort R in descending order by the required FSs λr
signed transmission requests due to the new transmission 2: Calculate the k-shortest routing paths for each pair of
request to ensure that they are still less than the XT nodes based on the K-SP algorithm [30]
threshold (lines 4–10). For example, since the previously 3: for each r in R do
assigned transmission request ar1 shares the same f 3 on 4: Set the starting FS of r equal to 1 as f s
adjacent cores of the common link e12 with the new trans- 5: for each ksp in KSPr do
mission request, the XT of ar1 on the f 3 FS will increase by 6: if FSs between f s ; f s λr − 1 are available on at
Δxtc1 ; c2 ; e12 . Thus, we should ensure that the sum of least one core (the core is selected according to
XTar1 ; f 3 (the current XT stored in the XT table) and the predefined reducing XT algorithm in
Δxtc1 ; c2 ; e12 is less than the XT threshold. Similarly, Ref. [31]) of each link along the ksp then
we should ensure that the XT for ar4 (f 3 on common link 7: Check XTs based on Algorithm 1
e23 ) is less than the XT threshold (line 15). Moreover, for 8: if Algorithm 1 returns 1 then
ar3, because it shares the same f 3 FS with the new trans- 9: Assign the FSs between f s ; f s λr − 1, light-
mission request on two common links (i.e., e12 and e23 ), the path ksp and the selected cores to serve trans-
XT of ar3 will increase twice [by Δxtc2 ; c3 ; e12 and mission request r
Δxtc2 ; c3 ; e23 , respectively] and should also be less than 10: Update usage table Ue; c; f and XT ta-
the XT threshold. We should repeat this process on each ble XTr; f
pre-assigned FS of the new transmission request. If any 11: break (Move to the next transmission
of the increased XTs of the previously assigned transmis- request)
sion requests are greater than the XT threshold, the pre- 12: end if
assigned FSs and cores cannot be successfully assigned 13: end if
to the new transmission request. 14: end for
Then, we check the XT of the new transmission request 15: f s 1
itself on each pre-assigned FS to ensure that it can be 16: return to line 5
17: end for
reducing XT algorithm in Ref. [31]) of each link along instances (simple network topology, 3-core MCF, small
the ksp (i.e., satisfying the contiguity and continuity set of offered demand requests, and so forth).
constraints), check the XT-limit constraint based on
1) Convergence Efficiency of our ILP/MILP Optimization
Algorithm 1. If the XT-limit constraint is also satisfied
Models: We employ the existing ILP optimization models
(Algorithm 1 returns 1), assign the FSs between
proposed in Refs. [18,21] as benchmarks to evaluate the
f s ; f s λr − 1, the routing path ksp and the cores to the
convergence efficiency of our ILP/MILP optimization mod-
transmission request r. However, if all the ksp in KSPr
els. For our MILP optimization model, we set jFj to be an
are not feasible on the current f s , we increase the starting
upper bound as the sum of all the required bandwidth (FSs)
FS f s by one and repeat the above processes until there is at
for all transmission requests.
least one ksp in KSPr that is feasible. Finally, we update
the usage table Ue; c; f and XT table XTar; f and move For the existing two arc-path-based ILP optimization
to the next transmission request. The pseudocode of our models, we consider three different cases of pre-calculated
Routing, Spectrum, and Core Assignment Algorithm is candidate paths, which are all the possible routing paths
shown in Algorithm 2. (AP), 5-shortest paths (5SP) and 3-shortest paths (3SP).
Moreover, the two existing ILP methods in Refs. [18,21]
and our ILP method consider the XT in a worst interference
IV. SIMULATIONS AND PERFORMANCE EVALUATIONS scenario. To ensure that the candidate path space will not
be bounded by the XT threshold (see Section II.C), we con-
sider a simple and geographically small six-node nine-link
In this section, we evaluate the performance of our ILP/ (n6s9) network, as shown in Fig. 4(a). In the n6s9 network,
MILP methods and our heuristic algorithm. Two types of the XTs (worst case) of all the paths are less than the XT
MCFs are considered in the simulation experiments, which threshold (set as −30 dB in this part of the simulation
are a 3-core MCF (for small-scale problem instances) and a experiments [18,21]).
7-core MCF [10,22] (for large-scale problem instances). The
parameters r, β, k, and ωtr of the MCFs are set as 50 mm, According to the statements above, the solutions ob-
tained by our node-arc-based ILP/MILP methods and the
4 × 106 m−1 , 4 × 10−4 , and 4.5 × 10−5 m, respectively
existing arc-path-based ILP methods (AP) can guarantee
[10,21,22]. The simulation experiments were performed
the optimality. A fair comparison should be considered
on Microsoft Windows 10 using a computer with an Intel
among these four methods, but for reference, we also evalu-
Xeon 8-core 3.5 GHz CPU and 64 GB of memory. We used
ate the performance of the existing ILP methods in the
the Gurobi Optimizer 7.0.1 [29] to solve the optimization
cases of 3SP and 5SP.
models, and for all the heuristic algorithms, we used the
same k-SP algorithm proposed in Ref. [30], where k is The number of transmission requests jRj is set to be 15
set to be 3. Moreover, in the simulation experiments, a unit (a transmission request between each node pair). The traf-
FS (i.e., the grid granularity) is assumed to be 12.5 GHz fic demand ranges considered are from 1–3, 2–4, and to a
bandwidth [30,32], and two neighboring optical paths maximum of 5–7 units of FSs. For each of the traffic
are separated by guardband that occupies 12.5 GHz demand ranges, we generate 10 different transmission re-
bandwidth (i.e., 1 FS) [33]. quest matrices R. The criterion for the convergence effi-
ciency measurement is the “runtimes” that are given by
AMPL/Gurobi [29].
A. Simulation Experiments of Small-Scale
Figure 5 shows the convergence efficiencies of the four
Instances optimization models in the n6s9 network. The x axis shows
different ranges of traffic demands. The y axis shows the
We first evaluate the performance of our ILP/MILP execution time. The results are averaged by 10 different
methods and our heuristic algorithm in small-scale transmission request matrices R.
Fig. 4. (a) Six-node nine-link (n6s9) simple network [21]. (b) 14-node 21-link NSF network [34]. (c) 28-node 34-link EON network [35].
204 J. OPT. COMMUN. NETW./VOL. 10, NO. 3/MARCH 2018 Yang et al.
Fig. 5. Convergence efficiency comparison among the four optimi- Fig. 6. Generalizability of different approaches for the n6s9
zation models for the n6s9 network. simple network.
As shown in Fig. 5, the execution time increases with in- the same FSs to adjacent cores on some common links
creasing traffic demand range for all the methods because along the lightpaths. Thus, as shown in Fig. 6, the required
the feasible solution spaces of the optimization models FSs become large as the XT threshold decreases. We list
expand with the increase in required FSs, leading to the number of blocked transmission requests by the
longer execution time. We can observe that our ILP existing XT-WC-based heuristic algorithms for each XT
method can solve the problem in less than 1 s, and even threshold in Table III as follows.
compared with the cases of 3SP and 5SP of the existing
3) Spectrum Efficiency of our Heuristic Algorithm in
ILP methods, it requires less execution time. However,
Small-Scale Problem Instances: In this paragraph, we
because our MILP method consider the XT strictly
evaluate the spectrum efficiency achieved by our heuristic
(i.e., XT-aware), it takes more execution time. The
algorithm in small-scale problem instances compared with
results are consistent with the theoretical analysis in
the heuristic algorithms proposed in the previous works
Tables I and II.
[18,21]. The solutions obtained based on our ILP optimiza-
tion model are set as the benchmarks. Because in geo-
2) Generalizability of our MILP Method and Heuristic graphically small networks, the solutions obtained based
Algorithm: To evaluate the generalizability of our MILP on our ILP optimization model are also the most optimal
method and our heuristic algorithm, we consider the pre- solution (see Section IV.A.1), but our ILP method requires
vious n6s9 network and let the XT threshold decrease from the least execution time. In this part, the XT threshold is
−30 dB to −40 dB (the different modulation formats [36] fixed to be −30 dB.
and the XT margin [37], which can be set by the network
operators, will lead to different XT thresholds). The de- Figure 7 presents the results of the spectrum efficiency
crease of the XT threshold can also be viewed as the geo- of each approach in the n6s9 network, in which the x axis
graphical size of the network becomes larger, but the XT shows the numbers of transmission requests jRj, and the
threshold remains the same, where the number of trans- source node and destination node for all transmission re-
mission requests jRj is fixed as 15 (between each node pair) quests are randomly selected. The required FSs for each
and the required FSs are randomly selected between 1 and transmission request are considered as a random number
5 FSs. For each XT threshold, we generated 10 different from 1 to 5 (FSs). For each jRj, we generate 10 different
transmission request matrices R, and the results are transmission request matrices R, and the results—
averaged as shown in Fig. 6. maximum required bandwidth (FSs) in the entire
network—are averaged, as shown on the y axis.
As shown in Fig. 6, there are no points for the two
existing XT-WC-based heuristic algorithms when the XT As shown in Fig. 7, in small-scale problem instances, our
threshold is reduced to −34 dB, because the two existing proposed heuristic algorithm performs better than the two
XT-WC-based heuristic algorithms become unable to as- existing heuristic algorithms, and the solutions are closer
sign the lightpaths to some transmission requests. That to the optimal solutions.
is, in the worst interference scenario, the XTs of the short-
est path for these transmission requests become greater TABLE III
than −34 dB (e.g., the transmission request between node NUMBER OF BLOCKED TRANSMISSION REQUESTS BY THE XT-WC-
pair [1,6]); thus, these transmission requests will be BASED HEURISTIC ALGORITHMS IN THE N6S9 NETWORK
blocked. In contrast, our XT-aware-based MILP method Θ (dB) −30 −32 −34 −36 −38 −40
and heuristic algorithm can assign the FSs adaptively to
Number of Blocked r 0 0 1 3 7 12
ensure successful transmissions by avoiding assigning
Yang et al. VOL. 10, NO. 3/MARCH 2018/J. OPT. COMMUN. NETW. 205
[22] A. Muhammad, G. Zervas, and R. Forchheimer, “Resource [30] K. Christodoulopoulos, I. Tomkos, and E. Varvarigos, “Elastic
allocation for space-division multiplexing: optical white box bandwidth allocation in flexible OFDM-based optical net-
versus optical black box networking,” J. Lightwave Technol., works,” J. Lightwave Technol., vol. 29, no. 9, pp. 1354–1366,
vol. 33, no. 23, pp. 4928–4941, Dec. 2015. May 2011.
[23] K. Morita and K. Hirata, “Dynamic spectrum allocation [31] S. Fujii, Y. Hirota, H. Tode, and K. Murakami, “On-demand
method for reducing crosstalk in multi-core fiber networks,” spectrum and core allocation for reducing crosstalk in multi-
in Int. Conf. on Information Networking (ICOIN), IEEE, core fibers in elastic optical networks,” J. Opt. Commun.
Jan. 2017, pp. 686–688. Netw., vol. 6, no. 12, pp. 1059–1071, Dec. 2014.
[24] W. DGrover, Mesh-Based Survivable Networks: Options and [32] ITU-T, “G.694.1:Spectral grids for WDM applications: DWDM
Strategies for Optical, MPLS, SONET, and ATM Networking, frequency grid,” Dec. 2012. Available: https://www.itu.int.
Upper Saddle River, NJ: Prentice Hall, Aug. 2003. [33] M. Klinkowski and K. Walkowiak, “A heuristic algorithm for
[25] A. Cai, G. Shen, L. Peng, and M. Zukerman, “Novel node-arc routing, spectrum, transceiver and regeneration allocation
model and multi-iteration heuristics for static routing problem in elastic optical networks,” in Int. Conf. on
and spectrum assignment in elastic optical networks,” J. Transparent Optical Networks, July 2012, pp. 1–4.
Lightwave Technol., vol. 31, no. 21, pp. 3402–3413, Nov. 2013. [34] B. G. Bathula and J. M. H. Elmirghani, “Constraint-
[26] Y. Wang, X. Cao, Q. Hu, and Y. Pan, “Towards elastic and fine- based anycasting over optical burst switched networks,”
granular bandwidth allocation in spectrum-sliced optical net- J. Opt. Commun. Netw., vol. 1, no. 2, pp. A35–A43,
works,” J. Opt. Commun. Netw., vol. 4, no. 11, pp. 906–917, July 2009.
Nov. 2012. [35] L. Velasco, A. Jirattigalachote, M. Ruiz, P. Monti, L. Wosinska,
[27] D. Siracusa, F. Pederzolli, P. S. Khodashenas, and J. M. Rivas- and G. Junyent, “Statistical approach for fast impairment-
Moscoso, “Spectral vs. spatial super-channel allocation in aware provisioning in dynamic all-optical networks,” J. Opt.
SDM networks under independent and joint switching Commun. Netw., vol. 4, no. 2, pp. 130–141, Feb. 2012.
paradigms,” in European Conf. on Optical Communication, [36] J. Perelló, J. M. Gené, A. Pagès, J. A. Lazaro, and S. Spadaro,
Oct. 2015, pp. 1–3. “Flex-grid/SDM backbone network design with inter-core
[28] P. S. Khodashenas, J. M. Rivas-Moscoso, D. Siracusa, F. XT-limited transmission reach,” J. Opt. Commun. Netw.,
Pederzolli, B. Shariati, D. Klonidis, E. Salvadori, and I. vol. 8, no. 8, pp. 540–552, Aug. 2016.
Tomkos, “Comparison of spectral and spatial super-channel [37] W. Klaus, B. J. Puttnam, R. S. Luís, J. Sakaguchi, J. M.
allocation schemes for SDM networks,” J. Lightwave Mendinueta, Y. Awaji, and N. Wada, “Advanced space
Technol., vol. 34, no. 11, pp. 2710–2716, June 2016. division multiplexing technologies for optical networks
[29] “Gurobi v7.0.1,” 2016 [Online]. Available: http://www.gurobi. [Invited],” J. Opt. Commun. Netw., vol. 9, no. 4, pp. C1–C11,
com. Apr. 2017.