Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/237020093

Shared Backup Path Protection in Elastic Optical Networks: Modeling and


Optimization

Conference Paper · March 2013

CITATIONS READS

35 375

2 authors:

Krzysztof Walkowiak Miroslaw Klinkowski


Wroclaw University of Science and Technology National Institute of Telecommunications
274 PUBLICATIONS 2,263 CITATIONS 170 PUBLICATIONS 2,413 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

NCN OPUS 11 "Modeling and optimization of space division multiplexing (SDM) elastic optical networks" View project

Special Issue "Novel Algorithms and Protocols for Networks" in Applied Sciences (IF: 2.474) View project

All content following this page was uploaded by Miroslaw Klinkowski on 19 May 2014.

The user has requested enhancement of the downloaded file.


Shared Backup Path Protection in Elastic
Optical Networks: Modeling and Optimization
Krzysztof Walkowiak, Member, IEEE and Mirosław Klinkowski

Abstract— Elastic Optical Network (EON) architectures are central frequency. The channel covers both the frequency
considered as a very promising solution for both huge bandwidth range occupied by the optical signal and a guard band required
and flexible connection provisioning in next generation optical for the roll-off filters.
networks. In EON, a basic problem in network design and In EON, the problem of finding unoccupied spectrum
operation is the problem of Routing and Spectrum Allocation
(RSA). In this article, we focus on offline RSA in a survivable
resources so that to establish a lightpath is called the Routing
EON scenario with shared backup path protection (SBPP). We and Spectrum Allocation (RSA) problem. RSA concerns
formulate RSA/SBPP as an Integer Linear Programming (ILP) assigning a contiguous fraction of frequency spectrum to a
problem. Since RSA is a difficult problem itself, we propose connection request subject to the constraint of no frequency
several heuristic algorithms including both new proposals called overlapping in network links. The RSA optimization problem
Adaptive Frequency Assignment with Shared Backup Path is NP-hard [7] and it is more difficult than the Routing and
Protection (AFA/SBPP) and Most Subcarriers and Average
Longest Path First (MSALPF) as well as existing RSA methods
Wavelength Assignment (RWA) problem in fixed grid
adapted to the SBPP scenario. We investigate the efficiency of all wavelength division multiplexing (WDM) networks. Offline
algorithms for a set of network scenarios and we show that the RSA has been addressed with both Integer Linear
proposed new algorithms outperform other reference algorithms. Programming (ILP) [7]-[10], meta-heuristics [8], [11], and
Moreover, numerical experiments show that the shared backup heuristic algorithms [7], [8].
path protection approach enables reduction of the spectrum usage So far, there have been proposed several solutions for
up to 28% comparing to the dedicated path protection approach
without sharing of backup capacity. However, the difference survivable EON in the literature. Regarding offline network
between these scenarios strongly depends on the network design, survivable RSA algorithms for dedicated path
topology and other parameters. protection (DPP) and shared backup path protection (SBPP)
in a ring network have been studied in [12]. DPP in a network
I. INTRODUCTION with generalized connectivity has been addressed with both
ILP formulation [13], metaheuristic [14], and heuristic
T HE evolution of optical communication networks leads
toward elastic optical networks (EONs) in which advanced
single-carrier modulation formats (such as m-PSK, m-QAM)
algorithms [13], [15]. Concurrently, an ILP formulation [16], a
metaheuristic algorithm [17], and an heuristic algorithm [16]
and multi-carrier modulation techniques (such as O-OFDM) have been proposed for SBPP. Besides, in [18] a MILP
are applied for adaptive and mixed-line-rate transmission and formulation and an heuristic algorithm have been presented for
where spectrum resources are allocated within flexible the so-called squeezed protection scheme in which the
frequency grids [1]. These components will allow EON to protection of a lightpath may not cover its entire bandwidth
utilize the spectrum more efficiently and will support elastic but it is allowed to be partial. Eventually, dynamic survivable
and on-demand bandwidth provisioning [2]. Due to space EON scenarios have been studied in [19]. All these solutions
limitations, we refer to [3]-[5] for more details on EON assure network survivability under single link failures.
architectures and proof-of-concept EON experiments. In this work, we focus on optimization of routing and
ITU-T has recently revised the G.694.1 recommendation spectrum assignment in survivable EON which is protected
and included the definition of a flexible DWDM grid (we call using the shared backup path protection approach called
it flexgrid) [6]. According to [6], the frequency spectrum in an RSA/SBPP. The objective is to minimize the width of
optical fiber link is divided into narrow frequency segments spectrum resources required in the network. As a failure
(we refer to them as slices). The optical path (lightpath) is scenario we consider a single link failure. We address two
determined by its routing path and a channel, which consists of possible cases of SBPP, namely with and without stub release.
a flexibly (ad-hoc) assigned subset of slices around a nominal We propose a novel ILP formulation of the problem, which is
less complex than the one presented in [16]. Moreover, we
propose effective heuristic algorithms to solve the problem
Krzysztof Walkowiak is with Wroclaw University of Technology, Wybrzeże
Wyspianskiego 27, PL-50-370 Wroclaw, Poland (phone +48713203539; including an extension of our original proposal for solving
e-mail: krzysztof.walkowiak@pwr.wroc.pl). RSA called AFA (Adaptive Frequency Assignment) and
Mirosław Klinkowski is with National Institute of Telecommunications, 1
Szachowa Street, 04-894 Warsaw, Poland. (e-mail: M.Klinkowski@itl.waw.pl).
MSALPF (Most Subcarriers and Average Longest Path First)
The work of K. Walkowiak was supported in part by the statutory funds of the as well as a set of heuristics developed for the classical RSA
Department of Systems and Computer Networks, Wroclaw University of and modified for the RSA/SBPP problem. Obtained results
Technology. The work of M. Klinkowski was supported in part by NCN under show that AFA provides the best results – on average the
Grant DEC-2011/01/D/ST7/05884.
results are about 4.71% and 2.14% from optimal solutions spectrum requirements in the network when comparing to SC.
yielded by CPLEX for two reference networks. However, the Therefore, in this work, we focus on modeling and
AFA needs significantly lower execution time and provides performance assessment of SBPP in the more efficient DC
radically better scalability (CPLEX solves only small problem scenario. However, presented below algorithms can be easily
instances). Using the heuristic, we conduct extensive adapted to the SC approach.
numerical experiments to investigate the problem of SBPP in In the following, we present ILP formulations of RSA/SBPP
EON. with NS and SR under the DC constraint, which are denoted,
The remainder of the paper is organized as follows. In respectively, as RSA/SBPP/NS/DC and RSA/SBPP/SR/DC.
Section 2, we present the details of our SBPP scenario and we
B. ILP formulation
formulate the optimization problem. In Section 3, we describe
heuristic algorithms. In Section 4, we present numerical The considered network is modeled as a directed graph.
results. Finally, in Section 5 we conclude the work. Links are labeled with e, where e = 1,2,…,E. In EON, the
number of slices to be allocated to a connection is a function
of the requested bandwidth, the modulation technique applied,
II. OPTIMIZATION MODEL the slice width in the flexgrid, and the guard band introduced
In this Section, we present an ILP model of an offline to separate two spectrum adjacent connections, among others.
problem of Routing and Spectrum Allocation with Shared For a given network scenario, where both the transmission
Backup Path Protection (RSA/SBPP) in an EON with static parameters and flexgrid definitions are given, there is a direct
traffic demands and subject to single-link failures. First, we relation between the requested spectrum and the requested
discuss the considered path protection scenario and then we bandwidth (e.g., see Sec. II in [8]). For instance, assuming the
formulate the model. The model will be used as a reference in QPSK modulation format with spectral efficiency 2 bit/s/Hz
the evaluation of heuristic algorithms proposed in Section III. and the slice width equal to 6,25GHz, a 100Gbit/s demand will
occupy a 50GHz spectrum segment consisting of 8 slices.
A. SBPP scenario
Therefore, without loss of generality, in this work the volume
Among network survivability schemes, connection recovery of demands is expressed in terms of the number of slices.
through path protection, in which backup network resources Apart from that, we assume that the requested number of slices
are provisioned in advance for each connection, is preferred is the same for both working and backup connections. It
due to its quick recovery time. In EON, network resources corresponds to an EON scenario in which either a pair of
correspond to spectrum resources in optical fiber links and dedicated working and backup transponders operates with the
path protection is equivalent with provisioning backup same modulation level, or a tunable transponder without
lightpaths for working lightpath connections. modulation adaptation capability is used. RSA/SBPP model
In SBPP, backup resources can be shared between the extensions which account for selection of different modulation
demands whose primary paths are not likely to fail at the same levels on working and backup paths are left for further study.
time, on the contrary to DPP, in which each connection has its Taking the above into account, demands are labeled with d,
own backup resources. Thanks to this property, resource where d = 1,…,D, and each demand corresponds to a lightpath
requirements in SBPP are usually lower than in DPP [20]. request with two end nodes and the volume nd expressed in
Regarding SBPP, the resources on the failure-affected paths terms of the number of slices.
can be either: a) left without use or b) released and used as We apply the link-path modeling approach [21] and we
spare resources for backup connections - the latter concept is consider that for each demand d a set of candidate pairs of link
called stub release. In this paper, we focus on both scenarios, disjoint routing paths indexed p = 1,2,…,Pd is given. In more
that is without and with stub release; they are denoted, detail, for each pair of end nodes we have k candidate pairs of
respectively, as NS and SR. link disjoint paths given. Thus, there are k candidate pairs of
In [13] and [15], there are distinguished two alternative paths for each demand (i.e., Pd = k).
scenarios for EON with path protection capability, namely: a) We take a similar approach as in [10] and [13] for
with Same Channel (SC) allocation and b) with Different formulating the RSA/SBPP problem as an ILP problem. In
Channel (DC) allocation. Case SC is a cost-effective scenario particular, slices are labeled with s, where s = 1,2,…,S. Then,
in which the transponders are shared between primary for demand d a set of candidate optical channels c = 1,2,…,Cd
(working) and backup connections and a traffic demand has is considered, where each channel consists of a subset of
allocated the same segment of optical frequency spectrum (i.e., adjacent slices of size nd slices, and RSA concerns selecting a
channel) on its primary and backup path. Such a solution path and assigning a channel to a demand. As shown in [10],
reduces the network cost and alleviates the connection this modeling approach is simple and computationally more
switching time [15]. In the DC case, the SC constraint of efficient when comparing with alternative ILP formulations of
having the same operating channel in both primary and backup RSA [7]-[9]. Consequently, the ILP formulation of RSA/SBPP
connections is relaxed, at the cost of installing either tunable presented in this paper is more efficient than the previous
or dedicated (for both working and backup lightpaths) formulation in [16], since [16] makes use of [7].
transponders. In [13], it is shown that DC allows to reduce In total, there are seven sets of binary decision variables in
the presented models. First, xdpc and zdpc denote the selection of ∑p∑c xdpc = 1, d = 1,2,…,D (2)
channel c for demand d, respectively, on a working and a
∑d∑p∑c γdpcsδedpxdpc = wes e = 1,2,…,E s = 1,2,…,S (3)
backup routing path p. The third variable is wes, which is 1, if
slice s on link e belongs to a working lightpath and 0 ∑d∑p∑c δgdpαdpcsβedpzdpc = begs e = 1,2,…,E g = 1,2,…,E
otherwise. Next, wegs says if slice s is occupied on link e for a g ≠ e s = 1,2,…,S (4)
working lightpath and the selected working lightpath is not
wes + begs ≤ yes e = 1,2,…,E g = 1,2,…,E
using link g. Similarly, begs denotes if slice s is occupied on g ≠ e s = 1,2,…,S (5)
link e for restoration in the case of link g failure. The sixth
variable is yes, which is 1, if slice s is occupied on link e and 0 ∑e yes ≤ Eys s = 1,2,…,S (6)
otherwise. Finally, variable ys denotes if slice s is occupied on ∑c xdpc = ∑c zdpc d = 1,2,…,D p = 1,2,…,Pd (7)
any network link. The objective is to minimize the width of
spectrum resources (denoted as Φ) required in the network, The objective (1) is to minimize the width of spectrum, in
similarly as in [7]-[9], [13], and [16]. In particular, Φ terms of the number of slices, required in the network.
corresponds to the largest slice index among all slices Equation (2) assures for each demand d exactly one candidate
allocated in network links and it determines the width of path and exactly one candidate channel are selected. To find
spectrum that the network should support in a green-field the allocation of slices to working paths and to meet the
network design [9]. guarantee that a slice on a particular link can be allocated to at
To present the models we use notation as in [21]. most one lightpath, we add equation (3) to the model.
Constraint (4) denotes the allocation of slices to backup paths.
RSA/SBPP/NS/DC (Without Stub Release) A single link failure (denoted as index g) is considered. Only
indices demands affected by the failure (i.e., δgdpzdpc = 1) are taken
s = 1,2,…,S slices into account. Next, the slice allocation following from
d = 1,2,…,D demands activation of backup paths is included in the definition of begs.
p = 1,2,…,Pd candidate pairs of link disjoint paths for To meet the guarantee that a slice on a particular link can be
demand d allocated to at most one lightpath considering both working
c = 1,2,…,Cd candidate channels for demand d and backup paths as well as sharing of backup resources, we
e,g = 1,2,…,E network links add equation (5) to the model. Constraint (6) defines that slice
constants s is used in the network (ys = 1) only when there is at least one
δedp = 1, if link e belongs to working path p realizing link on which the slice s is allocated. Constraint (7) allows to
demand d; 0, otherwise assign different channels to working and backup path of a
βedp = 1, if link e belongs to backup path p realizing particular demand, however the same pair of paths (associated
demand d; 0, otherwise with index p) must be used for the demand.
nd number of slices required for demand d
γdpcs = 1, if channel c associated with demand d on working RSA/SBPP/SR/DC (With Stub Release)
path p uses slice s; 0, otherwise (calculated according variables (additional)
to nd) wegs = 1, if slice s is occupied on link e for working paths
αdpcs = 1, if channel c associated with demand d on backup and the selected working path is not using link g; 0,
path p uses slice s; 0, otherwise (calculated according otherwise (binary)
to nd) objective (1)
variables subject to (2), (4), (6-7) and
xdpc = 1, if channel c on working path p is used to realize ∑d∑p∑c γdpcs(1–δgdp)δedpxdpc = wegs e = 1,2,…,E g = 1,2,…,E
demand d; 0, otherwise (binary) g ≠ e s = 1,2,…,S (8)
zdpc = 1, if channel c on backup path p is used to realize
demand d; 0, otherwise (binary) wegs + begs ≤ yes e = 1,2,…,E g = 1,2,…,E
wes = 1, if slice s is occupied on link e for working flows; g ≠ e s = 1,2,…,S (9)
0, otherwise (binary) Equation (8) defines variable wegs, which is 1, if slice s is
begs = 1, if slice s is occupied on link e for restoration in the occupied on link e by working paths excluding working path
case of link g failure; 0, otherwise (binary) that use link g. Therefore – comparing to (3) – on the left-hand
yes = 1, if slice s is occupied on link e; 0, otherwise side of (8) an additional term (1 – δgdp) is added.
(binary) Consequently, when link g fails, the slices allocated to serve
ys = 1, if slice s is occupied on any network link; 0, working paths of demands affected by the failure (i.e., selected
otherwise (binary) working paths of these demands use link g) are released and
objective can be used to serve for backup paths. Constraint (9) is a new
minimize Φ = ∑s ys (1) definition of variable yes and together with (8) guarantees the
stub release mode.
subject to Note that, similarly as in the RSA/DPP model [13], the
above RSA/SBPP models can be extended to the SC scenario
by substituting constraint (7) with the following constraint: paths). Next, all allocated demands for which the selected
xdpc = zdpc d = 1,2,…,D p = 1,2,…,Pd c = 1,2,…,Cd (10) working path shares at least one link with path p are identified
and slices allocated to backup paths of these demands are also
However, the evaluation of the SC model is out of scope of
excluded from the residual network. Let Allocate_W(d,p,s)
this paper.
denote a function that allocates in the network demand d to
working path p starting from slice s. Similarly,
III. HEURISTIC ALGORITHMS
Allocate_B(d,p,s) allocats the demand to a backup path. Let
In this Section, we describe heuristic algorithms for solving Sort_ALG(B) denote a sorting operation applied to a set of
the RSA/SBPP problem. The algorithms are based on demands B according to the selected ordering method ALG,
previously proposed RSA algorithms: e.g., in the case of the LPF method demands are sorted
− FF a Fixed-alternate routing and First-fit frequency according to the decreasing values of the length of routing
assignment algorithm [22]; paths.
− LPF (Longest Path First) algorithm based on a greedy Each algorithm presented in this Section is formulated in
sequential processing of demands according to the two versions. In the former one – called SA (separate
decreasing values of the length of routing paths [8]; assignment) – first only working paths of each demand are
− MSF (Most Subcarriers First) algorithm based on a greedy allocated in the network and next backup paths are analyzed.
sequential processing of demands according to the In the latter version – named JA (joint assignment) in the same
decreasing number of requested slices [8]; run both working and backup paths are allocated jointly.
− LPMSF (Longest Path and Most Subcarriers First) Below, we present the pseudocode of FF/SBPP/SA and
algorithm based on a greedy sequential processing of FF/SBPP/JA algorithms, which simply process all demands
demands according to the decreasing length of a routing without any special sorting and assigns the demand to the first
path, as an additional criterion the decreasing number of candidate path.
requested slices (demand size) is taken into account [16];
− AFA (Adaptive Frequency Assignment) algorithm Algorithm 1 FF/SBPP/SA
adaptively selects a sequence of processed demands in order 1: for d = 1 to D do begin
2: s ← MinFS_W(d, 1)
to minimize Φ [7].
3: Allocate_W(d, 1, s)
Moreover, we introduce a following algorithm similar to 4: end
LPMSF, however using another sorting method: 5: for d = 1 to D do begin
6: s ← MinFS_B(d, 1)
− MSALPF (Most Subcarriers and Average Longest Path 7: Allocate_B(d, 1, s)
First) algorithm is based on a greedy sequential processing 8: end
of demands according to decreasing number of requested
slices, in the case of a tie, demands are sorted according to Algorithm 2 FF/SBPP/JA
decreasing value of an average length of candidate paths. 1: for d = 1 to D do begin
Note that all algorithms except AFA are based on the same 2: s ← MinFS_W(d, 1)
3: Allocate_W(d, 1, s)
scheme consisting in sequential processing of demands which
4: s ← MinFS_B(d, 1)
are sorted according to a selected criterion.
5: Allocate_B(d, 1, s)
Now we introduce several functions and operators necessary 6: end
to describe all algorithms. Let MinFS_W(d,p) return the lowest
indexed and accessible slice when demand d is assigned to
Now, we focus on methods LPF, MSF, LPMSF and
working path with index p. In analogous way, MinFS_B(d,p)
MSALPF. All four methods process demands in a similar
returns the lowest indexed and accessible slice when demand d
greedy way, however, various ordering of demands is applied.
is allocated to a backup path with index p. Both functions
As an example pseudocode, we describe the LPF methods both
work in a residual network, i.e., all previously allocated
for SA and JA scenarios.
demands are taken into account and only available slices (not
allocated) are available for a new demand. In more detail, in
Algorithm 3 LPF/SBPP/SA
the case of working paths all previous allocations of working
1: B ← {d : d = 1,2,…,D}
and backup paths must be considered to find the residual
2: B ← Sort_LPF(B)
network. In the case of backup paths – according to the SBPP
3: for each demand d ∈ B do begin
approach – the residual network is calculated as follows. 4: smin ← Large_number
Without loss of generality, let assume that we consider that 5: for p = 1 to Pd do begin
demand d is assigned to path pair p. First, all slices allocated 6: s ← MinFS_W(d, p)
to working paths are excluded from the residual network (in 7: if s < smin then path(d) ← p and smin ← s
the case of the stub release scenario, the slices allocated to 8: end
working paths of all failed demands are available for backup 9: Allocate_W(d, path(d), smin)
10: end avoid potential collisions.
11: for each demand d ∈ B do begin Let B denote a set all demands. Set Bn contains all demands
12: p ← path(d) from B with requested nd equal to n, i.e., Bn = {d : d ∈ B,
13: s ← MinFS_B(d,p) nd = n}. For each link e = 1,2,…,E, the collision metric is
14: Allocate_B(d,p,s)
defined as ce = ∑d∑p δedpnd. Notice that ce estimates the number
15: end
of slices that might be allocated to link e taking into account
all candidate paths (both working and backup). Let lp = ∑e∈p ce
First, algorithm LPF/SBPP/SA sorts demands according to
denote the length of path pair p (both working and backup)
selected ordering (line 2). Since working and backup paths are
calculated according to metric ce. Metric ld = |Pd|-1 ∑p lp
examined in separate runs, there are two main loops in the
denotes the average length of candidate path pairs of demand d
algorithm: lines 3-10 and lines 11-15. In the first loop, for each
according to link metric ce.
demand d all candidate paths are analyzed to find the one
For the sake of notation simplicity, we introduce the
guaranteeing selection of the lowest slice. Since the candidate
following functions. Let MinDemandFS_W(d) return the
path is chosen in the context of the working path, for the
lowest accessible slice for allocation of demand d and the
backup path only slice selection is made (line 13).
index of a working path guaranteeing the lowest slice
alocation. The processing of MinDemandFS_W(d) is
Algorithm 4 LPF/SBPP/JA
analogous to lines 4-8 of the LPF/SBPP/SA algorithm. Let
1: B ← {d : d = 1,2,…,D}
MinDemandFS(d) return the lowest accessible slice for
2: B ← Sort_LPF(B)
allocation of demand d to both working and backup paths and
3: for each d ∈ B do begin
the index of the candidate path pair – similarly to lines 4-10 of
4: smin ← Large_number
5: for p = 1 to Pd do begin
LPF/SBPP/JA algorithm.
6: s1 ← MinFS_W(d, p)
7: s2 ← MinFS_B(d, p) Algorithm 5 AFA/SBPP/SA
8: s ← max(s1, s2) 1: B ← {d : d = 1,2,…,D}, n ← max{nd : d ∈ B}, i = 1
9: if s < smin then path(d) ← p, smin ← s, sw ← s1, sb ← s2 2: Bn ← {d : d ∈ B, nd = n}
10: end 3: for each d ∈ Bn do (sd, path(d)) ← MinDemandFS_W(d)
11: Allocate_W(d, path(d), sw) 4: d* ← arg min (sd). If more than one demand yields the
12: Allocate_B(d, path(d), sb) minimum value of sd, use metric ld as a tie-breaker
13: end 5: Ci ← d*, i ← i + 1
6: Allocate_W(d*, path(d*), sd*)
The JA version of the LPF/SBPP algorithm analyzes all 7: Bn ← Bn \ {d*}; if Bn = ∅, go to step 8, otherwise go to step 3
demands in a single run, thus there is only one main loop (lines 8: n ← n – 1; if n < 1 go to step 9, otherwise go to step 2
3-13). Again all candidate pair of paths are examined to find 9: for i = 1 to D do begin
10: d ← Ci
the one with the lowest slice index. When we compare SA
11: p ← path(d)
approach against JA approach the key difference is the number
12: s ← MinFS_B(d,p)
of function MinFS_B(d, p) calls. In the former algorithm this 13: Allocate_B(d,p,s)
function is called D times, while in the second case it is called 14: end
kD, where k denotes the number of candidate paths. Recall that
function MinFS_B(d, p) among others calculates the residual The AFA/SBPP/SA algorithm analyzes demands divided
network including available slices according to the SBPP into subsets according to the value of the number of slices
approach. Therefore, the JA approach on average needs more required for a demand. The process of generating the subsets is
computational time to find the solution comparing to the SA given in lines 1, 2 and 8. Each subset of demands is processed
approach. to allocate working paths (lines 3-7), next using the same order
Note that in the case of algorithms LPMSF and MSALPF in of demands the backup paths are selected (lines 9-14). In a
line 6 of SA version and line 8 of JA version as the tie-breaker currently analyzed subset, all remaining (not allocated)
the summary length of working and backup path is applied as demands are processed to find the lowest slice allocation (lines
in [16]. Consequently, algorithm LPMSF/SBPP/JA is the same 3 and 4). Afterwards, the selected demand d* is assigned to a
as the algorithm proposed in [16]. path guaranteeing a selection of the lowest possible slice. The
Finally, we present the AFA algorithm based on the method procedure to assign slices for backup paths is comparable to
that we have proposed in [7]. The key idea behind the AFA algorithm LPF/SBPP/SA.
algorithm is to adaptively divide demands into subsets and
process each subset individually. Moreover, a special collision
Algorithm 6 AFA/SBPP/JA
metric is used to select the best paths. In particular, we want to
1: B ← {d : d = 1,2,…,D}, n ← max{nd : d ∈ B
select paths that do not include links that can be potentially
2: Bn ← {d : d ∈ B, nd = n}
selected in a large number of demands and consequently to
3: for each d ∈ Bn do begin
4: (sdw ,sdb ,path(d)) ← MinDemandFS(d) We have used the same set of network and traffic scenarios
5: sd ← max(sdw ,sdb) as in the evaluation of dedicated path protection in our recent
6: end paper [13]. In particular, for all networks but INT9, the results
7: d* ← arg min (sd). If more than one demand yields the are averaged over 100 randomly generated demand sets. The
minimum value of sd, use metric ld as a tie-breaker results for INT9 are averaged over 88 demand sets since the
8: Allocate_W(d*, path(d*), sd*w)
ILP solver could not attain the optimality for some of DPP
9: Allocate_B(d*, path(d*), sd*b)
scenarios (see [13]). The use of common scenarios allows us
10: Bn ← Bn \ {d*}; if Bn = ∅, go to step 11, otherwise go to step 3
11: n ← n – 1; if n < 1 stop, otherwise go to step 2
to compare the efficiency of DPP and SBPP in EON.
All results presented below in subsections A, B and C refer
to the scenario without stub release. Only subsection D
Algorithm AFA/SBPP/JA works analogously to
includes results with stub release.
AFA/SBPP/SA, however both working and backup paths are
processed in the same run. As in the case of algorithms LPF, A. Comparison of Algorithms
MSF, LPMSF and MSALPF, also the JA version of AFA is The first goal of experiments is to evaluate performance of
requires larger number of residual network constructions heuristic algorithms. For smaller networks (SIMPLE6 and
comparing to the SA version. INT9) optimal results were provided by IBM ILOG CPLEX
All heuristic algorithms presented in this section are 12.4 (with default settings) [23]. When solving ILP problems,
polynomial time algorithms. Due to space limitations, we skip it was more difficult to find the solution of SBPP than of the
the detailed discussion on their complexity. corresponding DPP problem (see the relevant results for DPP
in [13]). To speed up the computation of SBPP, we set the
IV. RESULTS solution of DPP as an upper bound (UB) on Φ. As a result, the
In this Section, we present and discuss the results of computation time of SBPP has been reduced significantly,
computational experiments. Four networks topologies are from more than tens of minutes to some seconds (see Table I).
examined, namely: SIMPLE6 (6 nodes, 16 links), INT9 (9 In Table I, we report performance of heuristics in
nodes, 26 links), NSF15 (15 nodes, 46 links) and UBN24 (24 comparison to optimal results. We present results of SA and
nodes, 86 links), all shown in Fig. 1. We compare heuristic JA approaches for each heuristic. Average value of optimality
algorithms presented in Sec. III and, besides, for small gap and corresponding values of lengths of 95% confidence
networks, we provide optimal results obtained by solving ILP intervals are presented (the results are averaged over 100
models presented in Sec. II. Our main focus is on the occupied demand sets for SIMPLE6 and 88 demand sets for INT9). The
spectrum width (Φ), in terms of the number of slices, and execution time of each heuristic was always below 30ms in
computation time (T). The numerical experiments were small networks. We can easily notice that AFA outperforms
performed on an Intel i5 3.3GHz 16GB computer. other algorithms for both tested networks and provides results
close to optimal ones. Moreover, the JA approach yields better
INT9
SIMPLE6 results comparing to the SA approach for each method except
the FF algorithm.
For larger networks (NSF15 and UBN24) the CPLEX is not
UBN24
able to provide optimal results in reasonable times, therefore
NSF15 we compare only heuristics. In Table II, we show the average
values of Φ, average execution time, average distance to
minimum results and corresponding lengths of 95% confidence
intervals (the results are averaged over 100 demand sets for
each topology) To find the distance to minimum results, for
Fig. 1. Network topologies used in simulations: SIMPLE6, INT9, NSF15, and each unique demand set we run all algorithms and next for
UBN24. each algorithm we calculate a percentage gap to the best
Candidate pairs of primary-backup paths are link disjoint (minimum) obtained result among all tested methods. Again,
and they are calculated (and ordered) as shortest paths, taking the AFA method provides the best results, while the MSALPF
into account the overall length of both paths; we consider algorithm always is the second. Recall that both these methods
k∈{2,3,5,10,30}. As discussed in Sec. II.B, demands are are original algorithms proposed in this paper. The potential
expressed in terms of the number of requested slices. In drawback of the AFA method is much larger execution time
details, the requested spectrum nd is an even number generated comparing to other methods, especially in the JA approach.
with a uniform distribution and for randomly selected source The MSALPF algorithm needs computation time similar to
and destination pair of nodes. We assume the ITU flexgrid other greedy methods LPF, MSF, LPMSF, while yields better
definition [6], which requires to have the spectrum allocated results. As in the case of smaller networks, the JA approach
symmetrically around a central frequency. Therefore, in the provides better results than the SA approach. However – as it
evaluation, nd is considered to be an even number, in was pointed out in the previous section – the JA approach
particular, nd∈{2,4,..., nmax}, where nmax∈{8,16}. requires more operations what is reflected in higher values of
TABLE I
OPTIMALITY GAP OF HEURISTICS FOR SMALLER NETWORKS
Scenario ILP FF MSF LPF MSALPF LPMSF AFA
Network Heuristic Φ Time [sec] Average optimality gap
SA 22.00 3.7 12.81% 10.00% 10.80% 9.95% 10.45% 7.90%
SIMPLE6
JA 22.00 3.7 14.46% 6.55% 11.12% 5.76% 9.82% 4.71%
SA 31.75 6.3 7.06% 4.22% 4.92% 4.31% 4.48% 3.40%
INT9
JA 31.75 6.3 7.45% 3.30% 5.21% 2.78% 4.84% 2.15%
Lengths of 95% confidence intervals
SA - - 2.34% 2.20% 2.11% 2.18% 2.13% 1.91%
SIMPLE6
JA - - 2.54% 1.97% 2.23% 1.89% 2.29% 1.68%
SA - - 1.91% 1.51% 1.43% 1.44% 1.37% 1.38%
INT9
JA - - 1.86% 1.36% 1.63% 1.27% 1.56% 1.22%

TABLE II
COMPARISON OF HEURISTICS FOR LARGER NETWORKS
Scenario Φ Execution time [s]
Net. Heur. FF MSF LPF MSALPF LPMSF AFA FF MSF LPF MSALPF LPMSF AFA
SA 393.7 308.2 318.6 293.0 301.8 283.4 0.2 0.2 0.2 0.2 0.2 0.6
NSF15
JA 417.9 272.9 294.5 263.3 283.4 256.2 0.2 5.4 5.3 5.3 5.4 70.9
SA 918.6 651.4 640.8 632.3 632.9 585.4 2.6 2.0 2.0 2.0 1.9 7.8
UBN24
JA 942.6 567.0 559.5 542.9 550.3 521.3 2.0 49.2 47.7 48.3 47.4 1048.9
Average distance to minimum result Lengths of 95% confidence intervals
SA 28.64% 9.13% 12.05% 4.47% 7.18% 1.30% 1.02% 0.86% 0.92% 0.78% 0.94% 0.49%
NSF15
JA 38.90% 6.66% 13.50% 3.31% 10.11% 0.60% 0.69% 0.63% 0.75% 0.55% 0.75% 0.27%
SA 36.16% 10.12% 8.68% 7.44% 7.51% 0.11% 0.73% 0.81% 0.71% 0.82% 0.65% 0.09%
UBN24
JA 44.79% 8.20% 7.11% 4.20% 5.59% 0.41% 0.51% 0.97% 0.63% 0.81% 0.57% 0.18%

TABLE III
PERFORMANCE OF HEURISTICS AS A FUNCTION OF THE NUMBER OF CANDIDATE PATHS
Scenario Φ Execution time [s]
Net. k FF MSF LPF MSALPF LPMSF AFA FF MSF LPF MSALPF LPMSF AFA
2 417.9 371.4 392.0 368.2 381.9 366.7 0.2 0.6 0.6 0.6 0.6 5.7
3 417.9 364.5 382.8 358.2 372.1 356.2 0.2 0.8 0.8 0.8 0.9 8.9
NSF15 5 417.9 348.2 360.3 342.1 351.4 340.6 0.2 1.2 1.2 1.1 1.2 13.5
10 417.9 310.6 316.7 299.4 305.7 292.6 0.2 2.0 2.1 2.0 2.1 25.3
30 417.9 272.9 294.5 263.3 283.4 256.2 0.2 5.4 5.3 5.3 5.4 70.9
2 942.6 813.4 826.9 794.4 817.3 792.6 2.2 6.8 6.6 6.8 6.6 103.0
3 942.6 756.1 762.9 741.4 755.7 730.5 2.3 9.0 8.7 9.0 8.8 151.0
UBN24 5 942.6 681.7 684.0 664.7 675.1 639.3 2.3 12.7 12.4 12.7 12.6 237.3
10 942.6 605.8 604.9 586.6 595.1 560.1 2.3 21.3 20.9 21.2 21.1 429.7
30 942.6 567.0 559.5 542.9 550.3 521.3 2.0 49.2 47.7 48.3 47.4 1048.9

execution time.
C. SBPP versus DPP
B. Impact of the Number of Candidate Paths The last goal of experiments is to examine the benefits
Our next goal was to analyze the impact of parameter k following from the sharing of backup path resources.
(number of candidate paths) on the objective function Φ. In Therefore, we compare the SBPP approach against DPP
Table III, we report the average results and execution time of (Dedicated Path Protection) scenario. For comparison we use
all heuristics as a function of k for large networks NSF15 and the results of the DPP evaluation in our recent paper [12]. In
UBN24. Obviously, for the FF algorithm there is no impact of Table IV, we present the results obtained with both
different values of k since the algorithm always selects the first approaches, for smaller networks the results are optimal, for
candidate path. The improvement in Φ for other algorithms larger networks the results of AFA/SBPP/JA are reported. We
between k = 2 and k = 30 is in the range 24%-30% and 30%- can easily notice that the gap between DPP and SBPP depends
34% for NSF15 and UBN24, respectively. However, the on network topology and the number of candidate paths. The
execution time of algorithms grows approximately in a linear general trend observed for larger networks is that with the
way with the increase of k what is in harmony with the greedy increase of the number of candidate paths, the gap between
approach applied in the algorithms. The execution times are both approaches also increases.
still on acceptable levels and we can conclude that it is D. Stub Release
beneficial to use a large set of candidate paths since it has a
Up to now, all presented results were obtained for the
great impact on Φ.
scenario without stub release. However, ILP solver as well as
all heuristics were also run for the stub release approach.
Nevertheless, the results with stub release were almost the with constructive methods proposed in this work. Besides, we
same as without stub release. In more detail, only in two of would like to examine other objective functions, e.g., to
388 demand sets considering all four tested networks, there minimize the total maximum spectrum index of all links.
was a slight difference between two investigated scenarios. In
our opinion such results follow mainly from two facts. First, REFERENCES
the objective function Φ is a min-max function, what means [1] O. Gerstel et al., "Elastic optical networking: A new dawn for the
that in many cases some changes in network routing and slice optical layer?," IEEE Comm. Mag., vol. 50, no. 2, 2012.
[2] M. Klinkowski et al., "Elastic spectrum allocation for time-varying
allocation can have no impact on the objective. Second,
traffic in flexgrid optical networks," IEEE J. Sel. Areas Commun., 2013
comparing to classical survivability models formulated in the (accepted for publication).
context of MPLS and similar protocols, our EON models [3] M. Jinno et al., "Spectrum-efficient and scalable elastic optical path
include additional slice continuity constraint, which makes network: Architecture, benefits, and enabling technologies," IEEE
Comm. Mag., vol. 47, no. 11, 2009.
quite difficult using of slices released by working paths after a [4] D. J. Geisler et al., "The first testbed bemonstration of a flexible
single link failure. bandwidth network with a real-time adaptive control plane," in Proc. of
ECOC, Geneva, Switzerland, Sep. 2011.
TABLE IV [5] F. Cugini et al., "Demonstration of flexible optical network based on
SBPP VS. DPP path computation element," J. Lightw. Technol., vol. 30, no. 5, 2012.
[6] ITU-T Recommendation G.694.1 (ed. 2.0), "Spectral grids for WDM
Average 95% applications: DWDM frequency grid," Feb. 2012.
Network k DPP SBPP [7] M. Klinkowski and K. Walkowiak, "Routing and spectrum assignment
Gap Conf. Int.
in spectrum sliced elastic optical path network," IEEE Commun. Lett.,
SIMPLE6 3 25.28 22.00 12.64% 2.4%
vol. 15, no. 8, 2011.
INT9 2 33.61 31.75 5.56% 1.8% [8] K. Christodoulopoulos et al., "Elastic bandwidth allocation in flexible
2 449.6 366.7 18.4% 0.9% OFDM based optical networks," IEEE J. Lightw. Technol., vol. 29, no.
3 432.5 356.2 17.6% 0.9% 9, 2011.
NSF15 5 432.5 340.6 21.2% 1.0% [9] Y. Wang et al., "Towards Elastic and Fine-granular Bandwidth
10 382.2 292.6 23.4% 1.0% Allocation in Spectrum-sliced Optical Networks," OSA/IEEE J. of Opt.
30 358.1 256.2 28.3% 0.9% Commun. and Netw., vol. 4, no. 11, 2012.
2 970.0 792.6 18.2% 0.7% [10] L. Velasco et al., "Modeling the routing and spectrum allocation
3 882.1 730.5 17.1% 1.0% problem for flexgrid optical networks," Phot. Netw. Commun., vol. 24,
no. 3, 2012.
UBN24 5 882.1 639.3 27.4% 0.8%
[11] L. Gong et al., "A two-population based evolutionary approach for
10 737.8 560.1 24.0% 0.6%
optimizing routing, modulation and spectrum assignments (RMSA) in
30 677.8 521.3 22.9% 1.0% O-OFDM networks," IEEE Commun. Lett., vol. 16, no. 9, 2012.
[12] T. Takagi et al., “Algorithms for maximizing spectrum efficiency in
elastic optical path networks that adopt distance adaptive modulation,”
in Proc. of ECOC, Torino, Italy, Sep. 2010.
V. CONCLUSIONS [13] M. Klinkowski and K. Walkowiak, "Offline RSA algorithms for elastic
In this paper, we have focused on optimization of shared optical networks with dedicated path protection consideration," in Proc.
of RNDM, St. Petersburg, Russia, Oct. 2012.
protection in elastic optical networks. New ILP models of [14] M. Klinkowski, "A genetic algorithm for solving RSA problem in elastic
routing and spectrum assignment in EONs protected against optical networks with dedicated path protection," in Adv. in Intell. Syst.
single link failures were introduced. To the best of our and Comput. series, vol. 189, 2013.
[15] A. N. Patel et al., “Survivable transparent flexible optical WDM
knowledge, the models are the first ones that use the notion of (FWDM) networks,” in Proc. of OFC, Los Angeles, USA, Mar. 2011.
channels in the context of SBPP protection in EONs. Since the [16] A. Eira et al., "Optimized design of shared restoration in flexible-grid
ILP modelling can be applied efficiently only for relatively transparent optical networks," in Proc. of OFC, Los Angeles, USA,
2012.
small problem instances, we proposed several heuristic [17] A. Castro et al., "Path-based recovery in flexgrid optical networks," in
algorithms. Through extensive numerical experiments we Proc. of ICTON, Coventry, England, Jul. 2012.
showed that the best heuristic (original algorithm called AFA) [18] K. D. R. Assis, R. C. Almeida, and H. Waldman, "MILP formulation for
provides results very close to optimal results for small network squeezed protection in spectrum-sliced elastic optical path networks," in
Proc. of SPECTS 2012, Genoa, Italy, Jul. 2012.
topologies. Furthermore, we run tests for larger networks to [19] X. Shao, Y.-K. Yeo, Z. Xu, X. Cheng, and L. Zhou, "Shared-path
verify performance of algorithms as a function of number of protection in OFDM-based optical networks with elastic bandwidth
candidate paths and compare SBPP protection against non- allocation," in Proc. of OFC, Los Angeles, USA, 2012.
[20] K. Walkowiak, J. Rak, “Shared Backup Path Protection for Anycast and
sharing DPP. The main conclusions are: (i) increasing the Unicast Flows Using the Node-Link Notation,” in Proc. of ICC, Kyoto,
number of candidate paths can significantly improve the Japan, 2011.
objective function and (ii) benefit of SBPP over DPP is about [21] M. Pióro and D. Medhi, Routing, flow and capacity design
in communication and computer networks. Morgan Kaufmann, 2004.
28% and 22% for NSF15 and UBN24 networks, respectively. [22] M. Jinno, et al., "Distance-adaptive spectrum resource allocation in
Additionally, we discovered that SBPP with stub release spectrumsliced elastic optical path network," IEEE Comm. Mag., vol.
approach provides practically no additional savings comparing 48, no. 8, pp. 138–145, 2010.
[23] ILOG AMPL/CPLEX software: www.ilog.com/products/cplex/
to the SBPP without stub release.
In future work, we plan to develop soft optimization
algorithms (e.g., evolutionary, tabu search, simulated
annealing) to obtain results closer to optimal when comparing

View publication stats

You might also like