Routing Wavelength and Timeslot ReassignmentAlgorithms for TDM based Optical WDMNetworks - Multi rate traffic demands

Rajalakshmi P Ashok JhunjhunwalaTelecommunications and Networking Group (TeNeT)

Department of Electrical EngineeringIndian Institute of Technology Madras

Chennai 600036, India.Phone: 91-44-22574408, Fax: 91-44-22570120Email: {raji@tenet.res.in, ashok@tenet.res.in}

Abstract-In this paper, we consider the problem of maxi-mizing the time of first lightpath request rejection, T in thecircuit-switched time division multiplexed (TDM) wavelength-routed (WR) optical WDM networks. TDM is incorporated intoWDM, to increase the channel utilization when the carried trafficdoes not require the entire channel bandwidth. In TDM-WDMnetwork, multiple sessions are multiplexed on each wavelengthby assigning a sub-set of the TDM slots to each session. Thus,given a session request with a specified bandwidth, a lightpathhas to be established by using the routing, wavelength and time-slot assignment (RWTA) algorithms. If the lightpath cannot beestablished, lightpath request rejection or call blocking occurs.As each lightpath is substantial revenue and long-lived, lightpathrequest rejection is highly unfavourable in the opitcal backbonenetworks. In this paper, we are proposing an intelligent routing,wavelength and time-slot reassignment algorithm for multi ratetraffic demands, where, when a call gets blocked, the alreadyestablished calls in the network are rerouted, wavelength andtimeslot reassigned so as to accomodate the blocked call. Sincewe are talking of slow arrivals and long holding times for thelightpaths, it is possible to do this reassignment while provisioninga new call. Simulation based analyses are used to study theperformance of the proposed reassignment algorithm. The resultsshow that the proposed reassignment algorithm can be used tomaximize the time of first call blocking, thereby accommodatingmore calls in the network before upgrading the network capacity.

I. INTRODUCTION

Wavelength-division multiplexing (WDM) in optical net-works is a promising technology to utilize the enormousbandwidth of opitcal fiber and it offers the capability ofbuilding very large wide-area networks with throughputs ofthe order of gigabits per sec for each node. Current opticaltechnology demonstrations have shown the feasibility of upto 160 channels (wavelengths), each operating at 10 Gbps,per fiber [1]. Therefore, WDM technology is very attractivefor the backbone networks. In general, topology of a WDMnetwork may be an arbitrary mesh as shown in Figure 1.Traffic between pair of nodes is carried through lightpaths.A lightpath is a high-bandwidth end-to-end circuit betweentwo nodes, occupying a wavelength on each link of the pathbetween the nodes. Nodes are equipped with optical cross

9

2

Lightpath

4) A

*

Fiber Link

. .

Fig. 1. WDM network consisting of OXC nodes interconnected by fi berlinks

connects (OXCs) which can selectively drop and add some ofthe wavelengths locally, switch wavelengths from one inputport to another (wavelength routing , WR), and are able tochange the wavelength of an incoming lightpath to any ofthe outgoing wavelengths, called the wavelength conversioncapability of the OXCs [2]. This work considers the widearea backbone network architecture based on a wavelength-routed mesh topology, in which the nodes do not possess thewavelength conversion capability. Therefore, each lightpathmust occupy the same wavelength along all the links of thepath which is known as the wavelength continuity constraint[3].

In circuit-switched WR networks, each lightpath (circuitor session) is allocated an entire wavelength for its duration.If the actual bandwidth requirements of the session are lessthan the channel bandwidth, valuable bandwidth is wasted.This motivates the need for providing fractional wavelengthcapacity to the network traffic. Provisioning fractional wave-length capacity is achieved by incorporating time division

0-7803-9746-0/06/$20.00O2006 IEEE

multiplexing into WDM [5], [6]. In TDM-WDM networks, thebandwidth of the wavelength is partitioned into fixed-lengthtime slots. A session request specifies the number of time-slots required for the session duration. Thus, multiple sessionsare multiplexed on each wavelength by assigning a sub-setof the TDM slots to each session. Similar to the wavelengthcontinuity constraint, assigning a slot on a wavelength requiresthe same time slot on all links along the path on that particularwavelength. This is referred to as the time-slot continuityconstraint. In the wide area backbone networks, due to thelink propagation and node processing delays, the time-slotallocation on successive links will be shifted. In this work,we assume the link propagation and node processing delaysare zero.

Figure 2 shows the single fiber TDM-WDM network. Here,F, the number of slots per wavelength is assumed to be 4.Lightpath between 6 and 2, which has requested for two slots,is assigned slots 0 and 1 on the first link and the second link.The link propagation and node processing delay is assumedto be 0. Similarly, the time-slot allocations for the sessionrequest between nodes 3 and 5; and nodes 2 and 4 can beseen in the figure. We consider a TDM-WDM network sup-porting multirate circuit-switched sessions. A session requestis assigned one or more time slots forming an optical circuit,for the duration of the session. The rate of the session isproportional to the number of slots assigned to it, with oneslot corresponding to a minimum rate session.

In this paper, we consider the dynamic version of the circuit-switched network model, where session requests arrive to thenetwork based on some stochastic arrival process. When anew session request arrives with a specified bandwidth, an all-optical lightpath has to be established between the source andthe destination. In order to establish the lightpath, the networkhas to determine the route, assign the wavelength and allocatetime-slots within the wavelength that meets the given request.This is referred to as Routing, Wavelength and Time-slotassignment (RWTA) problem [6]. Once a connection requestarrives between a pair of nodes, the centralized algorithm,a single entity, such as a network manager, runs the RWTAalgorithms to establish the lightpath for the connection request.If the lightpath cannot be established due to the continuityconstraints explained earlier, or due to the capacity exhaustionon the links, then the lightpath request rejection or callblocking occurs.

For the optical backbone networks, the customers are gener-ally, the internet service providers (ISPs) or large enterprises,who request huge bandwidth for longer duration of time. Theypay huge money for the lightpath. The lightpaths are long-lived with the average holding time of several months orseveral years. Each lightpath is setup on a provisioning basis.The longevity, combined with the cost of a high-bandwidthlightpath, means that a network operator is unlikely to rejecta lightpath request, as it is a significant loss of revenue. Inother words, no call must be blocked or the operator wouldlike to increase the time of first lighpath request rejection orcall blocking so that, more calls can be accommodated in

Slotid: 0 1 2 3

Fig. 2. TDM-WDM network (F=4)

the network before upgrading the network by the additionof more capacity on the existing links. Therefore, we areproposing an intelligent routing, wavelength and time-slotreassignment algorithm, where, when the call gets blocked,the already established calls in the network can be rerouted,wavelength and timeslot reassigned, so as to establish theblocked call. The wavelength and timeslot reassignment al-gorithm removes blocking due to the continuity constraintson the link. Simulation based analyses are used to study theresulting improvement in the time of first lightpath requestrejection, T for the proposed reassignment algorithm. Wehave considered, single-fiber TDM-WDM network withoutwavelength conversion. The simulation results show that, whenthe call blocking occurs in the network, by employing theproposed reassignment algorithms, maximum traffic can beaccommodated in the network before the first call gets blocked.

II. RWTA ALGORITHM

The network topology is represented as a graph G(V, E),where V denotes the set of vertices (network nodes) and E theset of edges (links). When the session request arrives betweena pair of nodes, RWTA algorithm establishes the lightpath.We assume a non-contiguous time-slot allocation on a singlewavelength along the same path. In this section, we presentthe details of the various routing schemes, the wavelengthassignment and the timeslot assignment algorithm considered.

A. Routing

1) Fixed Routing (FR): For routing, the shortest path rout-ing algorithm such as Dijkstra's algorithm [4]-[6] is usedto find the shortest path between the source node and thedestination node. This is the primary path and this approachis called the fixed shortest-path routing (FR).

2) Aternate Routing (AR): For alternate path routing (AR),Link-disjoint algorithm is used [4]. The alternate path is thenext shortest path between the source and the destinationwhich does not share any links with the primary shortest path.

3) Dynamic routing (DR): In dynamic routing, the routefrom the source to the destination node is chosen dynami-cally, depending on the network state. The network state isdetermined by the set of all connections that are currentlyin progress. Based on the link loading, weight functions areassociated with each link. We consider the Least ResistanceWeigth (LRW) function [6], to decide on the path among theprimary and the alternate. For a given link (i, j) C E, let CAiidenote the amount of available capacity on the link whenthe link state information is gathered; and let CTaX be themaximum link total capacity in the network.

The LRW weigth function is given by:cT

wiJ C ,(i,j) CeE (1)ij

where 1 denotes the measure of resistance a link offersto establish a connection. The greater the number of available

cTlinks, the lower is the resistance offered. normalizes theresistance based on the maximum link total capacity of thenetwork.

B. Wavelength AssignmentFor wavelength assignment, the First Fit algorithm [4]-[6]

is used. In this scheme, all the wavelengths are numbered.When searching for available wavelength, a lower-numberedwavelength is considered first before higher numbered wave-length. The first available wavelength is then selected. Thisscheme packs all of the in-use wavelengths towards the lowerend of the wavelength space and has a low computation cost.

C. Time-slot AssignmentOnce the path and the wavelength are determined, then the

required time-slots are allocated. For time-slot assignment,the First Fit algorithm [6] is used. This scheme is similar tothe first fit wavelength assignment algorithm. This algorithmsearches the time-slots from lower-numbered slots to higher-numbered slot, and the first available time-slot is selected. Thisscheme packs all the transmission slots towards the start of theframe and has a low computation cost.

III. PROPOSED REASSIGNMENT ALGORITHM

In this section, we present the details of the proposedreassignment algorithm for the single-fiber TDM-WDM basedoptical backbone networks supporting dynamic traffic. Wedefine the following before giving the details of the algorithm.Link Status Matrix L1,K,F: The link status is a three dimen-sional matrix, where 1 is the number of links in the network, Kis the number of wavelengths per link and F is the number oftime-slots per wavelength. The link status matrix at any pointof time gives the call occupancy in the network, in other words,for the call to be established in the network, what are the linksinvolved in the path, which wavelength is assigned in those

links and which are the time-slots allocated in that particularwavelength.Link Utilization (U): Link utilization for the network at anypoint of time is given by

>KE El Time Slots in useLink Utilization * K * F (2)

A. MOLC Reassignment algorithmWhen a new session request arrives to the TDM-WDM

network, RWTA algorithm establishes a lightpath for therequest. If the lightpath cannot be established, MOLC reas-signment algorithm is used to remove the call blocking. In thisreassignment algorithm, we try to shift the already establishedcalls from the minimum overlapping wavelength/timeslot (ex-plained in step 2 and 3 of the pseudocode) to the leastcongested wavelength/timeslot and establish the blocked callin the freed wavelength/timeslot. We have presented the pseu-docode description of the algorithm for the single wavelengthwith multiple slots. This can also be extended for multiplewavelengths per link.When a new call gets blocked,Step 1: Find the links requiredfor the new call by using the

routing schemes discussed in section II.Step 2: From the Link status matrix Ll,K,F at that point

of time, determine how many of the required linksare occupied by the already established calls in eachtimeslot. Store this information in the occupancy table.

Step 3: Sort the occupancy table in the ascending order, sothat the timeslot, which has minimum overlap with, therequired links of the new call is at the top of the table.

Step 4: Try to reassign the already existing calls thatare occupying the required links in the minimumoverlapping timeslot (obtained from step 3) to theleast congested timeslot. If the calls could not bereassigned, try to move the call to the next leastcongested.If all the required links in the minimum overlappingtimeslot are freed by reassignment, then establish thenew call in the freed timeslot.Else

If there is a next minimum overlapping timeslotavailable from the occupancy table, go to step 4.Else

Declare that the new call is blocked At this timeof blocking, find URestore the original link status.

Step 5: Go back to the normal RWTA for the next call requestarrival.

IV. PERFORMANCE ANALYSIS

This section presents the performance analysis of the pro-posed reassignment algorithm for the various routing schemesbased on simulation.

A. Network ModelThe network model considered for the simulation is the

14 nodes, 21 links NSFNET backbone network shown in

Avg U at T for K=1; F=1280.7

- MR(1 ... 128)x*- MR(1 8)

0.6 SR

11

Fig. 3. 14 Nodes, 21 Links NSFNET

c 0.540

2

.02

< 0.2

0.1 _

Avg TforK=1; F=12810 20 30 40

Traffic Load per node (Erlangs)

Fig. 5. Averaged U at T for MR and SR traffi c demands, K=1;F=1280)

mU-e

0

E./F

Traffic Load per node (Erlangs)

Fig. 4. Averaged T for MR and SR traffi c demands, K=1;F=128

the 3. The network is assumed to be a single-fiber networkwith no wavelength conversions at the network nodes. Thedynamic network traffic is generated in terms of connectionrequests from a source to destination node. The connectionrequests arrive at every node according to a Poisson process

with rate Ar arrivals per unit time, and uniformly randomdestination address. Each established connection holds for an

exponential distributed period with a mean (,u) of 30 timeunits.We consider two types of traffic demands: Multi-rate(MR) and Single-rate (SR). In multi-rate traffic demand,the bandwidth demand of the session is uniformly distributedbetween 1 to F; and in the single-rate traffic demand, sessionsare allowed to request one time-slot only. Traffic load per node

y, expressed in Erlangs, is given by

a = Ar * * (3)

where Ar is the mean arrival rate, ,u is the mean holding time,r1 is the mean number of requested slots. Thus for single-rate,rj is one.

The metrics used for quantifying the performance are: thetime of first lightpath request rejection T, link utilization U,and the probability of blocking, Pb.

Avg Pb for K=1; F=1280.45

MR(1 128)

0.4 -_ MR(1. 8)SR

0.35

0)

0.3

m0.25-

0.2-

5

0.1

0.05

0

0 10 20 30 40 50 60

Traffic Load per node (Erlangs)

Fig. 6. Pb for MR and SR traffi c demands, K=1;F=128

B. Performance of Multi-rate versus Single-rateThe basic RWTA algorithm which uses fixed routing is

applied to the multi-rate (MR) and single-rate traffic demands(SR) and the performance interms of T, U and Pb is studied.We assume, K=1 and F=128 for the simulation purpose. Wehave plotted from that load, where the number of time thefirst blocking occured over the large number of trials i.e.,the probability that a first blocking will occur is greater than30°0. Below this value of the load, the curve is non-monotonicbecause of very less number of first blocking occuring over

the large number of trials.Figure 4 presents the comparison of time for first call

blocking, T under the varying traffic load per node for thedifferent traffic bandwidth demands. Similarly, in Figure 5and Figure 6, the link utilisation, U and the probability ofblocking, Pb are plotted respectively under the varying trafficload per node. From the figures, it is seen that the performanceof the single-rate traffic demand (session request is always1 timeslot) is far better than the multi-rate traffic demand

600 r50 60

oL_0

(session request uniformly randomly distributed between 1and 128, mean number of requested slot r = 64.5). Thisis because, in MR, the session request for higher bandwidth(more than one timeslot, particularly when the requested valueis towards F i.e., 128), leading to heavy loading of the linksin the path of the session, whereas in the SR, the requests aresingle slot and the loading is somewhat uniformly distributedamong all the links in the network for the same traffic loadper node. Also, in MR, it is difficult to assign large number ofsame timeslot on all the links of the path (timeslot continuityconstraint). Because of the heavy loading of some links andtimeslot continuity constraint, the performance of multi-ratetraffic demand is poorer than the single-rate traffic demand.However, the performance of the multi-rate traffic demand canbe improved by allowing the sessions to request for timeslotsonly upto a value, which is smaller than F. For simulation wehave allowed the timeslot request to vary uniformly between1 to 8, mean number of requested slot r = 4.5. This willreduce the loading on the links and also the timeslot continuityconstraint, thereby improving the performance compared tothe one where the request was uniformly randomly distributedbetween 1 to F (128) Figures 4, 5 and 6. Hence for all thefurther simulations, we allow the requested slot to vary upto avalue which is smaller than F for the multirate traffic demand.

C. Techniques derived using the proposed reassignment algo-rithm

In this section, we present the various techniques derivedusing the above proposed reassignment algorithm for thesimulation purpose. The techniques are

1) FR: This is the basic RWTA algorithm which uses thefixed routing for the multirate traffic demand. This is the sameas the MR proposed earlier with the requested number oftimeslot varying upto a value less than F.

2) FR-MOLC-AR-MOLC (FM-AM): When the new callgets blocked in FR, MOLC reassignment algorithm is usedto remove the blocking due to continuity constraints in theshortest path between the source and the destination. This isFR-MOLC. When the call gets blocked in FR-MOLC, use thealternate path to accomodate the blocked call. This distributesthe load on the alternate path. If the call is still blocked, MOLCreassignment algorithm is used to remove the blocking in thealternate path, which is the FR-MOLC-AR-MOLC technique.Thus in this technique, blocking due to continuity constraintsis removed in both the primary and the alternate path.

3) FR-SR: This is the basic RWTA algorithm which usesthe fixed routing for the singlerate traffic demand.

4) DR: The dynamic routing is used to establish the light-path for the multirate traffic demand where the sessions areallowed to request for timeslots upto a value which is smallerthan F.

5) DR-MOLC: When the new call gets blocked in DR, usethe MOLC reassignment algorithm to accomodate the blockedcall in the network. Thus this removes any blocking due to thecontinuity constraints.

6) DR-SR: The dynamic routing is used to establish thelightpath for the singlerate traffic demand.

D. Performance Evaluation of above Techniques

Simulation was carried out for K=1, F=128 and the re-quested number of slots by the sessions uniformly randomlydistributed between 1 and 8 (mean number of requested slotr1 = 4.5); and K=1; F=32 and the requested number of slotsuniform between 1 and 4 (mean number of requested slotr1 = 2.5). We describe the results for only K=1; F=128 sincethe performance trends with K=1; F=32 are very similar.

Figure 7 shows T under varying traffic load per node for thevarious techniques - FR, FR-MOLC-AR-MOLC (FM-AM),FR-SR, DR, DR-MOLC, DR-SR. It is seen that FR-MOLC-AR-MOLC (FM-AM) performs much better than the FR i.e.,it provides around 273% improvement in T from FR at theload per node of 43 Erlangs. This is because, in FR-MOLC-AR-MOLC technique, as explained in section IV.C, the linkloads are distributed among the primary and the secondarypath at the time of blocking and also the blocking due to thecontinuity constraint is removed in both the primary and thealternate path. Hence there is a tremendous improvement inT from that of FR. The performance of FR-SR (single ratetraffic) is better than FR-MOLC-AR-MOLC, i.e., at a loadper node of 43 Erlangs, FR-SR has 85.6% improvement in Tfrom the latter. Thus, it is seen that the performance of theMR can be made closer to SR-FR (upper bound as discussedin section IV.B), by using the reassignment algorithm in theprimary and the alternate path.

In Figure 7 the dynamic routing, DR performs better thanthe FR i.e., around 153% improvement in T from FR at aload per node of 43 Erlangs. This is because the routing isdone based on the link loading. However the DR performs lessthan FR-MOLC-AR-MOLC. The performance DR-MOLC isbetter than DR, i.e., it provides 153% improvement in T fromDR at load per node of 43Erlangs. DR also performs betterthan FR-MOLC-AR-MOLC, around 70% improvement at aload per node of 43 Erlangs. This is because MOLC removesany blocking due to the timeslot continuity constraint in thedynamic routing. However, the performance of SR-DR is seento be better than DR-MOLC, because of the already discussedreasons, around 170% improvement in T from that of DR-MOLC at a load per node of 53 Erlangs. Figure 8 and 9 showsthe U and Pb under varying traffic load per node. Similarimprovements can be seen interms of U and Pb also.

The reassignment algorithm when applied for the singlerate traffic demand with fixed routing gives performanceimprovement which can be found in [7].

V. CONCLUSION

Lightpaths carry data at very high rates and are long-lived.The longevity combined with the cost of a high-bandwidthlightpath today, means that lightpath request rejection is aloss of revenue for the network operator. In this paper, wehave proposed a reassignment algorithm, which removes the

Avg TforK=1; F=128600 r

0)

cn

E

j=-

35 40 45 50

Traffic Load per node (Erlangs)

Fig. 7. Averaged T for various techniques, K=1;F=128

Avg U for K=1; F=128

* FRe FM-AM

0.65 FR-SRA DRE DR-MOLG

0.6 -_-0DR-SRo

C/U= 0.55

.C: 0.5

0.45

0.4

0.35 _25 30 35 40 45 50

Traffic Load per node (Erlangs)55

blocking due to the wavelength and timeslot continuity con-straint. This reassignment algorithm has been applied for thedifferent routing schemes and the performance improvementhas been studied for the multi-rate traffic demands. It is seenthat by using the reassignment algorithm in both the fixedand alternate routing, the performance can be improved almostclose to the single-rate traffic demand with the fixed routingwhich is considered to be the upper bound. The performanceof the dynamic routing with the reassignment algorithm issimilar to the single-rate traffic with fixed routing, but poorerthan dynamic routing for single-rate traffic demand. Thesimulation results show that by employing these techniques, it

Elk is possible to postpone the time for first call blocking, thereby accomodating more calls in the network before going for

60 network upgradation. Since the lightpaths have slow arrivalsand long holding times, it is possible to do the reassignment ofthe already established calls in the network while provisioninga new call.

REFERENCES

[1] Paul Green, "Progress in optical networking," IEEE CommunicationsMagazine, vol. 39, no. 1, pp. 54-61, Jan. 2001.

[2] R. Ramaswami and K. N. Sivarajan, "Optical Networks: A PracticalPerspective," Morgan Kaufmann Publishers, San Francisco, 1998.

[3] C. Siva Ram Murthy and Mohan Gurusamy, "WDM Optical Network:Concepts, Design, andAlgorithms," Prentice-Hall, Inc., New Jersey, 2002.

[4] H. Zang, J. P. Jue, and B. Mukherjee,"A review of routing and wavelengthassignment approaches for wavelength-routed optical WDM networks,"Optical Networks Magazine, vol. 1, no. 1, pp. 47-60, Jan. 2000.

[5] Nen-Fu Huang, Guan-Hsiung Liaw, and Chuan-Pwu Wang, "A novel all-optical transport network with time-shared wavelength channels," IEEEJournal on Selected Areas in Communications, vol. 18, no. 10, pp. 1863-1875, Oct. 2000.

[6] Bo Wen and Krishna M. Sivalingam, "Routing, wavelength and time-slotassignment in time division multiplexed wavelength-routed optical WDMnetworks," IEEE Infocom 2002, pp. 1442-1450, 2002.

[7] P Rajalakshmi and Ashok Jhunjhunwala, "Intelligent routing, wavelength60 and timeslot reassignment algorithms for TDM based Optical WDM

networks," OptoElectronics and Communications Conference 2005, pp.128-129, 2005.

Fig. 8. Averaged U at T for various techniques, K=1;F=128

Pb for K=1; F=128

* FRe FM-AM

0.1 FR-SRA DR- DR-MOLC

-O-- DR-SR= 0.08-00QCUQ0. 0.06

0)0 0.04

.n

Traffic Load per node (Erlangs)

Fig. 9. Averaged Pb for various techniques, K=1;F=128

0)

c-

>1