Value-added proposition of the GMPLS control plane in IP optical networks

Value-added proposition of the GMPLS controlplane in IP optical networks

Qiang Song, Zhaoming Li, and Ibrahim Habib

City University of New York, New York, NY 10031

[email protected], [email protected], [email protected]

Wesam Alanqar

Sprint, 6220 Sprint PKWY, Overland Park, Kansas 66251

[email protected]

RECEIVED 8 AUGUST 2005; REVISED 14 OCTOBER 2005;ACCEPTED 21 OCTOBER 2005; PUBLISHED 6 DECEMBER 2005

We quantify the value-added proposition of the generalized multiprotocol labelswitching (GMPLS) distributed control plane vis à vis the centralized networkmanagement (CNM) plane. Our main objective is not to propose new protocolsor algorithms, but rather to provide guidance to network operators on how tomaximize the benefits of deploying a GMPLS control plane. We identify twomain goals for maximizing these benefits: (1) decreasing operations costs and (2)increasing revenues through enhancing the efficiency of existing network assetsas well as through enabling new services. Control plane functionalities suchas dynamic routing, user-initiated end-to-end connection setup, and distributedfault management in IP optical networks are analyzed to determine optimaloperating regions where their values are maximized. Extensive simulationsusing NSFNET topology have been carried out from which we have found thatthere are certain operating regions where (1) the distributed routing can improvethe efficiency of network assets by saving 15% of the number of wavelengthsper link, or increasing traffic loads by 20%. However, if the call interarrivaltime is shorter than 1 min, the efficiency (in terms of call blocking ratio) ofdistributed dynamic routing is reduced due to the OSPF convergence delay. (2)Enabling bandwidth on demand (BoD) services at finer bandwidth granularitiesconnections is more cost-effective to the operator than doing the same at coarsebandwidth granularities. (3) The connection setup delay is the limiting factorin determining the types of BoD services a carrier offers. The call blockingratio will increase quickly if the average call interarrival time is less thantens of seconds. (4) Assuming only single link failure, M:N shared protectioncan achieve 100% recovery at the expense of 10% increase in the number ofwavelengths compared with nonprotection schemes, while saving are as highas 48% wavelengths compared with 1:1 dedicated protection. © 2005 OpticalSociety of America

OCIS codes: 060.4510, 060.4250.

1. Introduction

Current optical networks, dominated by synchronous optical network/synchronous digitalhierarchy (SONET/SDH), use centralized network management for routing and connectioncontrol to provision services. A centralized database is used to maintain network inventoryand service information. The database is crucial to the process of provisioning services,since it is used to identify the availability of physical network assets such as circuits andports needed to support new services. In many cases, this process is not fully automated,

© 2005 Optical Society of AmericaJON 8407 December 2005 / Vol. 4, No. 12 / JOURNAL OF OPTICAL NETWORKING 838

owing to a lack of autodiscovery functions inside the physical network elements. This oftenleads to considerable errors in matching the information available in these databases withwhat is actually happening in the physical network. These errors in turn lead to considerablewaste in utilizing the resources of the network, thus increasing operating costs. The controlplane has autodiscovery capability that could be utilized to automate the above processesand hence decrease operating costs. There are also other important benefits that we identifyin the rest of the paper.

Today’s mode of centralized network management (CNM) operation does not allow op-erators to provide capabilities such as the rapid provisioning that is needed for new servicessuch as bandwidth on demand (BoD). This is caused by several problems, some of whichare related to the limitations of SONET rings, whereas others are due to the limitations ofcurrent generations of operations support systems (OSSs). For example, rapid provision-ing could be achieved by using OSSs; however, this requires the installation of SONETinterface cards and plug-ins in SONET add–drop multiplexers (ADMs) that will remainun-utilized in anticipation of requests for services that may or may not materialize. Thus toavoid the economical risks of keeping a large inventory of equipment underutilized; opera-tors prefer to wait for orders for services to arrive and, consequently, install the appropriateequipment and interface cards needed to provision the services. Furthermore, provisioninglarge capacity pipes (OC-12 or above) over today’s SONET rings is not an easy task, sincemost SONETs have a capacity of OC-12; hence a single OC-12 private line would con-sume the capacity of a whole ring. Therefore provisioning such services may take monthsbecause operators may need to upgrade existing infrastructure and install new equipment.

Current generations of OSSs are dominated by in-house or third-party software thatare proprietary, lack compatibility with other OSS systems, and are specifically developedfor certain technologies, or to suit specific needs of a carrier. In addition, current OSS im-plementations lack the capabilities to do auto-discovery of the installed equipments andsupported services in the network. This makes the provisioning process dependant uponmanual intervention, leading to errors in estimating the inventory of the installed infras-tructure, which in turn causes errors in deciding the available equipments for supportingrequests for new services. Scalability is also a concern in CNM approach. Although CNMallows a more optimal path computation using its centralized database, it is suitable only forsmall networks. As the network grows, the centralized approach will encounter scalabilityissues.

In an effort to overcome the above shortcomings, several standards organizations, suchas ITU, IETF, and OIF, have proposed an IP-based optical control plane based upon GM-PLS [1, 2]. It is based upon three main components: (1) the Open Short Path First withTraffic Engineering (OSPF-TE) protocol for distributed routing, (2) the Resources Reser-vation Protocol with Traffic Engineering (RSVP-TE) for signaling, and (3) the Link Man-agement Protocol (LMP) for autodiscovery of links and services. Through autodiscoveryand OSPF advertisements, each GMPLS engine within every switch builds a global topol-ogy of the network within its routing domain. Upon receiving a connection setup request,the RSVP-TE signaling process can either use hop-by-hop routing or source routing to setup an end-to-end path. This path is application-initiated and dynamic in the sense that theuser can set it up and tear it down similar to the manner in which circuit-switched telephonyis done today. Hence this dynamic nature of using communications resources makes it fea-sible for enabling the statistical sharing of communications resources among multiple usersbased upon time-varying demands. This, in turn, will make networking at large capacitiesmuch more cost effective compared with today’s mode of provisioning static fixed-capacitypipes that cannot be shared.

The promise of the control plane for optical networks is that it could enable capabilitiessuch as rapid provisioning, application-initiated call setup, and distributed fault manage-


ment by shifting or partly shifting some of the OSS functions, such as service activationand provisioning, as well as fault management to the control plane [3]. These capabilitiesare essential to enable new services, such as BoD. Compared with centralized approaches,the distributed approach is more scalable for BoD services, thus making it possible to set upand release end-to-end deterministic bandwidth pipes on a scale similar to how telephonecalls are placed today.

Despite the recent progress made by research communities in evaluating the GMPLScontrol plane performance [4–6], several important issues remain unresolved. Most of theissues pertain to quantifying the value-added proposition of the GMPLS control plane inenabling new services such as BoD compared with the traditional centralized managementplane approach. For example, what is the operating region where the impact of distributedrouting (e.g., OSPF-TE) is most valuable? What is the impact of the RSVP-TE signalingdelay on connection setup? How much benefit can be obtained by using the GMPLS controlplane for shared protection in optical mesh networks?

In the rest of the paper, we attempt to address these issues by focusing on quantifyingthe control plane’s main functionalities: routing, connection control, and failure recovery.In Section 2, we present a brief overview of the simulation methodology that we used toquantify the control plane value-added proposition. In Sections 3, 4, 5 through extensivesimulations studies, we present the results of our analysis. Finally, in Section 6, we sum-marize and conclude the paper.

2. Simulation Methodology

We used the simulation tool OPNET to build a model for an OXC node in which dynamicrouting with wavelength conversion and link recovery were simulated. Some related math-ematical analysis can be found in Ref. [7]. Our network node simulator consists of threemodules: GMPLS engine, call generator, and data collector. The GMPLS engine has twosubmodules: OSPF-TE for distributed routing and RSVP-TE for signaling. The OSPF-TEsubmodule supports OSPF Version 2 (RFC 2328), The OSPF Opaque LSA Option specifi-cation (RFC 2370), and TE Extensions to OSPF Version 2 (RFC 3630). The RSVP-TE sub-module supports RSVP—Version 1 Functional Specification (RFC 2205), RSVP—Version1 Message Processing Rules (RFC 2209), The Use of RSVP with IETF Integrated Services(RFC 2210), and GMPLS Signaling RSVP-TE Extensions (RFC 3473). The call generatorsimulates call requests from a node to other randomly selected nodes in the network. Thecall request parameters, such as call holding time, interarrival time, and requested band-width, are adjustable to simulate BoD and enhanced private line (EPL) service requests.The calls’ holding and interarrival times are both set to follow an exponential distribution.In the rest of the paper, if not specified parameters such as traffic load, call holding and in-terarrival times, always represent traffic from each node. The data collector reports the callblocking ratio (total number of calls blocked over total number of calls generated network-wide). Sufficient number of calls is generated to ensure that the call blocking ratio reachesa stable value before measurements are collected. Twenty simulations with different seedshave been used to get an accurate value for the average call blocking ratio. Table 1 lists theconfidence intervals of the call blocking ratios.

Table 1. Confidence Interval of Call Blocking RatioCall Block Ratio (mean) 0.001 0.01 0.1

Standard deviation 0.0002 0.0004 0.00295% Confidence interval 0.001±0.0001 0.01±0.0002 0.1±0.001


In this paper, we use the term distributed dynamic routing to mean that each node keepsa global view of the network topology. All the nodes are configured to be in the same OSPFsingle area routing domain. Traffic engineering (TE) database is updated as the link band-width changes. Path computation via Constraint Shortest Path First (CSPF) is done in eachnode to calculate the end-to-end path upon receiving a call request. All the calls requestthe same amount of bandwidth, i.e. a wavelength connection, and all the links have thesame number of wavelengths. Therefore, no traffic grooming is applied in the simulations.The OXC is assumed to have wavelength conversion capability. The wavelength in eachlink along the path is selected randomly from the unreserved wavelengths. There is no cen-tralized network element for path computation or TE algorithm. To evaluate the efficiencyof distributed dynamic routing, static (hop-by-hop) routing is introduced for comparison.By static routing we mean that no bandwidth information is updated; hence, each nodemaintains a static routing table to provide fixed next-hop information.

Three network topologies have been studied in our simulations. In topology 1, we usedthe NSFNET (Fig. 1) to obtain results based upon a real network topology. NSFNET has16 nodes and 25 links with an average node degree of 3. Topologies 2 (Fig. 2) and 3 (Fig. 3)are randomly generated topologies used to study a network with various node degrees. Forcomparison, both of them have the same number of nodes as that of NSFNET. Topology2 has 32 links with a node degree of 4, and Topology 3 has 40 links width a node degreeof 5. The propagation delay is considered in the simulations in all of the three topologies.The length of each link in topology 1 is set to its real value in the NSFNET (for example,Node 3 is San Diego, and node 14 is Huston). Topologies 2 and 3 both have a diameter of2000 km.

Fig. 1. NSFNET topology.

3. Distributed Dynamic Routing

In this section, we first quantify the value-added proposition of distributed dynamic routingcompared with static routing. We then analyze the impact of different call granularities onthe network utilization. Finally, we compare distributed dynamic routing with centralizeddynamic routing. By centralized dynamic routing we mean that the path computation isperformed in a centralized network element such as CNM. In addition, in the CNM ap-proach, a centralized routing database keeps track of the global topology of the networkacross multiple routing domains. The motivation for this analysis is to show the scalabilitylimitations of CNM approaches in supporting services such as BoD, as well as to furtheridentify the value-added proposition and limitations of distributed dynamic routing underdifferent traffic load scenarios and for different types of services.


Fig. 2. 16 nodes, 32 links (node deg = 4).

Fig. 3. 16 nodes, 40 links (node deg = 5).


3.A. Cost Advantages of Distributed Dynamic Routing over Static Routing

In the simulations, the number of wavelengths required to support a certain amount of thetraffic load in distributed dynamic routing and static routing approaches with different net-work topologies are compared. The required number of wavelengths is obtained by varyingthe number of wavelengths in the simulations such that the call blocking ratio is 0.01. Thecall interarrival time at each node is set to 10 min so that the OSPF convergence delay doesnot impact the call blocking ratio. The call holding time is varied from 20 to 2000 min togenerate a traffic load of 2 to 200 Erlangs. The above parameters are chosen to reflect aspecial type of BoD service called Provisioned Bandwidth (PB) where connections last fora few minutes to a few hours (e.g., a broadcast event or large data transfers encountered ineScience applications). This is in contrast to BoD services where calls’ interarrival timesare in the time range of seconds or fractions of seconds and holding times may be as smallas fractions of seconds (e.g., fast file transfers).

Two parameters are introduced to quantify the advantages of distributed dynamic rout-ing over static routing. One is the cost saving factor S which is the percentage of wave-lengths saved to accommodate a certain amount of traffic load:

S (%) = (Ns−Nd)/Ns×100%, (1)

where Ns is the number of wavelengths needed in static routing, Nd is the number ofwavelengths needed in dynamic routing.

The other parameter is the capacity improvement factor C which is the percentage ofincrease in traffic load accommodated by certain number of wavelengths:

C (%) = (Ld−Ls)/Ls×100%, (2)

where Ls is the traffic load accommodated in static routing and Ld is the traffic loadaccommodated in dynamic routing. Note that in the above definitions, one may replace thenumber of wavelengths by the number of time slots (e.g., such as in SONET TDM links).

Figure 4 shows that the number of wavelengths increases linearly as the traffic loadincreases for both dynamic and static routing. At a fixed amount of traffic load, the networkwith higher node degree (topology 3) requires fewer wavelengths than that with lower nodedegree (topologies 1 and 2). The advantages of dynamic routing compared with static, interms of number of wavelength per link, become more evident with the increase in trafficload. However, it is clear that increasing the node degree is a much more effective measurein terms of the cost-effectiveness of the network assets required to accommodate a certaintraffic load.

Figure 5 depicts the cost saving factor S of distributed dynamic routing over static rout-ing. This percentage indicates the cost savings to carriers. The network with higher nodedegrees exhibits higher savings than those with lower node degrees. For dynamic routing,a network with higher node degree means there are more alternative paths that could bechosen for routing the traffic. Figure 5 also indicates that for the same network, as the traf-fic load increases, the cost saving factor S drops. This is because in order to maintain thesame blocking ratio (0.01), the number of wavelengths has to increase to accommodate theincreased traffic load. With more wavelengths, the static routing tends to have more choicesin the wavelength selection at each hop. Therefore the performance difference between dy-namic routing and static routing tends to be narrowed. As the traffic load further increasesabove a certain value (about 30, 90, and 110 Erlangs for networks with node degrees being3, 4, and 5, respectively), the percentage of savings tends to saturate (around 15%, 16%,and 20%).

Figure 6 shows the percentage of capacity improvement in traffic load for distributeddynamic routing approach compared with the static one. This is an indication of the revenue


S = Static Routing, R = Dynamic Routing

0

50

100

150

200

250

300

0 20 40 60 80 100 120 140 160 180 200

Offered traffic load (Erlang)

No. o

f wav

elen

gths

per

link

ne

eded

Topology 1 (S) Topology 1 (D)Topology 2 (S) Topology 2 (D)Topology 3 (S) Topology 3 (D)

Fig. 4. Number of wavelengths per link versus. traffic load.

1015

202530

354045

5055

0 20 40 60 80 100 120 140 160 180 200Traff ic load per node (Erlang)

Cos

t sav

ing

fact

or S

(%

) Topology 1 (Node degree = 3)Topology 2 (Node degree = 4)Topology 3 (Node degree = 5)

Fig. 5. Wavelengths saved in distributed dynamic routing versus static routing.


increase to carriers obtained by deploying distributed dynamic routing. The percentageof capacity improvement decreases sharply as the number of wavelengths increases. Thishappens for the same reason that we have mentioned previously for Fig. 5, which is thatstatic routing has more choices in the wavelength selection at each hop as the number ofwavelengths increases. Therefore, the traffic load accommodated in static routing is closerto that of dynamic routing. For example, in NSFNET, at 7 wavelengths per link, the capacityimprovement is 100%; whereas, at 19 wavelengths per link it is only 40%. As the number ofwavelength increases above 30, the improvement in traffic load tends to be constant around20%.

20

40

60

80

100

120

140

160

0 20 40 60 80 100 120 140

No. of wavelengths per link

Capa

city

impr

ovem

ent f

acto

r C

(%)

Topology 1 (Node degree = 3)



Fig. 6. Capacity improvement in distributed dynamic routing versus static routing.

The above results imply that traffic load could increase at least 20% (i.e., 20% or moreincrease in revenue) if the distributed dynamic routing is used. Similarly, the number ofwavelengths could be reduced at least 15% (i.e., 15% or more cost reduction). Both costsavings and capacity improvement could be maximized by deploying dynamic routing innetworks with less than 30 wavelengths per link.

For simplicity, in the rest of the paper, only the NSFNET topology is used in the simu-lations.

3.B. Impact of OSPF Convergence Delay on the Distributed Dynamic Routing

The OSPF convergence delay is measured from the time a topology change occurs tillthe link state database of each node has been updated and synchronized. The measuredOSPF convergence time in NSFNET is around 10 to 15 seconds. If the network-wide callinterarrival time is shorter than the convergence time, the path computation might be donebased on outdated information.

Figure 7 shows the call blocking probability with the same traffic load (2 Erlangs for4 pairs of wavelengths), but with varying interarrival times. We can see that as the call in-terarrival time decreases, the performance of distributed dynamic routing (in terms of callblocking ratio) deteriorates. When the interarrival time is reduced down into the millisecondrange, distributed dynamic routing performs even worse than static routing. This is readilyunderstood, since when the interarrival time is in the millisecond range, the OSPF link statedatabases are nearly always outdated because of both the OSPF convergence delay and theRSVP signaling delay. Therefore, for such types of services (BoD) with extremely short


interarrival times, we need to use extremely fast signaling (preferably hardware implemen-tation) as well as other measures to decrease the OSPF convergence time (decreasing theamount of information to be flooded, small number of nodes per routing area, etc.) or keepthe link state database updated as quickly as possible.

Figure 8 shows the same trends of Fig. 7 but for traffic load/wavelengths equal to 0.75.The y axis shows the parameter delta, which is the difference between computing the callblocking ratio using distributed dynamic routing and stating routing. Positive values ofdelta imply that the call blocking probability is higher for distributed dynamic routing andvice versa. Interestingly, as the interarrival time increases above the one-minute range, thereis no additional gain achieved by using distributed dynamic routing. The gain curve staysflat, indicating that although there is a gain for using distributed dynamic routing, this gainsaturates for long interarrival times. A typical characteristic of private lines in comparisonwith BoD services is that the ratio of their interarrival times is extremely large. Hence, thiscurve tells us that for private line services, the gain of using dynamic over static tends tosaturate.

Traff ic load = 2 Erlang, 4 pair of w avelengths per link(NSFNET topology)

0

0.04

0.08

0.12

0.16

0.001 0.01 0.1 1 10

Call inter-arrival time (min), in log scale

Cal

l blo

ckin

g ra

tio

Dynamic routing Static routing

Fig. 7. Impact of OSPF convergence delay on distributed dynamic routing for calls withshort call inter-arrival times.

It is interesting to note that in Fig. 8, the 100 wavelength curve (around 10 min in-terarrival time) can be construed as an indication of private line services, whereas the 4wavelength (around 0.1 up to 1 min interarrival time) is an indication of the BoD service.For the 100 wavelength curve, if the traffic load/wavelength ratio is 0.75, the load is 75Erlangs, which means that the holding time is 750 min for an interarrival time of 10 min.Obviously, those values are indications of long-term pipes. If we examine the 4 wavelengthcurve at an interarrival time of 1 min, we obtain a holding time of 3 min. Clearly thosevalues are an indication of BoD services where both the holding times and interarrivaltimes are much less than private line services. The curve reinforces our conclusions thatthe gain achieved by deploying dynamic bandwidth for BoD services is significantly morethan that for private line services.


Traff ic load / No. of w avelengths = 0.75 (NSFNET)detla = call blocking ratio of (dynamic - static) routing

-0.08

-0.04

0

0.04

0.08

0.001 0.01 0.1 1 10Call inter-arrival time (min), in log scale

delta

4w v 10w v 100w v

Fig. 8. Differential advantage of dynamic versus static routing with different call interar-rival times.

3.C. Impact of Bandwidth Granularity

Figure 9 depicts the call blocking ratio B verses the traffic load A and the call granularity Naccording to Erlang B formula [8] in a single link point-to-point network.

B(N,A) =AN/N!

∑Nk=0Ak/k!

, (3)

where, N represents the number of channels in the link.

N: Number of channels per link

0

0.2

0.4

0.6

0.8

1

0 20 40 60 80 100Traff ic load A (Erlang)

Cal

l blo

ckin

g ra

tio B

N = 2

N = 4

N = 8

Fig. 9. Call blocking ratio versus traffic load and call granularity.

Each call requests the bandwidth of one channel. Thus, a channel could represent awavelength in the lambda switching capable network or a time slot in the TDM switchingcapable network. Clearly, traffic with finer call granularities leads to lower blocking ratiothan that with coarser granularities. In addition, higher utilization can be achieved for finergranularity calls at a certain call blocking rate.

In our simulations, to study the impact of call granularity, a link is divided into differ-ent number of equal capacity channels. Figure 10 shows that traffic with finer granularity


calls can achieve higher network utilization than traffic with coarser granularity calls. Thenetwork utilization tends to saturate as the number of channels in a link reaches 60. Furtherincrease in the number of channels will not result in a significant gain in the network utiliza-tion. This implies that the optimal granularity for each call request should be in the OC-3level for a network with OC-192 links or the number of wavelengths should be around 60in a WDM network in which the call granularity is in per wavelength level.

3.D. Distributed Versus Centralized Approaches

A simplified centralized routing agent is built to compare dynamic routing implemented ina centralized agent with the same implemented in a distributed approach. The centralizedagent has a centralized routing database that stores the global topology and TE informationand is updated periodically by the element management system (EMS) in each node. Uponreceiving a call request from the ingress node, the agent performs CSPF path computationand sends the message containing the computed path to the ingress node. After receivingthe message, the ingress node starts to use signaling protocol (e.g., RSVP-TE) to set up thepath.

0.4

0.5

0.6

0.7

0.8

0.9

0 20 40 60 80 100 120 140 160 180 200Call granularity (No. of channels per link)

Netw

ork

utili

zatio

n (%

)

Blocking ratio=0.001

Blocking ratio=0.01

Blocking ratio=0.1

Fig. 10. Network utilization versus call granularity.

Figure 11 quantifies the efficiency gain G of distributed dynamic routing over central-ized dynamic routing versus the centralized database update interval. The efficiency gain isdefined as

G(%) = (Ld−Lc)/Lc×100%, (4)

where Lc and Ld are the traffic load at 0.01 call blocking ratio in the centralized anddistributed approaches, respectively. The centralized database update interval is an indi-cation of the time delay needed to update the centralized routing table. This parameterdepends upon many factors such as the size of the network (e.g., number of nodes, links,wavelengths per link), calls’ interarrival times, as well as the type and efficiency of the al-gorithms used in updating the centralized routing database. Obviously, this is a dependantcomplex parameter which requires extensive analytical as well as empirical modeling toquantify, accurately. Our objective here is not to quantify this parameter, but rather to useit as a differentiator parameter that distinguishes centralized approaches from distributedones. Thus we use it as an independent parameter in these simulations.

The distributed approach performs better in terms of the amount of traffic load that itcan support, as long as the centralized approach has prolonged database update intervals.


4 wavelengths per link

-40

0

40

80

120

160

200

0.001 0.01 0.1 1 10 100 1000

Central database update interval (sec, in log scale)

Effic

ienc

y ga

in fa

ctor

G (%

) Inter-arrival time = 0.1 minInter-arrival time = 1 minInter-arrival time = 10 min

Fig. 11. Efficiency gain G versus centralized database update interva.l

The shorter the central database update interval, the closer is the performance of centralizedapproach to distributed approach.

As shown in Fig. 12, the efficiency gain factor varies with different call interarrivaltimes and exhibits a bell shape curve. The distributed approach only exhibits significantgain over centralized approach for certain types of traffic. Out of this range, distributedapproach shows no notable advantages. Consider the curve of a 1 min central database up-dating interval; this range is from 0.01 to 100 min call interarrival times, which falls intothe BoD services range. For calls with longer call interarrival times, which emulate the pri-vate line services, distributed approach exhibits little or no efficiency improvement. Notefor calls with very short interarrival times, e.g., less than 0.01 min, distributed dynamicrouting approach performs no better than centralized approach in terms of traffic load itaccommodated. At such short interarrival times, the CSPF path computations are based onoutdated databases due to the OSPF convergence delay. At the same traffic load, the callblocking ratio of dynamic routing is even higher than that of static routing as illustrated inSubsection 3.A. This phenomenon is further confirmed in the centralized approach wherethe network with 10 min update interval (more towards static routing) can accommodatemore traffic than that with 1 min update interval (more towards dynamic routing), suggest-ing that dynamic routing (both distributed or centralized) adds no value for such types oftraffic. Therefore, static routing is recommended for both approaches. Owing to the scala-bility issue of the centralized approach—as the network becomes larger—distributed staticrouting, i.e., hop-by-hop routing, is more efficient than CSPF-based approaches.

4. Connection Setup

We assume that a request for connection setup starts at time T1 at node 3 (see Fig. 1), andthe destination is to node 13. RSVP starts signaling setup at T1. In the worst case, if theRSVP signaling cannot find enough resources at node 3 at the time the RESV messagearrives to node 3, the call will be blocked and a TEAR message will be issued to releaseall the resources that have been already reserved along previous OXCs (i.e., the link fromnode 14 to node 9; node 9 to node 10; and node 10 to 13). Assuming a request for a secondcall arrives at time T2 at node 14 and the destination is node 9, this call will use one of theresources mentioned above (i.e., the link from node 10 to node 9). If the call interarrivaltime is very short, i.e., the time interval T2-T1 is close to, or shorter than, the first call’s


4 wavelengths per link

-50

0

50

100

150

200

0.01 0.1 1 10 100Call inter-arrival time (min, in log scale)

Effic

ienc

y ga

in fa

ctor

G (%

)

central database update interval = 1 mincentral database update interval = 10 min

Fig. 12. Efficiency gain G versus call interarrival time.

setup delay, then it is highly likely that the second call would be blocked, since that resourcemight not have been released by the first call. This will lead to a higher call blocking rates.

The connection setup up delay Ts is the time required to setup an optical path. It mainlyconsists of RSVP signaling propagation delay Tprop, path calculation delay Td , RSVP mes-sage processing delay Tp, and OXC delay Tx.

Ts = Tprop +Td +Tp +Tx. (5)

In this section we discuss the impact of the connection setup delay on the call block-ing ratio for traffic with short interarrival times (from hundreds of milliseconds to tens ofseconds). As we demonstrated previously, for this type of traffic, the OSPF convergencedelay also attributes to the higher call blocking ratio if dynamic routing is applied. To iden-tify the impact introduced only by the connection setup delay, static routing is used in thesimulations.

4.A. Impact of Propagation Delay

In the simulations of Fig. 13, only the propagation delay is considered. The study providesthe smallest blocking ratio a carrier could achieve at a fixed traffic load with different callinterarrival times. To analyze the impact of connection setup delay on traffic with shortinterarrival times, the call interarrival time is varied from 0.06 to 600 s while keeping thetraffic load constant. For the two 4-wavelength curves with different traffic loads, the callblocking ratios start to increase as the call interarrival time becomes shorter than 3 s. Sim-ilar observation is obtained with the two 8-wavelength curves. In 8-wavelength curves,however, the call blocking ratios start to increase as the call interarrival time becomes lessthan 0.6 s. The above result implies that the connection setup delay limits the calls withvery short interarrival time the carrier could provide, and in order to reduce the limitation,more wavelengths should be added.

4.B. Impact of RSVP Message Processing and OXC Delay

In Fig. 14, the RSVP message processing delay and OXC delay are considered. Since thestatic routing is used, the path computation delay is actually the routing table lookup de-lay and is ignored in our simulations. The OXC time is introduced in a serial-transaction


manner (the OXC will forward the RSVP signaling message to the next OXC only after itfinishes its cross connection).

To observe the impact introduced by only the RSVP message processing delay, theOXC time is set to 0 and the per node RSVP message process delay Tp is set to 1 and10ms, respectively. Results show that if the per node RSVP processing delay is at 1mslevel, the impact of RSVP message processing delay is quite limited. As Tp increases to10s, its impact on the call blocking ratio starts to appear as the call interarrival time is lessthan 6seconds. To investigate the impact of OXC delay, first, we set the OXC delay to50ms, and we vary the RSVP message processing delay from 0to10ms. These two curvesin Fig. 14 are almost identical, showing that the impact of the connection setup delay isnow dominated by the OXC delay. Next, we increase the OXC delay from 50msto100msand keep the RSVP message processing delay to 10ms. Consider the curve of 100ms OXCdelay for example, the impact of setup delay starts to be visible at 100 s call interarrivaltime.

W: number of w avelengths per link E: Traff ic load per node (Erlang)

0.008

0.012

0.016

0.02

0.024

0.028

0.032

0.01 0.1 1 10 100 1000Call inter-arrival time (sec, in log scale)

Cal

l blo

ckin

g ra

tio

W = 4, E = 0.75 W = 4, E = 1.00W = 8, E = 2.78 W = 8, E = 3.40

Fig. 13. The Impact of RSVP signaling propagation delay.

4 wavelengths per link, 1 Erlang per node

0.023

0.028

0.033

0.038

0.043

0.01 0.1 1 10 100 1000Call inter-arrival time (sec, in log scale)

Call

bloc

king

ratio

Tp = 0 ms, Tx = 0 msTp =1 ms, Tx = 0 msTp =10 ms, Tx = 0 msTp =0 ms, Tx = 50 msTp =10 ms, Tx = 50 msTp =10 ms, Tx = 100 ms

Fig. 14. Impact of connection setup delay with other delay parameters considered.


In summary, the connection setup delay limits the number of calls with the short in-terarrival time that the carrier could provide. The network performance in terms of callblocking ratio will significantly deteriorate as the call interarrival time is shorter than tensof seconds. The major contributors to the drop in performance are signaling propagationdelay, RSVP message processing delay and OXC delay. Among them, the current level ofOXC delay is the dominating factor.

5. Shared Protection

Each GMPLS engine in control plane has the capability to perform self-analysis and feedthe information to the network management system (NMS), which improves the accuracyand speed of the service assurance process, cuts down the provisioning time, and improvesthe carrier capabilities to meet customers’ SLA requirements by quickly identifying thesource of error and triggering restoration mechanisms. It provides the carrier with accuratemeans of automating service maintenance and repair records, as well as simplifying rootcause analysis. In addition, the GMPLS control plane makes it possible to enable the sharedprotection in a mesh network so that the network resources could be better utilized thanthose in a ring topology.

In this section, we compared the efficiency of M:N shared protection and 1:1 dedi-cated protection in terms of the required network resources, i.e. the number of wavelengthsneeded to accommodate certain amount of traffic load. In 1:1 dedicated protection, both theworking path and backup path are calculated upon receiving the connection request. Whensetting up the backup path, resources are reserved and the OXC is preconfigured along thebackup path. When a failure occurs, upon receiving the failure notification, the source nodeswitches traffic to the already established backup path. The 1:1 protection offers 100% re-covery and shortest recovery time, but the network resources are not efficiently utilized.In M:N shared protection, the working path and backup path are calculated upon receivingconnection request. In setting up the backup path, resources are reserved but the OXC is notcross connected along the backup path. Newly admitted backup paths may share the pre-reserved wavelengths of the existing backup path. When a failure occurs, upon receivingthe failure notification, the source node signals the backup path setup and cross connectsOXCs along the precalculated backup path. Since the backup path needs to be set up afterthe failure, the recovery time of shared protection is longer than the 1:1 dedicated protectionscheme.

In our simulations for M:N shared protection, we assume that only single link failureoccurs (the node failure is not considered). More complex algorithms and the cases formultiple failures can be found in Ref. [9]. To guarantee 100% recovery, the working pathand backup path should be SRLG disjoined. In addition, the backup paths can share thesame wavelengths only if their working paths are SRLG disjoined. For example, in Fig.15, the same wavelength in Link BC can be shared by the backup path 1 and 2, sincetheir working path 1 and 2 are SRLG disjoined. However, different wavelengths have to bereserved in link AB for backup paths 1 and 3 respectively, since their working paths 1 and3 use the same link AC.

In our simulations, all the wavelengths within a fiber link belong to the same SRLG,which is denoted by the link number. Each wavelength has an SRLG attribute to keep arecord of the links it is protecting. The Dijkstra algorithm is used to compute the workingand backup paths. During backup path computation, the links along the working path arepruned to guarantee that the backup path is SRLG disjoined with the working path. Foreach link, if no wavelength is prereserved for the protection purpose, an unreserved wave-length is selected and all the SRLG numbers of the working path are recorded in its SRLGattribute. Otherwise, the SRLG attribute is checked. If none of the SRLG numbers alongthe working path exists in the SRLG attribute, the working path is SRLG disjointed with


Working path 1Working path 2

Working path 3

Backup path 1

Backup path 2Backup path 3

A

B

C

D

Working path 1Working path 2

Working path 3

Backup path 1

Backup path 2Backup path 3

A

B

C

D

Fig. 15. Wavelength selection in shared protection.

the previous working paths protected by this wavelength; therefore the wavelength can beshared by the backup path. All the SRLG numbers of the working path are then added tothe SRLG attribute. Otherwise, a new unreserved wavelength is selected for the protection.

Figure 16 shows the relationship between the call blocking ratio and the number ofwavelengths per link at the fixed traffic load of 3 Erlangs per node. Clearly M:N sharedprotection has lower call blocking ratio than 1:1 dedicated protection. To compare theperformance of M:N shared protection with 1:1 dedicated protection and nonprotectionschemes in terms of number of wavelengths needed, we increase the traffic load to observethe number of wavelengths needed in each scheme while keeping the acceptable call block-ing ratio to 0.1. As indicated in Fig. 17, the number of wavelengths needed in each schemeincreases almost linearly with the increase of traffic load.

0

0.2

0.4

0.6

0.8

1

1 2 3 4 5 6 7 8 9 10

No. of Wavelengths per link

Call

bloc

king

ratio

1:1 ProtectionM:N ProtectionNon-protection

Fig. 16. Number of wavelengths per link versus call blocking ratio.

Figure 18 depicts the capacity saving factor for M:N protection compared with 1:1 pro-tection and non-protection schemes. The capacity saving factor is defined as the percentageof number of wavelengths saved in M:N protection compared with 1:1 and none-protection.For example, in Fig. 17, at 9 Erlangs offered traffic load, 12, 23, 11 wavelengths per link


are needed in M:N, 1:1 and none-protection networks, respectively. The shared protectionsaves 48% of wavelengths compared with 1:1 protection; whereas the shared protectionneeds extra 10% of wavelengths compared with none-protection. Both capacity saving fac-tors remain almost constant as the traffic load increases further. So for shared protection, itis better for the network to operate at traffic load at least of 9 Erlangs to take advantage ofthe maximum savings.

0

10

20

30

40

50

1 3 5 7 9 11 13 15 17 19

Offered traffic load per node (Erlangs)

No. o

f wav

elen

gths

ne

eded

per

link

1:1 protectionM:N protectionNon-protection

Fig. 17. Number of wavelengths needed versus traffic load.

-40

-30

-20

-10

0

10

20

30

40

50

1 3 5 7 9 11 13 15 17 19Offered traffic loads (Erlangs)

Capa

city

sav

ing

fact

or (%

)

M:N protection vs.1:1 protection

M:N protection vs. non-protection

Fig. 18. M:N shared protection performance evaluation.

6. Conclusions

It is important for carriers to quantify the value-added proposition of deploying GMPLS-based control plane. Deploying a complex control plane could result in an expensive propo-sition with little value added. The capabilities of the control plane will depend on the setof services offerings that a carrier wishes to enable. Therefore the advantages of the con-trol plane functionalities, such as distributed dynamic routing, connection control, shared


protection, rapid provisioning, autodiscovery, and distributed fault management, should beclearly quantified. In this paper, we focused on studying the first three functions and theirlimitations. Our findings through simulations on NSFNET reveal the following:

The number of wavelengths per link has an impact on the performance of distributeddynamic routing. The smaller the number of wavelengths, the better the distributed dynamicrouting performs over static routing. As the number of wavelengths per link increases above30, the performance difference between distributed dynamic routing and static routing tendsto be minimum and constant. Compared with static routing, distributed dynamic routing cansave 15% wavelengths per link, or increase the traffic load by 20%.

The OSPF convergence delay limits the type of services the carriers could provide. Asthe call interarrival time is shorter than 1min, the impact of the OSPF convergence delay onthe performance (in terms of call blocking ratio) of the distributed dynamic routing cannotbe ignored.

In bandwidth-on-demand (BoD) services, the level of call granularity plays an impor-tant role in achieving high network unitization, i.e., increases the operator’s revenues. Thereis an optimal region for the call granularity to achieve maximum utilization and keep thenetwork scalable. In NSFNET, the optimal call granularity is 1/60 of the link bandwidth.

The distributed approach has its advantages in boosting capacity over centralized ap-proach only in the BoD services, i.e., shorter interarrival times and holding times. As the in-terarrival times and holding times become longer, the distributed approach gradually losesits advantages. For EPL services, both distributed and centralized approaches exhibit al-most the same performance. If a carrier is interested only in EPLs, then doing routing via adynamic distributed approach is not a useful proposition.

The RSVP-TE message processing delay and OXC delay should be kept to a minimumto reduce the connection setup delay, which has an impact on the network call blockingratio. If the call interarrival time is within tens of seconds, the call blocking ratio willdramatically increase.

Assuming only single failure occurs, shared protection can achieve 100% recovery atthe expense of 10% extra wavelengths over the nonprotection scheme, and save up to 48%of wavelengths over the 1:1 dedicated protection.

References and Links[1] E. Mannie, “Generalized multi-protocol label switching (GMPLS) architecture, ” IETF RFC

3945, Oct. 2004.[2] E. Mannie and D. Papadimitriou, “Generalized multi-protocol label switching (GMPLS) ex-

tensions for synchronous optical network (SONET) and synchronous digital hierarchy (SDH)control, ” IETF RFC 3946, Oct. 2004.

[3] W. Alanqar and A. Jukan, “Extending end-to-end optical service provisioning and restorationin carrier networks: opportunities, issues, and challenges, ” IEEE Commun. Mag. 42(1), 52–60(2004).

[4] A. Jajszczyk, “Automatically switched optical networks: benefits and requirements, ” IEEECommun. Mag. 43(2), S10–S15 (2005).

[5] P. Iovanna, M. Settembre, R. Sabella, G. Conte, and L. Valentini, “Performance analysis of atraffic engineering solution for multilayer networks based on the GMPLS paradigm, ” IEEE J.Sel. Areas Commun. 22, 1731–1740 (2004).

[6] N. Ghani, S. Dixit, and T. S. Wang, “On IP-WDM integration: a retrospective, ” IEEE Commun.Mag. , Vol. 41, Iss. 9, (Sept. 2003) pp.42–45.

[7] Y. Luo and N. Ansari, “A computational model for estimating blocking probabilities of multi-fiber WDM optical networks, ” IEEE Commun. Lett. 8, 60–62(2004).

[8] R. B. Cooper, Introduction to Queuing Theory, 2nd ed. (North Holland, New York, 1981).[9] J. Zheng and H. T. Mouftah, “Dynamic lightpath restoration based on bidirectional initiation for

wavelength-routed WDM networks,” IEE Proc. Commun. 150, 409–413 (2003).


Documents

Value-added proposition of the GMPLS control plane in IP optical networks