Study of signaling effects on Dynamic Traffic Grooming in IP/MPLS over WDM network

Available online at www.sciencedirect.com

www.elsevier.com/locate/comcom

Computer Communications 30 (2007) 3586–3597

Study of signaling effects on Dynamic Traffic Groomingin IP/MPLS over WDM network

Sheng Chen, Gee-Swee Poo *

School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798, Singapore

Available online 13 October 2007

Abstract

Signaling is essential in a practical network for connection establishments. Previous works on Dynamic Traffic Grooming (DTG) didnot consider the signaling and related information update. The performance is poor when we incorporate the signaling requirement inprevious DTG algorithms. This shows that the effect of signaling cannot be ignored in a practical network. In this paper, we discuss theeffect of signaling on DTG, and propose a new technique called DTG-PRL to handle the DTG problem. In DTG-PRL, we divide theDTG into two steps: (1) pre-reserve some lightpaths based on statistical traffic observations and (2) dynamically groom the traffic basedon the established virtual topology. We have developed an ILP formulation and a heuristic algorithm for the purpose. The simulationresults show that the DTG-PRL outperforms previous DTG algorithms in IP Bandwidth Blocking Probability, Network Resource Uti-lization, Connection Setup Time and Control Message Efficiency. This demonstrates the usefulness of DTG-PRL in practical networks.� 2007 Elsevier B.V. All rights reserved.

Keywords: Signaling; Dynamic Traffic Grooming; Lightpath; ILP; WDM network

1. Introduction

As the Wavelength Division Multiplexing (WDM) tech-nology becomes more mature and widely used, the gapbetween the bandwidth requirement of a typical connectionrequest (e.g., OC-1, OC-3, or OC-12) and the bandwidth ofa single wavelength (capacity of a WDM channel, e.g., upto OC-192 or OC-768) becomes wider. In order to effi-ciently utilize the huge wavelength bandwidth resource,mechanisms of traffic grooming in IP over WDM networkhave been studied. By integrating low-speed traffic streamsinto high-capacity lightpaths, traffic grooming will enhancethe wavelength bandwidth utilization, improve the networkthroughput and minimize the network cost [1,2].

Problems of traffic grooming have been well studied onWDM SONET ring topology [3–5], as well as on meshWDM networks based on a static traffic demand [6]. Formore practical applications, Dynamic Traffic Grooming

0140-3664/$ - see front matter � 2007 Elsevier B.V. All rights reserved.

doi:10.1016/j.comcom.2007.10.009

* Corresponding author. Tel.: +65 67904512; fax: +65 67933318.E-mail addresses: [email protected] (S. Chen), egspoo@

ntu.edu.sg (G.-S. Poo).

(DTG) in IP over WDM network has been a researchhot spot recently. Most approaches on DTG [7–11] focusedmainly on developing algorithms and strategies for the IPtraffic grooming. The objectives include maximizing thenetwork throughput or resource utilization, minimizingthe wavelength cost, balancing the network load, and soon. However, the approaches did not consider traffic sig-naling and its influence on the grooming performance.Assumptions have been made that the signaling as well asthe related resource information update takes no time. Thisis not realistic in practical networks.

In reality, the time consumed by signaling and informa-tion update at each node will have a strong impact on thenetwork performance. In a practical network, signaling isneeded for lightpath establishment and graceful disconnec-tion. Typical signaling message processing time is about0.1 ms for a simple message and 0.3–0.4 ms for a complexmessage such as a connection request which includes aroute computation time. These timing figures wereobtained from measurements in the AT&T prototype test-bed [12]. Consequently, the IP router in each node of thetestbed [12] cannot process over 10000 events in 1 s. The

mailto:[email protected]

mailto:egspoo@ ntu.edu.sg

mailto:egspoo@ ntu.edu.sg

S. Chen, G.-S. Poo / Computer Communications 30 (2007) 3586–3597 3587

excess events that are beyond the node processing capacitywill be delayed. As time goes by, the delayed events accu-mulate and grow exponentially. They will either be dis-carded or waited in the queue depending on the queuingstrategy. Once the waiting time of IP connection requestexceeds a certain threshold, the request has to be rejectedor cancelled. On the other hand, the out-of-date link stateinformation and the delay of signaling will also lead to con-flicts in the resource allocation. From our study, we findthat the blocking performance is greatly influenced by thesignaling time and information update. New mechanismsare needed to include signaling and information updatein DTG formulation in order to provide reasonable perfor-mance of traffic grooming in real network. This has moti-vated the present work.

In this paper, we investigate the effect of signaling onDTG by comparing the performance of DTG with andwithout signaling. A strong impact will be shown that thesignaling can increase the bandwidth blocking probability.After the analysis of signaling considerations of DTG in IPover WDM network, we propose a new technique to com-bat the adverse effect of signaling. We divide the trafficgrooming in two steps, (1) pre-reserve some lightpathsbased on statistical traffic observations; (2) dynamicallygrooming the traffic with the pre-established virtual topol-ogy, where some mature algorithms could be used. We callthis scheme, Dynamic Traffic Grooming with Pre-ReservedLightpath (DTG-PRL). The traditional scheme that has noPre-Reserved Lightpaths is called Dynamic Traffic Groom-ing Basic (DTG-B). Details of the new technique will bepresented in subsequent sections.

The rest of the paper is organized as follows. The signal-ing consideration of DTG in IP over WDM network isintroduced in Section 2. The proposed DTG-PRL tech-nique together with its ILP formulation and heuristicapproach is presented in Section 3. Numerical simulationsand results are reported in Section 4. Finally, a conclusionis given in Section 5.

Node A

Node B

Node C

IP connection request:A B→

(1)

(4)

(2)

(3)

(6)

(5)

Virtual Link

Lightpath

Fiber Link

UNI

OXC

MPLS/IP router

Signaling (1)

0w

(a) No available lightpath for new request

Fig. 1. Signaling procedure for

2. Effect of signaling on DTG in IP over WDM Network

Signaling is needed for any resource allocation andrelease. In this section, we shall discuss the effect of signal-ing on DTG in the following sequence: (a) General signal-ing procedure of DTG; (b) Signaling algorithms for IPbandwidth allocation and lightpath provisioning; (c) Prob-lems of DTG due to the signaling, and finally (d) Signalingeffect as illustrated by simulation results.

2.1. Signaling Procedure in Dynamic Traffic Grooming

In IP over WDM network, there are two separate signal-ing procedures: one at IP layer and another at WDM layer.This is illustrated in Fig. 1 where a simple IP over WDMnetwork is used. We follow the node architecture asdescribed in [6]. Each node contains the functionality ofboth MPLS/IP router in IP layer and Optical Cross Con-nect (OXC) in WDM layer. The communication betweenIP and WDM layer in each node is through the User-to-Network-Interface (UNI). Two adjoining nodes are con-nected by a fiber link through the OXC component in eachof them. Each fiber link may consist of several fibers, andeach fiber may consist of several wavelength channels.

For illustration, we consider two scenarios: (1) no avail-able lightpath for the incoming connection request asshown in Fig. 1(a); (2) with available lightpaths for theincoming connection request as shown in Fig. 1(b).

Under the first scenario, when an IP connection requestarrives, e.g. IPReq: A fi B as shown in Fig. 1(a), there is noavailable lightpath for this connection request. The MPLS/IProuter in the IP Layer firstly needs to generate a lightpathrequest LPReq: A fi B to the control plane of OXC inWDM Layer through the UNI (step �). The WDM controlplane in Node A will carry out a lightpath provisioningprocedure to establish a lightpath from A to B throughthe physical fiber links. This involves the lightpath signal-ing to provide reservation and acknowledgement (step `

Node A

Node B

Node C

IP connection request:A C→

0w

1w

(1)

(2)

(b) With available lightpath for new request

Dynamic Traffic Grooming.

3588 S. Chen, G.-S. Poo / Computer Communications 30 (2007) 3586–3597

and ´). After the successful establishment of lightpathfrom A to B on wavelength w0 (LP: A fi B(w0)), theWDM control plane will send back a Lightpath Estab-lished Acknowledge (step ˆ) to the MPLS/IP router. Thusa new virtual link (VL: A fi B) is established in the virtualtopology on Node A, and the bandwidth allocation proce-dure for the IP Connection Request (IPReq: A fi B) willproceed based on the updated virtual topology. Signalingon bandwidth reservation and acknowledgement (step ˜

and ¯) will be performed.Under the second scenario, suppose a lightpath has

already been established, i.e. LP: A fi B fi C(w1) as shownin Fig. 1(b). Thus, there exists a virtual link VL: A fi C inthe virtual topology. Suppose VL: A fi C is available forthe incoming IP connection request IPReq: A fi C, onestrategy to process this request is that the MPLS/IP routerwill first find the available routes based on the virtualtopology. Then a bandwidth allocation will be directly car-ried out along the VL: A fi C, using the signaling to makereservation and acknowledgement (step � and `). In casethe VL: A fi C is not available for this connection or thebandwidth allocation procedure has failed, a new lightpathneeds to be provisioned. This requires the signaling proce-dure as described in the first scenario to be performed.

2.2. Signaling algorithms for IP bandwidth allocation and

lightpath provisioning

Various routing and signaling protocols can be imple-mented in IP bandwidth allocation and Lightpath provi-sioning. Different protocols and strategies may influencethe DTG differently. Meanwhile, for both peer modeland overlay model, the signaling algorithms will impactthe network performance significantly. For simplicity,we adopt the overlay model in IP over WDM networks[13–15], and adopt a specific routing and signaling proto-col in this study. We make use of OSPF-TE alike proto-col for the routing and RSVP-TE alike protocol for thesignaling in both IP bandwidth allocation and lightpathprovisioning. As the implementation of these protocolsin IP layer and WDM layer are quite similar, we gener-

Intermediate Nodes Source Node Node 1 Node N ... Destination Node

Connection Request

Check Global

RESERVE

RESERVENACK

Connection Failed NACK

NACK

NACK

NACK NACK

Set Devices

ACK

Connection Success

Set Devices

ACK

Set Devices ACK

RESERVE

ACKACK

ACK

Set Devices

Connection Failed

Connection Failed

(a) Resource Reservation

C

C

Fig. 2. Timing sequence o

ally call the procedures as Resource Reservation andResource Release. The signaling algorithms are illustratedin Fig. 2.

In Resource Reservation, the source node selects aroute and sends a reservation request. For lightpath pro-visioning in the WDM layer, a fixed wavelength is alsochosen. The selection of route as well as wavelength isbased on the distributed global link state table on thisnode, where the link state information of each link inthe network has been stored. For lightpath provisioning,the link state means the status of each wavelength ineach fiber link, whether it is ‘‘Occupied’’ or ‘‘Free’’.For IP bandwidth allocation, the link state means thestatus of each virtual link in the virtual topology, andthe status indicates the value of allocated bandwidthvarying from 0 to the maximum capacity of the wave-length. The reservation request is forwarded sequentiallybetween two neighboring nodes along the selected route.Each intermediate node tries to reserve the resource, for-wards the reservation message and then begins to set thedevices for the connection. The timing sequence of sig-naling procedure is illustrated in Fig. 2(a).

• The source node selects a route based on the distributedglobal link state table. For lightpath provisioning, awavelength is also selected according to a certainRWA scheme;

• The source node sends the reservation request to nextnode along the route;

• If an intermediate node is able to reserve the resource onthe link, it starts to set the device for this connection andmeanwhile forwards the reservation request to next nodealong the route, an ACK message will be sent back tothe source node as soon as the device is ready;

• If all the reservations are successful, the source willreceive all the ACK messages from the nodes alongthe route. The connection is successfully established.

• If any one of the reservations fails, this Resource Reser-vation procedure has failed and all the agreed nodes willreceive a NACK message from the failed node to releasethe reserved resources.

Intermediate Nodes Source Node Node 1 Node N ... Destination Node

onnection Stop

Request

UpdateGlobal

RELEASE

Set Devices

ACK

onnection Release Success

Set Devices

ACK Set Devices ACK

ACKACK

ACK

Set Devices RELEASE

RELEASE

(b) Resource Release

f signaling procedures.


During the Resource Release, the source node sends outa releasing message along the activated route after receiv-ing a disconnection request. The releasing message isforwarded sequentially from one node to another alongthe route. Each intermediate node releases the resource,forwards the releasing message and adjusts the deviceaccordingly, and an ACK message will be sent back tothe source node after the device adjustment. When all theACK messages from the nodes along the route come backto the source node, the connection is successfully takendown, and the global link state table in the source node willbe updated. Fig. 2(b) shows the timing sequence.

2.3. Problems of DTG due to the signaling

In practical networks, the functioning of DTG requiresthe signaling support. Although the signaling procedurecan be implemented in a variety of ways, some problemsare inherent in the process irrespective of which strategyis adopted. These problems may produce adverse effecton the network performance. We shall discuss four mostcrucial factors as follows.

2.3.1. Contention conflict

On receiving a reservation request, the source chooses asuitable route based on the link state information at thattime and sends the request. Nonetheless, the signaling,though fast, takes time. During the signaling trip, the cho-sen resource may have already been occupied or becomeunavailable. (The resource could be IP bandwidth orWDM wavelength). This situation is called contention con-flict. In distributed network, contention conflict is commonand unavoidable. However, previous works on DynamicTraffic Grooming did not consider signaling and ignorethis problem. They assume that the resource allocationcould be accomplished instantaneously. This is not realis-tic. In a practical network, especially when the traffic loadis relatively high, contention conflict can cause strongblocking which cannot be neglected.

2.3.2. Blocking due to outdated link state

In a network based on the distributed link state, such asOSPF, the source node may choose an unavailable route ifthe global link state in it is outdated. This will result in ablocking. This kind of blocking is highly dependent onthe updating strategies of link state. The purpose of linkstate update is to keep the global link state table up-to-dateand identical in network nodes. In reality, the updatingtakes time to cover the entire network. Therefore, theblocking due to outdated link state is unavoidable. Howand when to update and spread the information effectivelyare key issues to be handled. Frequent update will keepinformation afresh but the overheads of control messageswill be high. Infrequent update will have less overheadsbut the information is likely to be out-dated. Differentstrategies will produce different effects on DTG. This

problem is rather complex and requires some in-depthstudy in future.

2.3.3. Blocking due to node processing delay

The node processing delay has an effect on blocking. Ifthe node capacity is low, the excess events that are beyondthe node processing capacity will either be discarded orqueued depending on the strategy adopted. Once the wait-ing time of IP connection request exceeds a certain thresh-old that the user can endure, the request has to becancelled. As the traffic load increases, the blocking willbecome worse. This is so despite of the fact that somebandwidth resources may still be available. Ways to allevi-ate the problem include optimizing the queuing policy,decreasing control message overheads, enhancing the pro-cessing speed which is hardware dependent, and so on.

2.3.4. Double-layer signaling

For DTG, the signaling procedures are more complexdue to double-layer signaling, i.e. IP layer and WDMlayer signaling. Signaling in WDM layer is responsiblefor lightpath provisioning. Signaling in IP layer is neededfor bandwidth allocation. As shown in Fig. 1, the signal-ing procedures in these two layers are independent. Thismeans independent signaling protocols are used in eachlayer. Nonetheless, as the IP bandwidth allocationdepends on the virtual topology, the blocking probabilityof the IP connection requests is influenced by the varia-tion of the virtual topology, i.e. the establishment andreleasing of lightpaths. One key strategy of achieving ahigh IP bandwidth utilization is to reduce the blockingprobability of lightpath connections and maintain a rela-tively stable virtual topology.

2.4. Signaling Influence on DTG

In order to find out the effect of signaling on DTG, wecarry out a performance comparison between DTG-B with-out signaling and DTG-B with signaling. We evaluate theperformance by IP Bandwidth Blocking Probability (IBBP)and Network Resource Utilization (NRU). For a fair com-parison, a so-called Basic Strategy is applied, i.e. no retry isattempted. We assume that the event queue at each node isinfinite. The information update threshold is set to 1, whichmeans as soon as the link state information changes, anupdate message will be spread to the entire network. Moredetail information about the simulation is referred to Section4, and the topologies are shown in Fig. 4 in that section.

Fig. 3 shows the comparison results. It can be seen thatthe DTG-B without signaling has lower IBBP and higherNRU in both Simple6 and NSFNet networks. As the sim-ulations are conducted in exactly the same environment,the difference is only due to the signaling effect. ForDTG-B without signaling, the real-time link state updat-ing (threshold = 1) is assumed to be perfect, that will con-tribute to good performance. However, for DTG-B withsignaling, the real-time link state updating takes time,

0 50 100 150 2000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

IP Traffic Load for Node-Pair (Erlang)

IPB

andw

idth

Blo

ckin

g P

roba

bili

ty DTG-B without signalingDTG-B with signaling

(a) IBBP vs. TL, Simple6

0 10 20 30 40 50 60IP Traffic Load for Node-Pair (Erlang)

IPB

andw

idth

Blo

ckin

g P

roba

bili

ty DTG-B without signalingDTG-B with signaling

(b) IBBP vs. TL, NSFNet

0 50 100 150 2000

5

10

15

20

25

30

35

40


Net

wor

kR

eso

urce

Util

iza

tion

(%)

DTG-B without signalingDTG-B with signaling

(c) NRU vs. TL, Simple6

0 10 20 30 40 50 600

5

10

15

20

25


Net

wor

kR

eso

urce

Util

iza

tion

(%)

DTG-B without signalingDTG-B with signaling

(d) NRU vs. TL, NSFNet

Fig. 3. Performance comparison between DTG-B without signaling and DTG-B with signaling.


which aggravates blocking performance. Even at a lowload, the IBBP increases sharply and the NRU achievesits upper bound because of the strong processing delaymainly caused by a large number of updating messages.This demonstrates that the signaling and informationupdate are key factors affecting the DTG performanceand should be studied properly.

In summary, we have explained the signaling procedureand protocols, as well as the signaling influence on DTG inthis section. The simulation results show that the perfor-mance of DTG is greatly influenced by the signaling. Assignaling is required in a practical network, new mecha-nisms are needed to combat the negative impact of signal-ing. In the next section, we shall present the proposedmechanism, DTG-PRL, to solve the problem.

3. Dynamic Traffic Grooming with Pre-Reserved Lightpath

3.1. New algorithm For DTG with signaling

In this paper, we propose a new DTG-PRL mechanismto reduce the negative effect of the signaling. We make useof a pre-reserved virtual topology to carry out the DynamicTraffic Grooming. Two steps are involved: (1) pre-reservesome lightpaths based on statistical traffic observationsand (2) dynamically groom the traffic with the pre-estab-lished virtual topology.

The concept of pre-reserving resources has been appliedin lightpath protection and restoration [16]. The idea of an

envelope based on reserving wavelength links has beendeveloped in [16], which provides the provisioning overprotected capacity, rather than the provisioning to caterfor services. We extend this concept and apply it to theproblem of Dynamic Traffic Grooming with signaling. Inour approach, the lightpath is used as the minimum unitof the resources that will be pre-reserved. This differs fromprevious envelope concept, which uses separate wavelengthlinks as the minimum unit of resource. The reason for thisapproach lies in two aspects.

First, step (1) will reduce the conflicts brought in by theprocedure of lightpath establishment, as the Pre-ReservedLightpaths are carefully selected. Once reserved, the light-paths will not be released unless the statistical propertiesof the traffic are changed. However in DTG-B, the virtualtopology has to be established dynamically. Under low IPtraffic load, lightpaths are setup and released frequentlywhich is likely to increase the blocking probability of light-path provisioning. Under high IP traffic load, dynamiclightpath establishment faces resource shortage and theblocking probability becomes high. We hope a well pre-planned virtual topology will eliminate these drawbacks.

Second, as the virtual topology is well established, thecontrol messages for signaling in WDM layer and link stateupdating will be reduced correspondingly. Lower blockingprobability and shorter connection setup time are expectedin the DTG-PRL mechanism.

In this way, the Dynamic Traffic Grooming problem inpractical network can be divided to two sub-problems: (1)


How to select the Pre-Reserved Lightpaths, i.e. pre-estab-lish the virtual topology? (2) How to dynamically groomthe traffic on the established virtual topology?

Sub-problem (1) is a virtual topology design problemand some algorithms have been developed in previousapproaches (e.g. [6]) for static traffic grooming. ForDynamic Traffic Grooming, applying existing static algo-rithms is inappropriate. The challenging problems arehow to generate a well-balanced traffic matrix that matchesthe distribution of dynamic traffic load, and how to obtainthe optimal virtual topology suitable for the Dynamic Traf-fic Grooming.

We solve the sub-problem (1) by proposing an ILP formu-lation as well as a heuristic algorithm for scalable networks.As the Pre-Reserved Lightpaths will not consume all theresources in the whole network, the unused wavelength linksare still available for a DTG-B scheme. And some maturealgorithms in literature can be adopted to solve sub-problem(2). We will concentrate on solving the sub-problem (1) inthis paper, and compare the performance of DTG-PRLscheme with DTG-B scheme by means of simulation. TheILP formulation and the heuristic algorithm will be pro-posed in the following sub-sections. A comparison betweenthe ILP formulation and heuristic algorithm will be givenin the numerical analysis in Section 4.3.

3.2. ILP Formulation for DTG-PRL

How to select the Pre-Reserved Lightpaths for DTG-PRLcan be formulated as an ILP optimization problem. In thisoptimization procedure, both throughput and wavelengthcost are considered. The throughput is defined as the total

able 1P Notations

,n denote the endpoints of a physical fiber link that might occur in a lightpath;j denote the originating and terminating nodes for a lightpath. A lightpath may traverse single or multiple physical fiber links;d denote the source and destination of the end-to-end traffic request. The end-to-end traffic may traverse through a single or multiple lightpaths;

denotes the granularity of low-speed traffic requests. We assume x 2 {1,3,6,12,24,48}, which means that traffic demands between node pairs canbe any of OC-1, OC-3, OC-6, OC-12, OC-24 and OC-48;denotes the index of OC-x traffic request for any given node pair (s,d). For instance, if there are 5 OC-x requests between node pair (s,d), thenr 2 [1,5].denotes the number of nodes in the network;denotes the number of wavelengths per fiber. We assume all of the fibers in the network carry the same number of wavelengths;denotes the maximum capacity of each wavelength. We use OC-1 as the unit;

mn denotes the number of fiber interconnecting node m and node n. For our study in this paper, Fmn 2 {0,1}. Fmn = Fnm = 0 indicates no physicalconnection between node m and node n. Fmn = Fnm = 1 indicates there exists one unidirectional fiber link between these two nodes in both m fi n

and n fi m direction;x denotes the traffic matrix set of OC-x connection requests. Kx ¼ Kx

sd

� �, where Kx

sd is the number of OC-x connection requests between node pair(s,d);

wij denotes the number of lightpaths from node i to node j on wavelength w. P w

ij > 1 means the lightpaths between node i and node j on wavelength w

may take different paths;ij;wmn denotes the number of lightpath from node i to node j on wavelength w, employing fiber (m,n). Lij;w

mn 2 f0; 1g;wmn denotes whether the wavelength w is occupied on fiber (m,n) or not. 0 6 Qw

mn 6 F mn. For our study here, Qwmn ¼ 1 means w is occupied, otherwise

Qwmn ¼ 0;

sd;rij;x ksd;r

ij;x ¼ 1 means the r th OC-x low-speed traffic request from node s to node d employs lightpath (i,j) as an intermediate virtual link; otherwiseksd;r

ij;x ¼ 0;x;rsd Sx;r

sd ¼ 1 means the r th OC-x low-speed traffic request from node s to node d has been successfully routed; otherwise Sx;rsd ¼ 0.

TIL

m

i,s,x

r

N

W

C

F

K

P

LQ

k

S

traffic of successfully routed IP connection requests with dif-ferent bandwidth demands. The requested bandwidth of thelow-speed traffic (IP connections) varies and can be any oneof OC-1, OC-3, OC-6, OC-12, OC-24 and OC-48. The wave-length cost is defined as the number of occupied wavelengthson each fiber link in the entire network. We also make the fol-lowing assumptions in our study: (1) there is one fiber link ineach direction between neighboring nodes. Each fiber con-sists of 64 wavelengths; (2) no wavelength conversion is con-sidered; (3) the transceivers in a network node are tunable toany wavelength on the fiber; (4) Each node has sufficienttransceivers for input and output lightpaths; (5) The eventqueue in each node is large, that means no signaling eventswill be discarded.

The objective of the optimization procedure is to findout how the established lightpaths distribute based on agiven traffic matrix and a given network topology. Theoptimal lightpath distribution is defined as a situationwhere the established lightpaths may provide maximumtraffic throughput and occupy as few wavelength links aspossible. In order to achieve both of these objectives, weadopt a two-round ILP calculation. As maximum trafficthroughput is the main objective, it will be calculated inthe first round. In the second round to calculate the mini-mum wavelength link, the result of the first round will becounted as an additional constraint. The notations usedare indicated in Table 1.

The ILP is formulated as follows:Round 1 – The objective function aims at maximizing the

throughput:

Maximize:X

s;d;x;r

x � Sx;rsd ð1Þ

1. For each node pair (s,d), calculate the sum of unallo-

cated traffic request T(s,d) and the shortest hop distance

H(s,d) on the physical topology between (s,d);

2. For each node pair (s,d), calculate the wavelength utili-zation: WU(s,d) = T(s,d)/H(s,d);

3. Find out the maximum WUmax = T(s 0, d 0)/H(s 0,d 0);

4. Try to setup a lightpath between (s 0, d 0) using First-Fit

wavelength assignment and shortest-path routing, sub-

ject to the wavelength continuity constraint;

5. If the lightpath setup fails,, that means no lightpath can

accommodate any traffic request between (s 0,d 0), hence

set T(s 0,d 0) = 0, the traffic is blocked; otherwise, set

T(s 0, d 0) = Max[T(s 0,d 0) � C,0], where C denotes the

capacity of the wavelength.

6. Goto 1 until T(s,d) = 0 "s,d.


Constraints – on physical route variables:X

m

Lij;wmk ¼

X

n

Lij;wkn if k 6¼ i; j 8i; j;w; k ð2Þ

X

m

Lij;wmi ¼

X

n

Lij;wjn ¼ 0 8i; j;w ð3Þ

X

n

Lij;win ¼

X

m

Lij;wmj ¼ P w

ij 8i; j;w ð4ÞX

i;j

Lij;wmn ¼ Qw

mn 8m; n ð5Þ

Qwmn 6 F mn 8m; n ð6Þ

Lij;wmn 2 f0; 1g ð7Þ

Constraints – on virtual-topology traffic variables:X

i

ksd;rik;x ¼

X

j

ksd;rkj;x if k 6¼ s; d 8s; d r 2 ½1;Kx

sd � ð8ÞX

i

ksd;ris;x ¼

X

j

ksd;rdj;x ¼ 0 8s; d r 2 ½1;Kx

sd � ð9ÞX

j

ksd;rsj;x ¼

X

i

ksd;rid;x ¼ Sx;r

sd 8s; d r 2 ½1;Kxsd � ð10Þ

X

s;d;x;r

x � ksd;rij;x 6 C �

X

w

P wij 8i; j ð11Þ

Sx;rsd 2 f0; 1g ð12Þ

Round 2 – The objective function aims at minimizing thewavelength links:

Minimize :X

w

X

m;n

Qwmn ð13Þ

Additional Constraint for 2nd round:

Suppose the result of the object value in the first round isTmax, thenX

s;d;x;r

x � Sx;rsd ¼ T max 8s; d; x; r ð14Þ

The given traffic matrix Kx ¼ Kxsd

� �is randomly generated

according to the traffic model described in Section 4.2 Kxsd

denotes the average number of OC-x connection requestsbetween node pair (s,d) during the observation, and hasbeen normalized by mean service holding time.

Eq. (1) shows the optimization objective function formaximizing the traffic throughput. Eqs. (2)–(4) are themulticommodity equations (flow conservation) thataccount for the routing of a lightpath from its origin toits termination. Eqs. (5)–(7) ensure that wavelength w onone fiber link (m,n) can only be present in at most onelightpath in the virtual topology. Eqs. (8)–(12) are respon-sible for the routing of low-speed traffic requests on the vir-tual topology, and they take into account the fact that theaggregate traffic flowing through lightpaths cannot exceedthe overall wavelength capacity. Eq. (13) shows the optimi-zation objective function for minimizing the wavelengthlinks based on all the constraints from (2)–(12), as wellas additional Eq. (14) which indicates the total trafficthroughput being equal to the object value produced inthe first round.

3.3. Heuristic approach for DTG-PRL

It is well known that the traffic grooming problem in amesh network is an NP-complete problem [6]. The numberof variables and equations increases exponentially as thenode number and wavelength on each fiber increase. Con-sequently, we have to retort to heuristic approach forDTG-PRL for large networks. We develop a heuristic algo-rithm called Maximizing Wavelength Utilization (MWU)to both maximize the traffic throughput and minimize thewavelength cost – just like the ILP formulation.

The MWU algorithm follows a simple rule that thehigher wavelength utilization for the traffic between a nodepair, the higher is its priority for the resource allocation. Inthis way, MWU may find an optimal distribution of estab-lished lightpaths, i.e. large traffic is likely to be acceptedonly if the wavelength cost is as low as possible. MWU isdescribed as follows:

4. Simulation and numerical results

4.1. Simulation description

In the simulation we use the First Come First Serve(FCFS) strategy for processing each IP connection request,and First-Fit (FF) strategy for choosing the available light-path and wavelength. The common procedure for DTG,which is used to solve the sub-problem (2), can make useof any mature algorithm in previous works [7–11]. In thispaper, the performance comparison between DTG-B andDTG-PRL is the focus. Without loss of generality, weadopt a common DTG algorithm as described below forboth DTG-B and DTG-PRL schemes:

1. Get an IP connection request;

2. Find the available route based on the virtual topology, iffound, goto 3; otherwise goto 4;

3. Proceed with the signaling to allocate bandwidth

resources for this connection; if success, the connection

will be set up, goto 1; otherwise goto 4;

4. Request for a new lightpath and proceed with the signal-

ing for lightpath provisioning; if the new lightpath issuccessfully established, goto 3; otherwise the connec-

tion is blocked.


There are some differences between DTG-PRL andDTG-B when executing the above algorithm. For DTG-B, the virtual topology is empty at the beginning, and willhave to be dynamically changed as the lightpath is estab-lished and released during the simulation. However, forDTG-PRL, it has a well established virtual topology atthe beginning, and the Pre-Reserved Lightpaths will notbe released during the simulation time. Note that inDTG-PRL, the Pre-Reserved Lightpaths do not occupythe entire wavelength resources, the spare capacity is avail-able for dynamic lightpath set up just like DTG-B. Conse-quently, the virtual topology in DTG-PRL is contributedby both the Pre-Reserved Lightpaths and the dynamicallyestablished lightpaths.

4.2. Topology and assumptions

We adopt a simple 6 nodes topology, namely Simple6 inFig. 4(a), and the NSFNet in Fig. 4(b) for simulation. Thefollowing assumptions are made for both of them:

0

21

5 4

3

1

1

1 1

11

1

1

(a) Simple network with 6 nodes, 12 links

1

6

11

1

Fig. 4. Topology of (a) Simple6

NRU ¼PðSuccessful IP Connection RequestÞðReques

ðSimulation PeriodÞ �PðLinksÞ

Pð

• Each link represents two unidirectional fibers, and eachfiber contains 64 wavelengths. Wavelength capacity isOC-192 (10 Gbps). The number on the link denotesthe distance in 100 km.

• As the Pareto distribution has been widely used in thestudy of inter-Ethernet traffic [17], we adopt the Paretodistribution on the IP traffic in order to make the com-parison easier and the signaling effects highlighted. Thetraffic over WDM layer is dominated by the demandsfrom the IP layer, according to the specific routingand signaling strategies. Both the IP request arrive inter-val (1/k) and the IP service holding time (1/l) follow thePareto distribution with shape = 1.2 (typical value forEthernet traffic). Mean service holding time is 1.0 sec-ond, and traffic load for each node-pair is calculatedby Erlang (k/l).

• We set an observation period of over 100,000 IP connec-tion requests on each node. The request bandwidth fol-lows the ratio as OC-1:OC-3:OC-6:OC-12:OC-24:OC-48 = 1:1:1:1:1:1.

• Four performance metrics are employed, the definitionsof which are given below:

1) IP Bandwidth Blocking Probability (IBBP): This is theratio of the requested IP bandwidth which is not suc-cessfully accommodated to the total requested IP band-width of all IP connections:

IBBP ¼ 1�PðSuccessful IP Connection RequestÞðRequested BWÞ

PðExecuted IP Connection RequestÞðRequested BWÞ

ð15Þ

2) Network Resource Utilization (NRU): This is the ratioof effectively utilized IP bandwidth to the total usablelink capacity in the entire network during thesimulation:

0

2

54

36

16

12

28

0

11

6

8

7

9

10

11

12

13

20

20

7

24

8

8

8

79

3

35

5

(b) NSFNet with 14 nodes, 21 links

and (b) NSFNet network.

ted BW� Service Holding TimeÞ

WavlengthsÞðWavelength CapacityÞ ð16Þ


3) IP Connection Setup Time (ICST): The average timeperiod from the arrival of the IP connection requestto the time when the IP connection is successfully setup. Note that, only successful IP connections arecounted:

ICST ¼PðSuccessful IP Connection RequestÞðtsuccessfully set up � trequest arrivalÞP

ðSuccessful IP Connection RequestÞð17Þ

4) Control Message Efficiency (CME): This is the ratio ofIP connection throughput in bits to control messageoverhead in bits. CME denotes how much effectivetraffic is successfully accommodated by a certain lengthof control messages overhead. Higher CME shows bet-ter efficiency performance of the control messageoverhead:

CME ¼PðSuccessful IP Connection RequestÞðRequested BW� Service Holding TimeÞðbitsÞ

PðControl MessagesÞðLength of MessageÞðbitsÞ ð18Þ

• Different strategies on information message updatingand signaling retry policies in IP or WDM layermay impact the network performance strongly. Com-parisons among different strategies are rather complex,and need further study. For a clear and fair compar-ison between DTG-B and DTG-PRL, we adopt thefollowing strategies. The information updating thresh-old is set to 1 in the WDM layer and 1, 5, 10 in theIP layer, respectively; no retry is allowed during thesignaling in both IP and WDM layer. As the IP con-nection requests are more dynamic than the lightpathconnection requests, the control overhead in the IPlayer is heavier than that in the WDM layer. Thus,we set 1 as the updating threshold in the WDM layer,and several different values in the IP layer, in order tosimultaneously compare the influence brought in bythe information updating.

4.3. Numerical results

4.3.1. Comparison between ILP formulation and Heuristic

approach

Fig. 5 shows the comparison results of ILP formulationand heuristic algorithm MWU on the Simple6 network.The simulations are based on scenarios of DTG-PRL withsignaling. We focus only on the curves of ILP and MWUunder various conditions. It can be seen that, the curvesof MWU have a good match with those of ILP. The aver-age absolute errors of MWU compared with ILP in IBBPare only about 2.1%, 0.81% and 0.72% for U = 1, 5, 10,

respectively. And in NRU, the average absolute errorsare only about 0.55%, 0.20% and 0.17% for U = 1, 5, 10,respectively. Thus, the proposed heuristic algorithmMWU provides fairly comparable results as the ILP for-mulation. Subsequently, when we employ MWU for the

simulation of the larger network, NSFNet, the resultsshould be reliable.

4.3.2. Comparison between DTG-B and DTG-PRL

Fig. 6 shows the performance comparisons using theSimple6 and NSFNet networks. To keep the figures simple,for Simple6, we show only the ILP results whereas for

NSFNet, we show the heuristic results. We make compar-isons between scenarios of DTG-PRL and DTG-B withand without signaling in order to find the differences. Forthe strategies with signaling, the information updatingthreshold in IP layer, U, is set to 1, 5 and 10. For the strat-egies without signaling, the ICST and the CME are mean-ingless, so that only strategies with signaling are shown infigures of ICST and CME.

It can be seen that in both Simple6 and NSFNet, DTG-PRL schemes outperform DTG-B schemes under the samestrategy in all four metrics. This means DTG-PRL schemeshave lower IP bandwidth blocking probability, higher net-work resource utilization, lower IP connection setup timeand higher control message efficiency. Meanwhile, thecurves of Simple6 and NSFNet show similar trends in eachmetric.

Figs. 6(a) and (b) show the IBBP and Figs. 6(c) and (d)show the NRU, respectively. Under the same strategy, thedifference between DTG-B and DTG-PRL is only due tothe Pre-Reserved Lightpaths. The performance improve-ment by Pre-Reserved Lightpaths is obvious. For instance,when the IBBP of DTG-PRL U = 1, 5, 10 is around 10%,the IBBP of DTG-B U = 1, 5, 10 is up to 20–30% in Sim-ple6 and 20–25% in NSFNet. Besides, in NSFNet forexample, the highest NRU of DTG-B U = 1, 5, 10 arearound 5%, 11%, 15% at 15, 30 and 40 Erlang respectively,whereas the highest NRU of DTG-PRL U = 1, 5, 10 arearound 7%, 13%, 17% at 20, 40 and 50 Erlang, respectively.This illustrates that the gain from DTG-PRL is up to 40%,20% and 13% for strategies of U = 1, 5, 10 in comparisonwith the DTG-B in NSFNet.

0 50 100 150 2000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8


IPB

an

dw

idth

Blo

ckin

g P

rob

ab

ility ILP, U=1

Heu, U=1ILP, U=5Heu, U=5ILP, U=10Heu, U=10

(a) IBBP vs. TL

0 50 100 150 2000

5

10

15

20

25

30


Net

wo

rk R

eso

urc

e U

tiliz

atio

n (%

) ILP, U=1Heu, U=1ILP, U=5Heu, U=5ILP, U=10Heu, U=10

(b) NRU vs. TL

0 50 100 150 20010

-4

10-3

10-2

10-1

100


IPC

onne

ctio

n S

etu

p T

ime

(s)

ILP, U=1Heu, U=1ILP, U=5Heu, U=5ILP, U=10Heu, U=10

(c) ICST vs. TL

0 50 100 150 2000.5

1

1.5

2

2.5

3

3.5

4x 10

12


Co

ntro

l Me

ssa

ge

Effi

cie

ncy

ILP, U=1Heu, U=1ILP, U=5Heu, U=5ILP, U=10Heu, U=10

(d) CME vs. TL

Fig. 5. Comparison between ILP and MWU with various updating threshold, U values.


Fig. 6(e) and (f) show the ICST and Fig. 6(g) and (h)show the CME, respectively. As the load increases, theICST of DTG-B decreases first and increases later, andthe CME of DTG-B increases first and decreases later.Nevertheless, the ICST and CME of DTG-PRL maintainconsistent increasing and decreasing trend respectively.This can be explained as follows. In DTG-PRL, some care-fully selected lightpaths are pre-reserved and will not bereleased, whereas in DTG-B, the lightpaths are establishedor released as soon as they are requested or freed. At lowload, lightpaths vary frequently in DTG-B, which increasesthe ICST. As the load increases, established lightpaths arenot easily released because they are likely to be occupiedfor a longer period. The lightpath variation decreases, aswell as the ICST. As the load increases further, the ICSTincreases again due to the processing delay. For DTG-PRL, ICST is barely influenced by the lightpath variation,so that it will increase consistently. For CME, the reason issimilar. High lightpath variation induces more control mes-sage overhead, which lowers the control message efficiency.However, as load increases, the effective throughputdecreases, so does the CME. It can also be shown thatthe difference of ICST between DTG-B and DTG-PRL isbigger in NSFNet. This may be concluded that in a morescalable network, the average time saving by Pre-ReservedLightpaths during the IP connection setup could be more.

In addition, it can be observed that IP layer informationupdating threshold has a distinct impact on the network

performance. For either DTG-B or DTG-PRL, the sce-nario without signaling has the best performance becauseof its unrealistic assumption that the information updatingand signaling takes no time. Among the scenarios with sig-naling, the U = 10 strategy has the best performance andU = 1 strategy has the worst performance. The reasoncan be explained as follows. Generally, a lower updatingthreshold means a more frequent updating, thus the virtuallink (lightpath) information in the IP layer should beexpected up-to-date. However, on the contrary, the fre-quent updating of virtual link information brings a highercontrol message overhead, which tends to congest the sig-naling procedure of connection establishment, and eventu-ally results in high blocking. The appropriate setting of theupdating threshold is a complex issue. A good strategy isneeded here.

5. Conclusion

Previous research works on Dynamic Traffic Groomingdid not consider the effect of signaling and related informa-tion update. The performance is poor when we incorporatethe signaling in previous DTG algorithms. This shows thatthe effect of signaling cannot be ignored in practical net-works. In this paper, we discuss the effect of signaling onDTG and propose a new technique called DTG-PRL tosolve the DTG problem in practical network. DTG-PRLconsists of two steps, (1) Pre-reserve some lightpaths based

0 20 40 60 80 100 120 140 160 180 2000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9


IP B

an

dw

idth

Blo

ckin

g P

rob

ab

ility

IP B

an

dw

idth

Blo

ckin

g P

rob

ab

ilityNS,U=1,DTG-B

NS,U=1,DTG-PRLWS,U=1,DTG-B

WS,U=1,DTG-PRL

WS,U=5,DTG-B

WS,U=5,DTG-PRLWS,U=10,DTG-B

WS,U=10,DTG-PRL

(a) IBBP vs. TL, Simple6

0 10 20 30 40 50 60


NS,U=1,DTG-B


WS,U=1,DTG-PRL

WS,U=5,DTG-B


WS,U=10,DTG-PRL

(b) IBBP vs. TL, NSFNet

0 20 40 60 80 100 120 140 160 180 2000

5

10

15

20

25

30

35

40


Ne

two

rk R

eso

urc

e U

tiliz

atio

n (

%) NS,U=1,DTG-B


WS,U=1,DTG-PRL

WS,U=5,DTG-B


WS,U=10,DTG-PRL

(c) NRU vs. TL, Simple6

0 10 20 30 40 50 600

5

10

15

20

25


Net

wo

rk R

eso

urc

e U

tiliz

atio

n (%

) NS,U=1,DTG-B


WS,U=1,DTG-PRL

WS,U=5,DTG-B


WS,U=10,DTG-PRL

(d) NRU vs. TL, NSFNet

0 20 40 60 80 100 120 140 160 180 20010

-4

10-3

10-2

10-1

100

10-4

10-3

10-2

10-1

101

100


IP C

on

ne

ctio

n S

etu

p T

ime

(s)

WS,U=1,DTG-BWS,U=1,DTG-PRL

WS,U=5,DTG-B

WS,U=5,DTG-PRL


(e) ICST vs. TL, Simple6

0 10 20 30 40 50 60IP Traffic Load for Node-Pair (Erlang)

IP C

on

ne

ctio

n S

etu

p T

ime

(s)


WS,U=5,DTG-B

WS,U=5,DTG-PRL


(f) ICST vs. TL, NSFNet

0 20 40 60 80 100 120 140 160 180 2000

0.5

1

1.5

2

2.5

3

3.5

4x 10

12x 10

12


WS,U=1,DTG-B

WS,U=1,DTG-PRL


WS,U=10,DTG-B

WS,U=10,DTG-PRL

(g) CME vs. TL, Simple6

0 10 20 30 40 50 600

0.5

1

1.5

2

2.5


Co

ntr

ol M

essa

ge

Effi

cie

ncy

Co

ntr

ol M

essa

ge

Effi

cie

ncy


WS,U=5,DTG-B

WS,U=5,DTG-PRL


(h) CME vs. TL, NSFNet

Fig. 6. Performance comparison on Simple6 (a, c, e, and g) and NSFNet network (b, d, f, and h) (WS-with signaling, NS-without signaling, U-Information updating threshold in the IP layer).



on statistical traffic observations; (2) Dynamically groom thetraffic based on the established virtual topology. We havedeveloped the ILP formulation and the MWU heuristic algo-rithm for the purpose. It has been shown that the results ofMWU have a good match with those of ILP. This establishesthe reliability of the MWU algorithm when applied to scal-able networks. Subsequently, we carry out the comparisonsbetween DTG-PRL and DTG-B. The results show that theDTG-PRL outperforms DTG-B when comparing the met-rics of IP bandwidth blocking probability (IBBP), networkresource utilization (NRU), IP connection setup time(ICST), and the control message efficiency (CME). Mean-while, information updating strategy is identified to be oneof the key factors affecting the performance of DynamicTraffic Grooming. This deserves further study in future.

References

[1] E. Modiano, P.J. Lin, Traffic grooming in WDM networks, IEEECommunication Magazine 39 (2001) 124–129.

[2] R. Dutta, G.N. Rouskas, Traffic grooming in WDM networks: Pastand future, IEEE Network 16 (2002) 46–56.

[3] X. Zhang, C. Qiao, On scheduling all-to-all personalized connectionsand cost-effective designs in WDM rings, IEEE/ACM Transactionson Networking 7 (1999) 435–443.

[4] A.L. Chiu, E.H. Modiano, Traffic grooming algorithms for reducingelectronic multiplexing costs in WDM ring networks, IEEE/OSAJournal of Lightwave Technology 18 (2000) 2–12.

[5] P.J. Wan, G. Calinescu, O. Frieder, Grooming of arbitrary traffic inSONET/WDM BLSRs, IEEE Journal on Selected Areas in Commu-nications 18 (2000) 1995–2003.

[6] K. Zhu, B. Mukherjee, Traffic grooming in an optical WDM meshnetwork, IEEE Journal on Selected Areas in Communications 20(2002) 122–133.

[7] K. Zhu, H. Zang, B. Mukherjee, A comprehensive study on next-generation optical grooming switches, IEEE Journal on SelectedAreas in Communications 21 (2003) 1173–1186.

[8] H. Zhu, H. Zang, K. Zhu, et al., A novel generic graph model fortraffic grooming in heterogeneous WDM mesh networks, IEEE/ACMTransactions on Networking 11 (2003) 285–299.

[9] K. Zhu, H. Zhu, B. Mukherjee, Traffic engineering in multigranular-ity heterogeneous optical WDM mesh networks through dynamictraffic grooming, IEEE Network 17 (2003) 8–15.

[10] B. Chen, W. Zhong, S.K. Bose, A path inflation control strategy fordynamic traffic grooming in IP/MPLS over WDM network, IEEECommunication Letters 8 (2004) 680–682.

[11] S. Huang, M. Bo, J. Zhang, et al., Dynamic traffic grooming withadaptive routing in optical WDM mesh networks, in: Proc. of SPIENetwork Architectures, Management, and Applications III, 6022(2005) 60222Y.

[12] M. Goyal, J. Yates, G. Li, W. Feng, Benefits of restoration signalingmessage aggregation, in: Proc. of OFC2003, vol. 1, pp. 203.

[13] C. Assi, A. Shami, M.A. Ali, Optical networking and real-timeprovisioning: an integrated vision for the next-generation Internet,IEEE Network 15 (2001) 36–45.

[14] D.O. Awduche, MPLS and traffic engineering in IP networks, IEEECommunication Magazine 37 (1999) 42–47.

[15] A. Banerjee, J. Drake, J.P. Lang, et al., IEEE CommunicationMagazine 39 (2001) 144–150.

[16] W.D. Grover, The protected working capacity envelope concept: analternate paradigm for automated service provisioning, IEEE Com-munications Magazine 42 (2004) 62–69.

[17] V. Paxson, S. Floyd, Wide area traffic: the failure of poisson modeling,IEEE/ACM Transactions on Networking 3 (1995) 226–244.

Chen Sheng received the B. Eng. and M. Eng.degrees in Engineering of Electronics & Infor-mation from Huazhong University of Science &Technology (HUST), P.R.China, in 1999 and2002, respectively. He is currently pursuing thePh.D degree in the School of Electrical & Elec-tronic Engineering, Nanyang TechnologicalUniversity, Singapore. His research interestsinclude dynamic traffic grooming, multicastoptical network, GMPLS optical network, Pol-icy-based management, etc.

Gee-Swee Poo received the M.S. degree from
Imperial College, UK and the Ph.D. degree fromUniversity of Leeds, UK. He is currently anAssociate Professor at the School of Electricaland Electronic Engineering, Nanyang Techno-logical University, Singapore. Previously, he wasan Associate Professor at the School of Com-puting, National University of Singapore, and aPrincipal Researcher at the Standard Telecom-munication Laboratories in UK. He is the authorof more than 110 technical papers in the areas of
communication networks published in international journals, conferenceproceedings and books.

His current research interests are in the areas of optical networks,GMPLS, multicast, VPN and QoS. Dr. Poo is a member of the AdvisoryEditorial Board of Computer Communications, an international journalpublished by Elsevier Science.

Study of signaling effects on Dynamic Traffic Grooming in IP/MPLS over WDM network

Documents

Transcript of Study of signaling effects on Dynamic Traffic Grooming in IP/MPLS over WDM network