Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir...

13
Cross-Layer Approach to Survivable DWDM Network Design Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce Abstract—All-optical networks, in which the elec- trical regeneration bottlenecks are removed, are seen as the next-generation backbone networks. Any link failure in these high-speed environments, if not dealt with promptly, is catastrophic and can cause the loss of gigabits of data. While techniques to improve the survivability of optical networks are now well- established, such is not the case with all-optical net- works. In these environments, the absence of regen- eration implies that physical impairments accumu- late over long paths. So-called cross-layer techniques mitigate the physical impairments’ impact on the net- work layer performance. In this work, we apply cross- layer techniques, previously successfully applied to the impairment-constrained routing and wavelength assignment problem [IEEE J. Sel. Areas Commun., vol. 26, p. 32, 2008], to the problem of improving the survivability of all-optical networks facing link fail- ures. To the best of our knowledge, cross-layer surviv- ability of all-optical networks has never been studied before. We present algorithms that improve the net- work survivability over non-cross-layer algorithms by decreasing both the blocking probability and the vulnerability of the network to failures. Our mecha- nisms are evaluated with extensive simulations for a realistic regional-sized network. The cross-layer algorithms are computationally intensive, and to alle- viate this issue we propose two new compound resto- ration algorithms as well as two novel quality-of- transmission-aware protection schemes that exhibit low blocking probability and have a moderate vulner- ability ratio and time complexity. Index Terms—All-optical networks; Network survivability; Wavelength assignment; Wavelength routing. I. INTRODUCTION T he always-increasing demand for high- throughput transmission in the backbone of data networks has drawn attention to all-optical dense wavelength division multiplexing (DWDM) networks. In such ultra-high-speed environments, the effect of a component failure becomes much more severe and survivability considerations can prevent significant service interruptions. But the design procedure in all- optical networks has particular challenges and, as shown in this paper, neglecting the physical layer is- sues can lead to unacceptable performance. This is mainly due to the removal of optical–electrical–optical (OEO) conversion, which is the main speed bottleneck but helps ameliorate signal quality. Consequently, the quality of the received signal can be below the ac- cepted level for the receiver—thus causing the connec- tion or call to be dropped [1]. We refer to this event as quality of transmission (QoT) or physical blocking. In this work, we measure QoT in terms of bit error rates (BERs), which could increase above an acceptable level (e.g., BER=10 -9 , a threshold value set by the network operator). Also, since all-optical wavelength conversion is not yet mature for commercial deploy- ment, we assume no wavelength conversion in our study. In this case, a call may also be blocked due to unavailability of a wavelength-continuous path or simply wavelength blocked. Although the network survivability problem in DWDM networks has been extensively studied in the past [2–4], the physical layer has not been taken into account. Our studies show that in many cases the physical blocking can be the dominating component in the total blocking probability for a connection. These results call for a cross-layer approach in survivable network design. We show that cross-layer (or QoT- aware) designs can greatly increase the network per- formance in terms of lowering the blocking probability and vulnerability ratio (a metric we formally define later to quantify the survivability of the network to a random link failure). There has been a significant amount of research on routing and wavelength assignments that consider physical impairments in DWDM networks [5–15], but Manuscript received August 1, 2008; revised January 26, 2010; ac- cepted April 2, 2010; published May 12, 2010 Doc. ID 123025. A. Askarian (e-mail: [email protected]) and S. Subramaniam are with and Y. Zhai was with the Department of Electrical and Computer Engineering, George Washington University, Washington, DC 20052, USA. Y. Pointurier is with Alcatel-Lucent Bell Labs, Nozay, France. M. Brandt-Pearce is with the Charles L. Brown Department of Electrical and Computer Engineering, University of Virginia, Charlottesville, Virginia 22904, USA. Digital Object Identifier 10.1364/JOCN.2.000319 Askarian et al. VOL. 2, NO. 6/ JUNE 2010/ J. OPT. COMMUN. NETW. 319 1943-0620/10/060319-13/$15.00 © 2010 Optical Society of America Authorized licensed use limited to: The George Washington University. Downloaded on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Transcript of Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir...

Page 1: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

Askarian et al. VOL. 2, NO. 6 /JUNE 2010/J. OPT. COMMUN. NETW. 319

Cross-Layer Approach to SurvivableDWDM Network Design

Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

TnwIcssossm(bqctqt(lncmsus

Dpaptrnafalr

rp

Abstract—All-optical networks, in which the elec-trical regeneration bottlenecks are removed, are seenas the next-generation backbone networks. Any linkfailure in these high-speed environments, if not dealtwith promptly, is catastrophic and can cause the lossof gigabits of data. While techniques to improve thesurvivability of optical networks are now well-established, such is not the case with all-optical net-works. In these environments, the absence of regen-eration implies that physical impairments accumu-late over long paths. So-called cross-layer techniquesmitigate the physical impairments’ impact on the net-work layer performance. In this work, we apply cross-layer techniques, previously successfully applied tothe impairment-constrained routing and wavelengthassignment problem [IEEE J. Sel. Areas Commun.,vol. 26, p. 32, 2008], to the problem of improving thesurvivability of all-optical networks facing link fail-ures. To the best of our knowledge, cross-layer surviv-ability of all-optical networks has never been studiedbefore. We present algorithms that improve the net-work survivability over non-cross-layer algorithmsby decreasing both the blocking probability and thevulnerability of the network to failures. Our mecha-nisms are evaluated with extensive simulations for arealistic regional-sized network. The cross-layeralgorithms are computationally intensive, and to alle-viate this issue we propose two new compound resto-ration algorithms as well as two novel quality-of-transmission-aware protection schemes that exhibitlow blocking probability and have a moderate vulner-ability ratio and time complexity.

Index Terms—All-optical networks; Networksurvivability; Wavelength assignment; Wavelengthrouting.

Manuscript received August 1, 2008; revised January 26, 2010; ac-cepted April 2, 2010; published May 12, 2010 �Doc. ID 123025�.

A. Askarian (e-mail: [email protected]) and S. Subramaniam arewith and Y. Zhai was with the Department of Electrical andComputer Engineering, George Washington University, Washington,DC 20052, USA.

Y. Pointurier is with Alcatel-Lucent Bell Labs, Nozay, France.M. Brandt-Pearce is with the Charles L. Brown Department of

Electrical and Computer Engineering, University of Virginia,Charlottesville, Virginia 22904, USA.

Digital Object Identifier 10.1364/JOCN.2.000319

1943-0620/10/060319-13/$15.00 ©

Authorized licensed use limited to: The George Washington University. Download

I. INTRODUCTION

he always-increasing demand for high-throughput transmission in the backbone of data

etworks has drawn attention to all-optical denseavelength division multiplexing (DWDM) networks.

n such ultra-high-speed environments, the effect of aomponent failure becomes much more severe andurvivability considerations can prevent significantervice interruptions. But the design procedure in all-ptical networks has particular challenges and, ashown in this paper, neglecting the physical layer is-ues can lead to unacceptable performance. This isainly due to the removal of optical–electrical–optical

OEO) conversion, which is the main speed bottleneckut helps ameliorate signal quality. Consequently, theuality of the received signal can be below the ac-epted level for the receiver—thus causing the connec-ion or call to be dropped [1]. We refer to this event asuality of transmission (QoT) or physical blocking. Inhis work, we measure QoT in terms of bit error ratesBERs), which could increase above an acceptableevel (e.g., BER=10−9, a threshold value set by theetwork operator). Also, since all-optical wavelengthonversion is not yet mature for commercial deploy-ent, we assume no wavelength conversion in our

tudy. In this case, a call may also be blocked due tonavailability of a wavelength-continuous path orimply wavelength blocked.

Although the network survivability problem inWDM networks has been extensively studied in theast [2–4], the physical layer has not been taken intoccount. Our studies show that in many cases thehysical blocking can be the dominating component inhe total blocking probability for a connection. Theseesults call for a cross-layer approach in survivableetwork design. We show that cross-layer (or QoT-ware) designs can greatly increase the network per-ormance in terms of lowering the blocking probabilitynd vulnerability ratio (a metric we formally defineater to quantify the survivability of the network to aandom link failure).

There has been a significant amount of research onouting and wavelength assignments that considerhysical impairments in DWDM networks [5–15], but

2010 Optical Society of America

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 2: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

tatcpl

Aotiptpobfmb

abttarpH

ppmIrIcnam

A

tTan

wtl

s

320 J. OPT. COMMUN. NETW./VOL. 2, NO. 6 /JUNE 2010 Askarian et al.

network survivability has not been considered inthese papers. The most closely related work to ours is[16], which investigated path protection routing andwavelength assignment (RWA) algorithms consideringtransmission impairments with the goal of achievingmaximum resource sharing, but it resorts to OEOplacement in the network to achieve the desired qual-ity for the received signal. Our paper considers fullytransparent networks and addresses this problemwith a cross-layer approach and also defines a metricto evaluate the vulnerability of the algorithms to afailure. To the best of our knowledge, routing andwavelength assignment algorithms that strive to im-prove the resilience of all-optical networks to link fail-ures without sacrificing low call blocking probabilitiesare proposed and evaluated for the first time in thispaper.

The main contributions of this paper are as follows.First we evaluate the performance of different exist-ing survivability algorithms in all-optical networkswith realistic physical layer impairments. Then weapply a cross-layer algorithm to protection and resto-ration schemes and show the considerable perfor-mance improvement they provide. To mitigate thehigh time complexity of the cross-layer algorithms, wealso propose other simpler and yet powerful algo-rithms.

Survivability in DWDM networks can be achievedby protection or restoration, which in turn can bepathwise or linkwise; furthermore, protection algo-rithms can use either a shared protection path or adedicated protection path [2]. We begin by looking atthe dedicated path protection �1+1�. In the dedicatedpath protection, every connection has two link-disjointlightpaths to handle single-link failures, a primarypath and a backup path. In networks with regenera-tion [such as a synchronous optical network (SO-NET)], both the primary and backup paths are simul-taneously used, and the receiving node monitors eachcopy of the signal and uses the best one (lowest BER).This ensures quick traffic restoration in case one ofthe paths fails. However, in transparent DWDM net-works that are transmission impaired, keeping thebackup path dark versus lighting it up has an impacton the QoT of other lightpaths in the network. Light-ing up the backup path worsens the impairments forother lightpaths due to added crosstalk and thus in-creases the blocking probability of lightpaths. On theother hand, keeping the backup path dark (until it isneeded) can lead to increased traffic restoration times(due to additional needed signaling between transmit-ting and receiving nodes). Therefore, it is of interest tostudy the effect of lighting up the backup path on net-work performance. Besides blocking probability, wealso study the impact of dark and lit backup paths onthe vulnerability of connections to failures. A link pro-

Authorized licensed use limited to: The George Washington University. Download

ection approach in such environments is also evalu-ted. Path and wavelength provisioning for link pro-ection schemes leads to long lightpaths for theonnections. We show that link protection does noterform well in all-optical networks with physicalayer impairments [17].

We also look at path and link restoration schemes.s expected, they exhibit better performance in termsf the blocking probability than their protection coun-erparts. An interesting result of our study shows thatn realistic transmission-impaired DWDM networks,ath restoration has a failure vulnerability close tohat of path protection. This shows that a naive ap-roach to reserve resources for failure recovery notnly wastes the network resources and increases thelocking probability, but also cannot ensure recoveryrom a failure any more than a simple restorationethod that uses resources efficiently and has low

locking probability.

We use our previously proposed1 cross-layer RWAlgorithm called highest Q or HQ [5] and apply it tooth protection and restoration schemes. We will seehat this cross-layer approach significantly improveshe performance in terms of both blocking probabilitynd failure vulnerability. The drawback of this algo-ithm is its high computational complexity; we pro-ose two additional algorithms and two compoundQ–non-HQ algorithms that alleviate this issue.

The paper is organized as follows. In Section II weresent the network model and assumptions for thehysical layer and define the metrics used for perfor-ance evaluation of the RWA algorithms. In Section

II we look at the proposed cross-layer protection andestoration algorithms in DWDM networks. In SectionV we propose modifications to achieve high-speedross-layer algorithms for survivability in all-opticaletworks. In Section V we present simulation resultsnd evaluation of the proposed algorithm perfor-ances. Finally we conclude the paper in Section VI.

II. NETWORK MODEL AND PERFORMANCE METRICS

. Network Model

In this section, we present the model and assump-ions for the physical layer used throughout the paper.his model was previously proposed by us in [5,11],nd we summarize it here for clarity and complete-ess.

We consider circuit-switched all-optical networksith no wavelength conversion. On a call arrival, ei-

her one or two new circuits are tentatively estab-ished depending on the RWA algorithm being used.1This algorithm was originally developed without considering

urvivability.

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 3: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

[[ccndbatit

wAr

eoglaleNlsmfitnctcW[nwinlc

Askarian et al. VOL. 2, NO. 6 /JUNE 2010/J. OPT. COMMUN. NETW. 321

The RWA algorithms are presented in Section III.Physically, a circuit corresponds to a lightpath [18],that is, the combination of a route (sequence of nodescalled optical cross-connects or OXCs, separated byspans of fibers) and a channel (a wavelength). Notethat by lightpath establishment we mean that the re-sources are reserved, whether the correspondingwavelength is lit or not. We assume that all links arebidirectional and carry exactly �Total wavelengths ineach direction. Due to the absence of wavelength con-version, lightpaths must respect the wavelength con-tinuity constraint and remain on the same wave-length end to end.

The physical components of a lightpath (see Fig. 1)are a transmitting laser, optical cross-connects, spansof fibers, and a receiver. We model amplifiers as non-saturating, and the receiver as a wideband optical fil-ter (for demultiplexing purposes) and a photodetectorfollowed by a narrow electrical filter. In this work, wedo not assume that transmission at the physical layeris error free: error-free transmission is a valid as-sumption only for small networks and large networkswhere signals are periodically regenerated electroni-cally. In the context of regional or even metropolitanall-optical networks, the distances involved are solarge that physical impairments are no longer negli-gible. We measure the QoT of a lightpath by its BER,which should remain below a threshold set by the net-work manager to ensure almost error-free data trans-mission.

To estimate BERs, we use the relation between theBER and the so-called corresponding Q factor (an elec-trical signal-to-noise ratio) for on–off-keying modula-tion: BER= 1

2 erfc�Q /�2�. The Q factor for a signal on alightpath is given by, assuming Gaussian distribu-tions for the “0” and “1” samples after photodetection[19]:

Q =�1 − �0

�0 + �1, �1�

where �0 and �1 are the means of the “0” and “1”samples, respectively, and �0 and �1 are their stan-dard deviations.

sourcecall

OXC

interchannel crosstalk

OXC and amplifiersfiber spans

amplifiers, Ofiber span

node crosstalknode crosstalknoise

Amplifier

Fig. 1. (Color online) Model of a transmission path used to computin fiber spans causes nonlinear crosstalk, while leaks in the OXCs

Authorized licensed use limited to: The George Washington University. Download

Here, we account for four dominating impairments20]: intersymbol interference (ISI), amplifier noiseamplified spontaneous emission (ASE) noise], inter-hannel nonlinear effects (also called nonlinearrosstalk), and optical leaks at the nodes (also calledode crosstalk). A fifth impairment, polarization modeispersion (PMD), is negligible at 10 Gbps but shoulde incorporated at faster data rates (40 Gbps/channelnd more); we chose to ignore it in this work. Each ofhe four aforementioned effects can be accounted forn the Q factor as noiselike terms (variances), suchhat

�12 = �i

2 + �n2 + �nl

2 + �nx2 , �2�

here �i2, �n

2, �nl2 , and �nx

2 are the variances due to ISI,SE noise, nonlinear crosstalk, and node crosstalk,espectively.

ISI is caused by the interplay between fiber nonlin-arity and dispersion characteristics, and ASE noiseriginates from the amplifier medium; therefore, for aiven lightpath, ISI and ASE noise depend only on theightpath’s physical and topological properties (suchs the number of spans of the lightpath and theirengths). Fast techniques using precomputed tablesxist in the literature to estimate �i and �n [21,22].onlinear crosstalk is the result of interplay between

it channels in fiber spans, while node crosstalk con-ists of leaks inside the nodes, whether it is at the de-ultiplexers (port crosstalk) or inside the switching

abric (fabric crosstalk). Demultiplexer crosstalk cann turn be either adjacent port crosstalk (the channelshat interfere are adjacent in the optical spectrum) oronadjacent port crosstalk. The intensity of fabricrosstalk and demultiplexer crosstalk vary accordingo the OXC implementation; however, nonadjacentrosstalk is always weaker than adjacent crosstalk.e presented a detailed model for node crosstalk in

11], which we reuse here. Contrary to ISI and ASEoise, nonlinear and node crosstalk depend on the net-ork status: lighting more paths increases nonlinear

nteractions within fiber spans, thereby causing moreonlinear crosstalk, and increases the number of

eaks in the OXCs, thereby causing more noderosstalk. Since crosstalks are network-status-

OXC

node crosstalk

s photodetector diagrameye

power power

probability density

the ‘‘0’’s ‘‘1’’sdistributions of

time

| · |2σ0

σ1

µ1

µ0

e Q factor. Amplifiers inject ASE noise, interplay between channelsse node crosstalk.

XCs,

e thcau

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 4: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

wdaclirhu

p

fipappibQtsgcrpno

tsatoim

t

322 J. OPT. COMMUN. NETW./VOL. 2, NO. 6 /JUNE 2010 Askarian et al.

dependent effects, it is not possible to precomputetheir standard deviations �nl and �nx. However, it ispossible to precompute the standard deviations for asingle term of each kind of crosstalk [22,23], and ap-propriate summation of these variances over the set ofinterfering lightpaths makes it possible to design fastQoT estimators, for which the only online computa-tions consist of determining which lightpaths inter-fere and summing their respective effects. Such esti-mators pave the way for the design of online QoT-aware RWA algorithms, as shown in the next section.

B. Performance Metrics

We investigate RWA algorithm performance fromtwo perspectives, blocking probability in the call ad-mission process and vulnerability to a random failureduring transmission. After taking the physical layerimpairments into consideration, there are two types ofblockings: wavelength blocking due to the unavailabil-ity of a continuous wavelength on the chosen path(wavelength continuity constraint not met) and QoTblocking due to the unsatisfactory Q factor of the pathafter network resources have been allocated to it (QoTconstraint not met).

Since the purpose of survivability schemes is to pre-vent connections from breaking down because of fail-ures, in addition to various types of blocking in thesystem, we are also interested in the network behav-ior when a random failure occurs. Single-link failureis considered in this paper: at any time, at most onelink failure is allowed in the entire network. In ourlink failure model, we consider that (single) link fail-ure location and time of failure are randomly (uni-formly) distributed over their respective domains.

We define the vulnerability ratio as a metric to de-scribe the performance of our algorithms in the con-text of random single-link failures. The vulnerabilityratio is defined as the probability that a randomlypicked ongoing connection (at the time of failure) can-not be restored because of lightpath unavailability (inthe case of restoration algorithms) or unacceptableQoT (for both protection and restoration algorithms) ifa random link fails at a random point of time duringthe operation of the network. To compute the vulner-ability ratio, we note that the vulnerability of a con-nection stays the same between network statechanges (i.e., connection admissions and departures).Therefore, we can calculate the vulnerability ratio byaveraging the vulnerability over all network states.

For a failure on link j in network state i, the prob-ability that a random ongoing connection fails is

Authorized licensed use limited to: The George Washington University. Download

Pij =

Dij

Ti, �3�

here Dij is the number of the connections that are

ropped (due to unacceptable QoT or lightpath un-vailability), and Ti is the total number of ongoingonnections in state i. We denote by M the number ofinks and by S the total number of network states dur-ng network evaluation. For each network state pe-iod, any of the links can fail with equal probability,2

ence the following average over possible link fail-res:

Pi =1

M�j=1

M

Pij =

1

M�j=1

M Dij

Ti. �4�

Then, averaging over the entire network operationeriod, the vulnerability ratio is

V =1

�i=1

S

�i

�i=1

S

Pi�i =1

M

1

�i=1

S

�i

�i=1

S

�j=1

M Dij�i

Ti, �5�

where �i is the duration of state i.

III. SURVIVABLE RWA ALGORITHMS

Traditional RWA algorithm design assumes a per-ect physical layer, which leads to downgraded block-ng probability performance when physical layer im-airments are taken into consideration. Theselgorithms typically have low wavelength blockingrobability but have high QoT blocking probability—articularly when there are strong physical layermpairments—and hence an unsatisfactory totallocking probability. To deal with this situation, newoT-aware RWA algorithms have been designed and

heir performances have been verified through exten-ive simulations. The idea behind QoT-aware RWA al-orithms is to take physical layer impairments intoonsideration while choosing the wavelength andoute for a connection request in the admission controlrocess with the hope that this new connection doesot significantly degrade the QoT performance ofther ongoing connections.

We look at two non-QoT-aware RWA algorithms. Inhe first, a connection is routed according to thehortest-path algorithm and the first (lowest-index)vailable wavelength on that path (if any) is assignedo that connection. We refer to this method as first fitr FF in short. In the second approach, for each work-ng wavelength we try to find the shortest path (none

ay be available, due to wavelength unavailability2We assume this in this paper. Other failure probability distribu-

ions can easily be incorporated if needed.

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 5: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

mapscpcbb

wopfii

B

OFftampaHtrhuawrrwt

ct

C

olorrtrs

nrg

Askarian et al. VOL. 2, NO. 6 /JUNE 2010/J. OPT. COMMUN. NETW. 323

somewhere along that path) and assign the shortestone among all to the connection. We refer to this ap-proach as best fit or BF.

As a QoT-aware algorithm, we consider the HQ al-gorithm [5]. In HQ, a shortest-path algorithm is runon each wavelength to find a candidate path on eachwavelength. Then the end-to-end Q factor is calcu-lated for all candidate paths and, among all the can-didate paths, the path with the highest Q factor ischosen for the current connection. In our previouswork [5], we have shown that this algorithm leads tolow average BER and high fairness among connec-tions with different path lengths.

A. Path Protection

In the context of dedicated path protection, three ofthe aforementioned RWA algorithms are consideredhere: shortest-path routing with FF wavelength as-signment, BF RWA, and the HQ RWA algorithm.These RWA algorithms using either a lit or darkbackup path are investigated under dedicated pathprotection schemes using the same network topologyand physical layer parameters.

First let us explain the connection admission proce-dure for the FF RWA. With the lit backup path protec-tion scheme, the FF algorithm is run twice in order tocompute two link-disjoint paths. If the wavelengthcontinuity constraint cannot be met on either of thetwo paths, the connection is wavelength blocked. Oth-erwise, both paths are assumed to be lit up and QoTblocking verification starts (the wavelengths for thetwo lightpaths may be different). In this scheme, onlyone of the two paths of any ongoing connection needsto meet the BER threshold requirement for the receiv-ing end to correctly receive the data. Thus, the inter-ference brought into the network by the establish-ment of the two paths of the new connection requestshould be limited enough so that no connection in thenetwork sees both its lightpaths disrupted (due to anunmet QoT constraint) at the same time. If both theprimary and the backup path of any ongoing connec-tion (or those of the new connection) do not meet theQoT constraint, then the requested connection is QoTblocked. If the requested call is admitted, then bothlightpaths are lit up.

In the dark backup path protection scheme, thewavelength blocking check is the same as with the litbackup path protection scheme; however, since thedark backup path protection scheme only lights upone path during the whole transmission period, theQoT verification phase differs. One of the paths (firstshortest path) is assumed to be lit and the QoT con-straint is checked for the primary path of every ongo-ing connection in the network, including the new con-nection itself. If all primary paths in the network

Authorized licensed use limited to: The George Washington University. Download

eet the QoT constraint, then the new connection isdmitted; the path chosen to be lit is the primaryath, while the other path (for which the QoT con-traint is not checked) is the backup path. If the QoTonstraint is violated, then the same procedure is re-eated with the second shortest path of the incomingonnection. If the QoT constraint cannot be met foroth the first and the second shortest path, then QoTlocking occurs.

The procedure detailed above is similar in the casehere the chosen RWA algorithm is BF or HQ insteadf FF, except that the candidate paths for the roles ofrimary and backup path are not required to be therst and second shortest path, but are chosen accord-

ng to the BF or HQ algorithms.

. Path Restoration

Path restoration can be achieved in several ways.ne approach is to use the shortest-path routing andF wavelength assignment scheme to set up a path

or a requested connection and to use the same to findhe restoration path between the connection sourcend destination in case of link failure. We refer to thisethod as FF-FF3 path restoration. To improve the

erformance of this method, we introduce a QoT-ware path restoration scheme, in which we use theQ algorithm to find the path for the arrived connec-

ion and to restore a connection from a link failure. Weefer to this approach as HQ-HQ path restoration. Toave a fair comparison with the HQ algorithm thatses the shortest paths on all available wavelengthss the candidates to find the one with highest Q factor,e also look at the shortest-path best-fit (BF-BF) path

estoration method in which, both for primary andestoration paths, the shortest path on every availableavelength is found and then the shortest among

hem is chosen for the connection.

Table I summarizes our pathwise algorithms, in-luding the fast QoT-aware algorithms that will be in-roduced in Section IV.

. Link Protection and Restoration

In link protection schemes, there is a need for anffline algorithm that finds a protection path for eachink and reserves wavelengths along it, so that in casef link failure the protection path for that link is al-eady known and the required wavelength is alreadyeserved. The algorithm we use for link protection inhis paper is derived from [24]. In that work, an algo-ithm is presented to find a two-connected directedubgraph of the network graph. Let us call this di-3In our terminology for path restoration schemes, the first acro-ym refers to the wavelength assignment algorithm to the primaryoute and the second refers to the restoration route. The routing al-orithm is shortest path in both cases.

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 6: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

ri

A

ltaatswuTcsdwsscf

nfttmabnatuHra

B

g

324 J. OPT. COMMUN. NETW./VOL. 2, NO. 6 /JUNE 2010 Askarian et al.

rected subgraph the blue digraph (directed graph). Atthe same time another subgraph, which we refer to asthe red digraph, is generated that is exactly the sameas the blue digraph except that the edge directions arereversed. Half of the available wavelengths, say set�1, are assigned to be used by primary paths on theblue digraph and by protection paths on the red di-graph. The remaining set of wavelengths, say set �2,are used to carry primary data on the red digraph andprotection data on the blue digraph. Upon arrival, aconnection can be routed on either of the two di-graphs, using the wavelength set assigned for the pri-mary data on that digraph. In this paper, we use theshortest-path routing algorithm with first-fit wave-length assignment to find the primary path. In case ofa link failure, those connections that were using thatlink on the blue digraph direction, which are using awavelength in set �1, would be routed on the backuppath on the red digraph around that link. Since set �1was reserved for protection in the red digraph, itwould be available for all those connections. The sameapproach would be followed to protect data on theother digraph.

The backup path for each link is static and can befound offline; hence in case of failure the backup pathis already known around each link. In our approach,we assign the shortest path around each link (on theother digraph) for protecting a connection passingthat link in each direction. Notice that all the connec-tions on the failed link would follow the same backuppath; therefore there is no need for demultiplexingand multiplexing these connections at either end ofthe failed link.

In the link restoration scheme we study here, arriv-ing calls are routed according to the shortest-pathrouting algorithm with first-fit wavelength assign-ment. In case of link failure, for each affected connec-tion we find the shortest available path around thatlink on the same wavelength the connection is alreadyusing. It is obvious that different connections mayneed to use different restoration paths.

Clearly in both link and path restoration, there ex-ists the possibility that a restoration path cannot befound due to wavelength unavailability. This contrib-utes to the vulnerability ratio of the restoration algo-

TABPATH PROTECTION AND

Non-QoT-Aware

Protectionschemes

FF (Lit/Dark) Backup (Sec. III.A)BF (Lit/Dark) Backup (Sec. III.A)

Restorationschemes

FF-FF (Sec. III.B)BF-BF (Sec. III.B)

Authorized licensed use limited to: The George Washington University. Download

ithms and numerical results for it are presented latern the paper.

IV. FAST FAILURE-RECOVERY RWA ALGORITHMS

. Compound Path Restoration Algorithms

As discussed in the previous section—and the simu-ation results in the next section support this fact—he HQ-HQ path restoration algorithm has a desir-ble performance in terms of both blocking probabilitynd vulnerability ratio, but it is computationally in-ensive and can be slow for time-sensitive applicationsuch as streaming voice and video. To have a networkith seamless connectivity even in the face of a fail-re, a high-speed restoration algorithm is desirable.oward this goal, we look at the combination of HQonnection setup and non-QoT-aware restorationchemes. The idea behind this approach is that someelay in connection setup phase can be acceptable, buthen a failure happens, connections with time-

ensitive traffic should be restored as quickly as pos-ible. For this reason we use the HQ algorithm in theall setup phase to gain low blocking probability, butor the restoration phase we use faster schemes.

In the first algorithm we look at, every arriving con-ection is routed based on the HQ method. In case of a

ailure, the restoration path for every affected connec-ion is found according to the FF algorithm. We refero this algorithm as HQ-FF. As opposed to the HQ-HQethod in which the restoration paths are also found

ccording to the HQ algorithm, HQ-FF is expected toe considerably faster. Our simulation result in theext section supports this expectation. In the secondpproach, we establish the primary path according tohe HQ algorithm, and for the restoration paths wese the BF approach. We refer to this algorithm asQ-BF. The performances of these two proposed algo-

ithms in terms of blocking probability and vulner-bility ratio are discussed in Subsection V.D.

. Path Protection Algorithms With Low Complexity

To avoid the computational complexity of the HQ al-orithm even in the connection setup phase, for the

IESTORATION SCHEMES

oT-Aware Compound

Q (Lit/Dark) Backup (Sec. III.A)PALW (Lit/Dark) Backup (Sec. IV.B)C (Lit/Dark) Backup (Sec. IV.B)

N/A

Q-HQ (Sec. III.B) HQ-FF (Sec. IV.A)HQ-BF (Sec. IV.A)

LER

Q

HSMH

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 7: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

aO

udl

castetenttflt

wrtnnacw

tisprb

A

SaNbCtn

Askarian et al. VOL. 2, NO. 6 /JUNE 2010/J. OPT. COMMUN. NETW. 325

applications that need rapid bandwidth provisioning,we first propose the idea of including informationabout the physical layer in link weights [25]. In thisway, a simple shortest-path algorithm can implicitlytake into account the physical layer characteristics.We introduce the shortest path with adaptive linkweights algorithm (SPALW), a QoT-aware schemethat runs a shortest-path algorithm on an adaptivelink weight network. The link weights are chosen torepresent the physical layer interference. Three fac-tors contribute to link weights: the physical length ofthe link, the wavelength availability of the link, andthe number of established connections passingthrough the link’s end nodes.

The physical length of the link accounts for the ASEnoise from amplifiers. The longer the physical lengthis, the more the number of amplifiers are and thestronger the ASE noise is. We denote this factor by L.

Wavelength availability is a parameter that ac-counts for the wavelength continuity constraint. Intraditional routing techniques, links are weighted ac-cording to their length. However, by assigning higherweights to links with fewer available wavelengths,connections tend to be routed on the links that havemore available wavelengths, thereby increasing thechance to meet the wavelength continuity constraint.We define the wavelength availability � of a link asfollows:

� =�Used

�Total − �Used.

Thus, when there are no used wavelengths on a link,�Used=0 and �=0. When all the wavelengths on a linkare in use, �Used=�Total and �=�.

In our physical layer model, we include impair-ments due to node crosstalk (optical leaks at the de-multiplexers or within the switching fabric in nodes).In [11], three types of crosstalk were introduced,namely, switch port crosstalk, self-crosstalk, andneighbor crosstalk. Switch port crosstalk comes fromthe interaction between two connections that traversethe same node on the same wavelength. The other twotypes of crosstalk (adjacent crosstalk) are only pos-sible if several connections using adjacent wave-lengths traverse the same node; the more connectionstraversing a node, the higher the crosstalk impair-ments experienced by those calls. For this reason, weinclude in the link weights the number of establishedconnections at the link’s end nodes. We define the es-tablished connection quantity at the head node asQHead and at the tail node as QTail.

To make these three factors comparable and havethe same relative influence on the link weights, threecoefficients are defined. We call � the physical lengthcoefficient, the wavelength availability coefficient,

Authorized licensed use limited to: The George Washington University. Download

nd the established connection quantity coefficient.verall, we define link weights as

Link Weight = �L + � + �QHead + QTail� = �L

+ �Used

�Total − �Used+ �QHead + QTail�.

SPALW selects the path with the minimum weightsing a shortest-path algorithm in the weighted graphefined above and assigns the first available wave-ength to the connection.

We call our second RWA algorithm minimumrosstalk (MC). This algorithm is adopted from [11]nd is adjusted here for protection purposes. MC isimilar to HQ, with a different wavelength pickingechnique. MC runs a shortest-path algorithm forach wavelength (with constant link weights equal tohe physical link lengths) to find candidate routes. Forach candidate route, the number of crosstalk compo-ents along the route is calculated. Since the twoypes of crosstalk we consider have different leak ra-ios, we use two coefficients to differentiate their in-uences. For each route candidate, the crosstalk in-ensity (CI) on wavelength j is defined as

CIj = �i=1

Nr

�Nis + �Ni

a,

here Nr is the number of nodes on the consideredoute and wavelength j, Ni

s is the number of connec-ions on the same wavelength at node i, Ni

a is theumber of connections on adjacent wavelengths atode i, � is the switch port crosstalk ratio, and � is thedjacent wavelength crosstalk ratio. Among all theandidate routes, the MC algorithm chooses the routeith the minimum crosstalk intensity: CI=minj�CIj�.

It should be noted that the low-complexity protec-ion algorithms proposed here only improve the speedn the connection setup phase. In case of a failure,witching impacted connections to their protectionaths takes the same time as in path protection algo-ithms presented earlier (for lit backup or darkackup schemes, respectively).

V. SIMULATION RESULTS

. Simulation Model

To evaluate our algorithms, we use the Nationalcience Foundation (NSF) topology depicted in Fig. 2s an example of a mesh topology. We downscale theSF topology (originally a continental-size network)y a factor of 10, resulting in a regional-size network.ontinental-size networks require intermediate elec-

rical regeneration: indeed, even considering ISI andoise only and ignoring network-state-dependent im-

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 8: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

idnrat=SPhsrtlw

B

sinnhltptapwuwlcwodIl

Ft

326 J. OPT. COMMUN. NETW./VOL. 2, NO. 6 /JUNE 2010 Askarian et al.

pairments (nonlinear and node crosstalks), it is notpossible to transmit signals over more than roughly1000 km with standard techniques4 while achievingadequate QoT (BER 10−9, corresponding to a Q fac-tor of 6) [5]. The regional network we consider, on thecontrary, exhibits milder impairments that are lowenough to guarantee that, at low loads (and hencewhen no or low crosstalk occurs), any node is reach-able from any other node while maintaining adequateQoT. At higher loads, interchannel and nodecrosstalks become disruptive, but their effects aremitigated by QoT-aware RWA algorithms such as HQ,as shown below. For simplicity, we modified the NSFtopology such that each link consists of an integernumber of 70 km fiber spans.

The physical parameters for the simulated networkare summarized in Table II; the values used are typi-cal for modeling next-generation regional-size all-optical networks. The high attenuation for nonadja-cent port crosstalk we used essentially means that we

4Note that link distances longer than 1000 km are achievable us-ing optimized long-haul link design and components.

1

22

2

4

4

1

1 2

1

2

1

1 121

1

4

2

2 1

Fig. 2. (Color online) Topology used in the simulations. We used adownscaled version of the NSF net topology (14 nodes, 21 bidirec-tional links) to perform our simulations. The link weights on the fig-ure correspond to the number of 70 km long fiber spans.

TABLE IIPHYSICAL PARAMETERS FOR THE SIMULATED NETWORK

Description Value

Span length 70 kmSignal peak power 2 mW

Bit duration 100 ps �10 Gbps�Pulse shape NRZ

Fabric crosstalk −40 dBAdjacent port crosstalk −30 dB

Nonadjacent port crosstalk −60 dBAdjacent wavelength crosstalk −25 dB

Fiber loss 0.2 dB/kmNonlinear coefficient 2.2 �W km�−1

Linear dispersion 17 ps/nm/kmDispersion compensation 100% post-DC

ASE noise factor 2Receiver electrical bandwidth 7 GHz

Number of wavelengths 8Minimum Q factor 6

Authorized licensed use limited to: The George Washington University. Download

gnored it; indeed, in practice, the main leaks at theemultiplexers come from adjacent wavelength chan-els. The typical value for the switch port crosstalkatio and adjacent wavelength crosstalk ratio are −30nd −25 dB, respectively; therefore the parameters ofhe MC algorithm are �=10−30 dB/10=0.001 and �10−25 dB/10=0.0032. We used �=1, =100, and =2 inPALW. Calls are assumed to arrive according to aoisson process and have exponentially distributedolding times with unit mean. The results for eachimulation run are averaged over 5000 connection ar-ivals. The source and destination nodes of a connec-ion are randomly (uniformly) selected. The networkoad is thus the total arrival rate of calls to the net-ork.

. Performance Evaluation: Blocking

First we look at the path and link protectionchemes. The wavelength blocking probability, whichs only due to unavailability of wavelength and doesot take into account the quality of the received sig-al, is shown in Fig. 3. One noticeable observationere is that the link protection scheme has much

ower wavelength blocking probability than path pro-ection schemes. This can be explained as follows. Inath protection schemes, for each requested connec-ion a working path and a link-disjoint backup pathre reserved. This means that considering all shortestaths between all source–destination pairs in the net-ork, only half of the wavelengths are available to setp a working path. But in the link protection schemee consider here, on each digraph half of the wave-

ength set is available for the working path, and inase no free wavelength is found, the connectionould search on the working wavelength set on thether digraph (notice that considering any of the twoigraphs alone, the network is still fully connected).ntuitively speaking, in the link protection scheme aarger number of wavelengths are available to an ar-

2 4 6 8 10 12 14 16 18 2010

−4

10−3

10−2

10−1

100

Load (Erlang)

Wav

elen

gth

Blo

ckin

gP

roba

bilit

y

FF, Dark Backup Path ProtectionFF, Lit Backup Path ProtectionHQ, Dark Backup Path ProtectionHQ, Lit Backup Path ProtectionBF, Dark Backup Path ProtectionBF, Lit Backup Path ProtectionLink Protection

ig. 3. Wavelength blocking probability versus traffic load for pro-ection algorithms.

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 9: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

pis

QcbttatbmatH

tritlbaBcstsoiwct

C

t

Ft

Askarian et al. VOL. 2, NO. 6 /JUNE 2010/J. OPT. COMMUN. NETW. 327

riving connection. This is achieved at the cost oflonger paths and, as we will see next, has a negativeresult on the quality of the received signal and in-duces a high vulnerability to failure.

One other observation is the lower wavelengthblocking of the lit backup scheme in all RWA algo-rithms. The reason is that in lit backup schemes (aswe will see in Fig. 4) more connections are blocked dueto the low quality of their received signals; thereforemore free wavelengths are available for the new arriv-als. In this graph, we also see that the wavelengthblocking probabilities of BF and HQ path protectionalgorithms are in the same range and all are lowerthan that of FF schemes. This is due to the fact thatBF and HQ search for the shortest path on all avail-able wavelengths, but FF finds the shortest path, thentries to find an available wavelength on that path.

Next we look at the total blocking probability5 ofprotection schemes, depicted in Fig. 4. Consideringthe lower offered load values, the lit backup protectionscheme has much worse performance than the darkbackup one. This is due to the increased interferencecaused by lighting up the backup paths, and conse-quently deteriorating the quality of newly arrivingconnections. The lower blocking probability of thedark backup scheme can be weighted against itsslower traffic restoration compared with the litbackup scenario.6 We can also see that while the linkprotection scheme has a better wavelength blockingprobability, it exhibits much higher blocking probabil-ity when we take QoT into account. This is due to thefact that in this scenario the route of a connection is tobe found in one or the other of the two digraphs; there-fore it may not be the shortest path. Increasing the

5Note that P�blocking�=P�wavelength blocking�+ �1−P�wavelength blocking�� P�QoT blocking�. Thus, the total blockingprobability is not the sum of the wavelength and QoT blockingprobabilities.

6Restoration time analysis is out of the scope of this paper.

2 4 6 8 10 12 14 16 18 2010

−4

10−3

10−2

10−1

100

Load (Erlang)

Blo

ckin

gP

roba

bilit

y

Link ProtectionFF, Lit Backup Path ProtectionFF, Dark Backup Path ProtectionBF, Lit Backup Path ProtectionHQ, Lit Backup Path ProtectionHTQ, Lit Backup Path ProtectionBF, Dark Backup Path ProtectionHQ, Dark Backup Path Protection

Fig. 4. Total blocking probability versus traffic load for protectionalgorithms.

Authorized licensed use limited to: The George Washington University. Download

ath length leads to more noise and crosstalk in morentermediate nodes and finally lower quality of theignal at the receiver.

In this graph, we also consider the cross-layer (oroT-aware) path protection algorithm HQ, and as we

an see, it significantly improves the performance ofoth the lit and dark backup schemes. We also seehat the HQ lit backup algorithm even outperformshe FF dark backup scheme. The best-fit dark backuplgorithm also shows very good performance, which ishe consequence of combining the low wavelengthlocking of the BF algorithm and the good perfor-ance of the dark backup scheme. At higher loads, all

lgorithms become wavelength-blocking limited, andhere is little that QoT-aware algorithms, includingQ, can do to improve performance.

The wavelength blocking probability of the restora-ion algorithms are shown in Fig. 5. The FF-FF pathestoration and link restoration have the same block-ng probability because they behave the same way inhe call setup phase. This graph shows that BF wave-ength assignment leads to much lower wavelengthlocking than FF (notice that regarding wavelengthvailability, BF and HQ have the same performance).ut as for the total blocking probability, which in-ludes QoT-blocking, we see from Fig. 6 that the BFcheme is much less advantageous. It can be seen thathe proposed HQ-HQ path restoration algorithmtrongly outperforms other algorithms in the presencef physical layer impairments, which supports thedea of using cross-layer approaches in all-optical net-ork design. This advantage comes at a cost of in-

reased time complexity, which is discussed later inhe paper.

. Performance Evaluation: Vulnerability

We now look at the vulnerability ratio as an indica-ion of the capability of these algorithms to recover

8 10 12 14 16 18 2010

−4

10−3

10−2

10−1

Load (Erlang)

Wav

elen

gth

Blo

ckin

gP

roba

bilit

y

FF−FF (Path,Link) Restoration(HQ−HQ,BF−BF) Path Restoration

ig. 5. Wavelength blocking probability versus traffic load for res-oration algorithms.

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 10: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

smfbwisftpgskteoHpoptspaj

nmnmmhppncashpFriro

ptrtlQtfpc

328 J. OPT. COMMUN. NETW./VOL. 2, NO. 6 /JUNE 2010 Askarian et al.

from a failure. Recall that the vulnerability ratio isthe probability that a random ongoing connection (atthe time of failure) cannot be restored due to unac-ceptable QoT (or even unavailability of a lightpath inthe case of restoration) if a random link fails at a ran-dom point of time during the operation of the network.

First we look at the vulnerability ratio of protectionschemes, shown in Fig. 7. When connections affectedby a failure start using their backup paths, thebackup paths for a fraction of these connections maynot have adequate QoT (the only reason why a connec-tion would not be restored in the lit backup path case)or they may even influence other lightpaths (in thecase of dark backup path protection and link protec-tion).

In the low offered load range (below 10 Erlangs), wesee that the link protection scheme has the highestvulnerability. This is again due to the long primarypaths used in this scheme, which makes them highlyexposed to the interference caused by lighting up thebackup path for the failed link. In this load range, we

2 4 6 8 10 12 14 16 18 2010

−4

10−3

10−2

10−1

Offered Load (Erlang)

Blo

ckin

gP

roba

bilit

y

FF−FF (Path,Link) RestorationBF−BF Path RestorationHQ−HQ Path Restoration

Fig. 6. Total blocking probability versus traffic load for restorationalgorithms.

2 4 6 8 10 12 14 16 18 2010

−3

10−2

10−1

100

Load (Erlang)

Vul

nera

bilit

yR

atio

Link ProtectionBF, Lit BackupHQ, Lit BackupFF, Lit BackupBF, Dark BackupHQ, Dark BackupFF, Dark Backup

Fig. 7. Vulnerability ratio versus traffic load for protectionalgorithms.

Authorized licensed use limited to: The George Washington University. Download

ee that again the HQ algorithm improves the perfor-ance due to its capability of providing a higher Q

actor margin for each connection in the setup phasey carefully spreading wavelengths out over the net-ork. When a random link failure happens, although

nterference may increase due to the protectioncheme, the path signal quality represented by its Qactor could still remain above the required threshold,hough decreased. From another perspective, HQ im-roves the vulnerability ratio over FF by a larger mar-in for the lit backup scheme than for the dark backupcheme. This is because in the lit backup scheme, HQnows the backup path’s signal quality, so that it canake measures to alleviate its interference with oth-rs. In the dark backup scheme, the backup path is litnly when a failure happens and it is impossible forQ to predict the interference that the dark backupaths are going to introduce when they are lit. More-ver, we see that the BF algorithm provides no im-rovement over the FF algorithm, due to the fact thathe non-QoT-aware BF RWA is not able to intelligentlypread out the connections over the network, as op-osed to HQ. This shows that cross-layer algorithmsre capable of reducing the vulnerability more thanust through a more efficient wavelength assignment.

We see from Fig. 7 that, as the offered load to theetwork increases toward 9 Erlangs and there areore ongoing connections spread out over the entire

etwork, the HQ algorithm cannot improve the perfor-ance by finding better paths because there is toouch interference on all candidate paths. Indeed, at

igher offered loads, HQ has a negative impact. Thishenomenon can be explained by the fact that HQ, byroviding lower blocking probability, admits more con-ections into the network, which in turn greatly in-reases the vulnerability of the network to a failure,nd since there are already too many connectionspread all over the network, HQ cannot be of muchelp in the recovery phase. The same argument ex-lains the higher vulnerability of BF compared withF as the load increases. In general, in these loadanges, the network does not have a good performancen terms of both blocking probability and vulnerabilityatio and the network operators should avoid theseperating regions.

The vulnerability ratio for restoration algorithms islotted in Fig. 8. One interesting observation here ishat even the non-QoT-aware path restoration algo-ithms have vulnerability ratios at the same level ashe dark backup path protection and significantlyower than the lit backup protection schemes (eitheroT-aware or non-QoT-aware schemes; see Fig. 7). In

heory we can argue that by reserving some resourcesor protection, we may encounter higher blockingrobability, but the network can be guaranteed to re-over from failure. But we see that in fact where the

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 11: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

D

tbdpp

htraatsp4m(ts

trgasa

r

mtpcaoorc

Fa

Askarian et al. VOL. 2, NO. 6 /JUNE 2010/J. OPT. COMMUN. NETW. 329

physical layer has impairments, this argument is nottrue and a naive approach to resource reservationwould even increase the vulnerability of the networkto failure. In the same way, we can see that link res-toration has lower vulnerability compared with linkprotection.

These results also show that adding QoT awarenessto the path restoration scheme by using HQ for initialcall setup and finding the restoration path, in otherwords using HQ-HQ path restoration, further de-creases the vulnerability ratio of the network signifi-cantly. It is obvious that the restoration algorithmsmay not be able to recover the affected connectionsfrom a failure due to the unavailability of wave-lengths. The probability of such an event, called thewavelength vulnerability ratio, is shown in Fig. 9.Note that this probability is zero for the protection al-gorithms (because backup wavelengths are reservedat the time of call setup) and is included in the com-putation of the vulnerability ratio shown in Fig. 8 forthe restoration algorithms.

2 4 6 8 10 12 14 16 18 2010

−3

10−2

10−1

100

Offered Load (Erlang)

Vul

nera

bilit

yR

atio

Link RestorationFF−FF Path RestorationBF−BF Path RestorationHQ−HQ Path Restoration

Fig. 8. Vulnerability ratio versus traffic load for restorationalgorithms.

2 4 6 8 10 12 14 16 18 2010

−7

10−6

10−5

10−4

10−3

10−2

10−1

Load (Erlang)

Vul

nera

bilit

yR

atio

(wav

elen

gth) Link Restoration

FF−FF Path RestorationHQ−HQ Path RestorationBF−BF Path Restoration

Fig. 9. Wavelength vulnerability ratio versus traffic load for resto-ration algorithms.

Authorized licensed use limited to: The George Washington University. Download

. Performance of the High-Speed Algorithms

So far we have seen the advantages of HQ in pro-ection and restoration schemes in terms of bothlocking probability and vulnerability ratio. As wasiscussed earlier, the drawback of HQ is its high com-utation time. Next we look at the performance of theroposed high-speed algorithms.

Figure 10 shows the blocking probability for theigh-speed algorithms proposed in Section IV. We seehat the HQ-FF and HQ-BF as well as HQ-HQ pathestoration algorithms enjoy a very low blocking prob-bility. Note that these three algorithms behave ex-ctly the same way in the call setup phase; thereforehey have the same blocking probability. Also we ob-erve that SPALW and MC both have lower blockingrobability than FF path protection schemes (see Fig.) and in the higher traffic load region, their perfor-ance is similar to the HQ path protection schemes

in dark or lit backup cases, respectively; the curve forhe dark backup case is repeated here for compari-on).

Looking at the vulnerability ratio in Fig. 11, we seehat the HQ-BF and HQ-FF path restoration algo-ithms, SPALW, MC dark backup path protection al-orithms perform in the same range and have vulner-bility ratios close to that of the HQ dark backupcheme, while lit backup path protection algorithmsre more vulnerable to a failure.

Figure 12 shows the processing time7 for the resto-ation algorithms. Here we can see the high time com-7The processing time here is the total simulation time. As wasentioned earlier, in order to measure the vulnerability ratio, any

ime the network state changes (either a connection arrival or de-arture occurs) we fail all the links in the network one by one andompute the number of ongoing connections that cannot be restorednd then average over all link failures. Thus, this measured time isnly for comparing the time complexity of the algorithms to one an-ther and does not directly measure call setup times of each algo-ithm. The simulation is run on an unloaded general purposeomputer.

2 4 6 8 10 12 14 16 18 2010

−4

10−3

10−2

10−1

100

Load (Erlang)

Blo

ckin

gP

roba

bilit

y SPALW, Lit Backup Path ProtectionMC, Lit Backup Path ProtectionSPALW, Dark Backup Path ProtectionMC, Dark Backup Path ProtectionHQ, Dark Backup Path Protection(HQ−HQ,HQ−FF,HQ−BF) Path Restoration

ig. 10. Blocking probability versus traffic load for high-speedlgorithms.

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 12: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

sccsrh

nvltgppfflmw

awaat

sflwccf

Ap(Yb0[

330 J. OPT. COMMUN. NETW./VOL. 2, NO. 6 /JUNE 2010 Askarian et al.

plexity of HQ-HQ path restoration and that the com-pound algorithms produce significant improvementsand are almost as fast as non-QoT-aware algorithms.Figure 13 shows the processing time for the protection

2 4 6 8 10 12 14 16 18 2010

3

104

105

106

107

Offered Load (Erlang)

Pro

cess

ing

Tim

ein

seco

nds

(500

0co

nnec

tions

)

HQ−HQ Path RestorationHQ−BF Path RestorationHQ−FF Path RestorationBF−BF Path RestorationFF−FF Path RestorationLink Restoration

Fig. 12. Time complexity versus traffic load for restorationalgorithms.

2 4 6 8 10 12 14 16 18 2010

−5

10−4

10−3

10−2

10−1

100

Load (Erlang)

Vul

nera

bilit

yR

atio

MC, Lit Backup Path ProtectionSPALW, Lit Backup Path ProtectionMC, Dark Backup Path ProtectionHQ−FF Path RestorationHQ−BF Path RestorationSPALW, Dark Backup Path ProtectionHQ, Dark Backup Path ProtectionHQ−HQ Path Restoration

Fig. 11. Vulnerability ratio versus traffic load for high-speedalgorithms.

2 4 6 8 10 12 14 16 18 2010

3

104

105

Offered Load (Erlang)

Pro

cess

ing

Tim

ein

seco

nds

(500

0co

nnec

tions

)

HQ, Dark Backup Path ProtectionMC, Dark Backup Path ProtectionLink ProtectionSPALW, Dark Backup Path ProtectionFF, Dark Backup Path Protection

Fig. 13. Time complexity versus traffic load for protectionalgorithms.

Authorized licensed use limited to: The George Washington University. Download

chemes. Here also we see the improvements in timeomplexity caused by the SPALW and MC algorithmsompared with the HQ path protection. These resultshow the performance trade-offs between these algo-ithms and, depending on the particular application,ow the right one should be chosen.

VI. CONCLUSIONS

In this work, the problem of survivable all-opticaletwork design has been addressed from a newiewpoint—namely, the consideration of physicalayer impairments—that raised some questions abouthe validity of the traditional approaches in next-eneration all-optical networks. We saw some unex-ected results, such as similar vulnerability ratios forath protection and restoration schemes, which stemrom neglecting the physical layer impacts on the per-ormance of higher-level algorithms. The new cross-ayer algorithms proposed in this work significantly

itigate the physical layer impairment effects on net-ork survivability.

The QoT-aware algorithms, if not designed deliber-tely, can be computationally intensive. This issueas also addressed in this work and compound QoT-ware setup/non-QoT-aware restoration algorithmsnd two simpler yet powerful QoT-aware path protec-ion algorithms were proposed.

Our work can be extended to include other schemes,uch as shared path protection. The optimizationramework in this case should consider the physicalayer characteristics in order to lead to algorithmsith good performance in all-optical networks. Time

omplexity analysis for restoration algorithms thatonsiders the necessary signaling delays is also a topicor future study.

ACKNOWLEDGMENTS

. Askarian, Y. Zhai, and S. Subramaniam were sup-orted in part by the National Science FoundationNSF) under grants CNS-0519911 and CNS-0915795.. Pointurier and M. Brandt-Pearce were supportedy NSF under grants CNS-0520060 and CNS-916890. Parts of this paper were presented in17,26–28].

REFERENCES

[1] J. Strand, A. Chiu, and R. Tkach, “Issues for routing in the op-tical layer,” IEEE Commun. Mag., vol. 39, no. 2, pp. 81–87,Feb. 2001.

[2] D. Zhou and S. Subramaniam, “Survivability in optical net-works,” IEEE Network Mag., vol. 14, no. 6, Nov./Dec. 2000.

[3] S. Ramamurthy and B. Mukherjee, “Survivable WDM mesh

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.

Page 13: Cross-Layer Approach to Survivable DWDM Network Designsuresh/Publications_files/2010... · Amir Askarian, Yuxiang Zhai, Suresh Subramaniam, Yvan Pointurier, and Maïté Brandt-Pearce

Askarian et al. VOL. 2, NO. 6 /JUNE 2010/J. OPT. COMMUN. NETW. 331

networks, part I—protection,” in IEEE Conf. on ComputerCommunications (INFOCOM), 1999, pp. 744–751.

[4] S. Ramamurthy and B. Mukherjee, “Survivable WDM meshnetworks, part II—restoration,” in IEEE Int. Conf. on Commu-nications (ICC), 1999, pp. 2023–2030.

[5] Y. Pointurier, M. Brandt-Pearce, S. Subramaniam, and B. Xu,“Cross-layer adaptive routing and wavelength assignment inall-optical networks,” IEEE J. Sel. Areas Commun., vol. 26, no.6, pp. 32–44, Aug. 2008.

[6] B. Ramamurthy, D. Datta, H. Feng, J. Heritage, and B.Mukherjee, “Impact of transmission impairments on the tele-traffic performance of wavelength-routed optical networks,” J.Lightwave Technol., vol. 17, no. 10, pp. 1713–1723, Oct. 1999.

[7] J. Martins-Filho, C. Bastos-Filho, E. Arantes, S. Oliveira, L.Coelho, J. de Oliveira, R. Dante, E. Fontana, and F. Nunes,“Novel routing algorithm for transparent optical networksbased on noise figure and amplifier saturation,” in IEEE Int.Microwave and Optoelectronics Conf. (IMOC), vol. 2, 2003, pp.919–923.

[8] D. Penninckx and C. Perret, “New physical analysis of 10-Gb/stransparent optical networks,” IEEE Photon. Technol. Lett.,vol. 15, no. 5, pp. 778–780, May 2003.

[9] A. Jukan and G. Franzl, “Path selection methods with multipleconstraints in service-guaranteed WDM networks,” IEEE/ACM Trans. Netw., vol. 12, no. 1, pp. 59–72, Feb. 2004.

[10] J. He and M. Brandt-Pearce, “RWA using wavelength orderingfor crosstalk limited networks,” in Optical Fiber Communica-tions Conf. and Expo. and the Nat. Fiber Optic Engineers Conf.,Anaheim, CA, USA, 2006, paper OFG4.

[11] T. Deng, S. Subramaniam, and J. Xu, “Crosstalk-aware wave-length assignment in dynamic wavelength-routed optical net-works,” in 1st Int. Conf. on Broadband Networks (Broadnets),2004, pp. 140–149.

[12] J. He and M. Brandt-Pearce, “Dynamic wavelength assign-ment using wavelength spectrum separation for crosstalk lim-ited networks,” in 3rd Int. Conf. on Broadband Networks(Broadnets), San Jose, CA, USA, 2006, pp. 1–9.

[13] Y. Huang, J. Heritage, and B. Mukherjee, “Connection provi-sioning with transmission impairment consideration in opticalWDM networks with high-speed channels,” J. Lightwave Tech-nol., vol. 23, no. 3, pp. 982–993, Mar. 2005.

[14] I. Tomkos, S. Sygletos, A. Tzanakaki, and G. Markidis, “Im-pairment constraint based routing in mesh optical networks,”in Optical Fiber Communication Conf. and Expo. and the Nat.Fiber Optic Engineers Conf., Anaheim, CA, USA, 2007, paperOWR1.

[15] N. Zulkifli and K. Guild, “Moving toward upgradeable all-optical networks through impairment-aware RWA algorithms,”in Optical Fiber Communication Conf. and Expo. and the Nat.Fiber Optic Engineers Conf., Anaheim, CA, USA, 2007, paper

OWR3.

Authorized licensed use limited to: The George Washington University. Download

[16] X. Yang, L. Shen, and B. Ramamurthy, “Survivable lightpathprovisioning in WDM mesh networks under shared path pro-tection and signal quality constraints,” J. Lightwave Technol.,vol. 23, no. 4, pp. 1556–1567, Apr. 2005.

[17] A. Askarian, Y. Zhai, S. Subramaniam, Y. Pointurier, and M.Brandt-Pearce, “Protection and restoration from link failure inDWDM networks: a cross-layer study,” in IEEE Int. Conf. onCommunications (ICC), Beijing, China, 2008, pp. 5448–5452.

[18] I. Chlamtac, A. Ganz, and G. Karmi, “Lightpath communica-tions: a novel approach to high bandwidth optical WANs,”IEEE Trans. Commun., vol. 40, no. 7, pp. 1171–1182, July1992.

[19] G. Agrawal, Fiber-Optic Communications Systems. Wiley,2002.

[20] B. Mukherjee, “WDM optical communication networks:progress and challenges,” IEEE J. Sel. Areas Commun., vol.18, no. 10, pp. 1810–1824, Oct. 2000.

[21] B. Xu and M. Brandt-Pearce, “Analysis of noise amplificationby a CW pump signal due to fiber nonlinearity,” IEEE Photon.Technol. Lett., vol. 16, no. 4, pp. 1062–1064, Apr. 2004.

[22] Y. Pointurier and M. Brandt-Pearce, “Study of crosstalk en-hancement by fiber nonlinearity in all-optical networks usingperturbation theory,” J. Lightwave Technol., vol. 23, no. 12, pp.4074–4083, Dec. 2005.

[23] B. Xu and M. Brandt-Pearce, “Comparison of FWM- and XPM-induced crosstalk using the Volterra Series Transfer Functionmethod,” J. Lightwave Technol., vol. 21, no. 1, pp. 40–53, Jan.2003.

[24] M. Medard, R. A. Barry, S. G. Finn, W. He, and S. Lumetta,“Generalized loop-back recovery in optical mesh networks,”IEEE/ACM Trans. Netw., vol. 10, no. 1, pp. 153–164, Feb.2002.

[25] T. Deng and S. Subramaniam, “Adaptive QoS routing in dy-namic wavelength-routed optical networks,” in 2nd Int. Conf.on Broadband Networks (Broadnets), 2005, pp. 184–193.

[26] Y. Zhai, Y. Pointurier, S. Subramaniam, and M. Brandt-Pearce,“QoS-aware RWA algorithms for path-protected networks,” inOptical Fiber Communication Conf. and Expo. and the Nat. Fi-ber Optic Engineers Conf., Anaheim, CA, USA, 2007, paperOWR5.

[27] Y. Zhai, Y. Pointurier, S. Subramaniam, and M. Brandt-Pearce,“Performance of dedicated path protection in transmission-impaired DWDM networks,” in IEEE Int. Conf. on Communi-cations (ICC), Glasgow, Scotland, 2007, pp. 2442–2347.

[28] A. Askarian, Y. Zhai, S. Subramaniam, Y. Pointurier, and M.Brandt-Pearce, “QoT-aware RWA algorithms for fast failure re-covery in all-optical networks,” in Optical Fiber Communica-tion Conf. and Expo. and the Nat. Fiber Optic Engineers Conf.,San Diego, CA, USA, 2008, paper OWN7.

ed on July 12,2010 at 16:07:38 UTC from IEEE Xplore. Restrictions apply.