ARTICLE IN PRESS

OpticsOptikOptikOptik 121 (2010) 1412–1417

0030-4026/$ - se

doi:10.1016/j.ijl

�CorrespondE-mail addr

www.elsevier.de/ijleo

Comparison analysis of optical burst switched network architectures

Amit Kumar Garga,�, R.S. Kalerb

aSchool of Electronics and Communication Engineering, Shri Mata Vaishno Devi University (J&K), IndiabDepartment of Electronics and Communication Engineering, Thapar University (Punjab), India

Received 6 October 2008; accepted 5 February 2009

Abstract

Optical burst switching (OBS) is a promising paradigm for the next-generation Internet infrastructure. In this paper,a novel efficient network architecture for OBS has been presented and compared with conventional OBS architectures.To enhance OBS system performance, the architecture employs a novel proposed burst assembly algorithm, fiber delaylines (FDLs) and dynamic route selection technique. A queuing model is used to predict the system behavior for bothclassless and prioritized traffic. Simple closed-form expressions are obtained for the burst-loss probability of bothclassless and prioritized traffic. Numerical results show that the proposed architecture provides an accurate fit for theperformance of the highest traffic class and lower bounds for the other traffic classes that are tighter than earlier knownresults.r 2009 Elsevier GmbH. All rights reserved.

Keywords: Optical burst switching; OBS architectures; Burst assembly; Route selection

1. Introduction

The exponential growth of the Internet trafficdemands a high speed transmission technology tosupport rapidly increasing bandwidth requirements.Currently, the dense wavelength-division multiplexing(DWDM) technology achieves multiplexing of 160–320wavelengths in one fiber with 10–40Gb/s transmissionrate per wavelength. In order to efficiently utilize theraw bandwidth in DWDM networks, an all-opticaltransport system that can avoid optical buffering whilehandling bursts traffic, which can also support fastresource provisioning and asynchronous transmission ofvariable sized packets, must be developed. Optical burst

e front matter r 2009 Elsevier GmbH. All rights reserved.

eo.2009.02.010

ing author.

ess: garg_amit03@yahoo.co.in (A.K. Garg).

switching (OBS) [1] is a switching technique thatoccupies the middle of the spectrum between thewell-known circuit switching and packet switchingparadigms, borrowing ideas from both to deliver acompletely new functionality (as shown in Table 1).

OBS is a compromise between optical circuit switch-ing (OCS) and optical packet switching (OPS), since itallows for a data burst to be sent in an all-opticalmanner over the network, although the network switch-ing and input/output resources are reserved by asignaling message electronically interpreted at eachnode, sent prior to the burst in a separate channelnamed control channel. Network resources such aswavelength converters or data channels (l channels) arereserved at a node after the setup message is processedfollowing a given signaling protocol. These protocolsmay be classified as one-way reservation, termed

ARTICLE IN PRESS

Table 1. Comparison of optical switching schemes.

Optical switching (paradigm) Bandwidth utilization Latency (setup) Optical buffer Traffic adaptivity

Circuit Low High Not required Low

Packet/cell High Low Required High

Burst High Low Not required High

H

Ingress Edge Node

Core Switch

EgressEdge Node

Burst

Offset time H---header

Fig. 1. Optical burst switched (OBS) network architecture.

A.K. Garg, R.S. Kaler / Optik 121 (2010) 1412–1417 1413

tell-and-go (TAG–OBS), such as just-in-time (JIT) andjust-enough-time (JET), or two-way reservation, termedtell-and-wait (TAW–OBS) [1–3]. TAG protocols arefaster since they do not wait for a resource reservationconfirmation message, but have performance problemscaused by concurrent attempts to reserve the samenetwork resources. TAW protocols need a longer setuptime and the packets in the burst experience a longerdelay, but the probability of burst loss is smaller sincetransmission of the burst is done only after all theresources have been successfully reserved. In OBSnetworks, the traffic management decisions are per-formed at the edge nodes, keeping the core nodes assimple as possible. Thus, when an edge node transmits aburst into the network, its control packet (CP) alreadyincludes information on the path for the burst. Theinformation in OBS nodes is used only locally; thus thenetwork as a whole system does not benefit from theinformation available on each of the individual nodes.Some attempts have been made to solve this problem,such as the architecture using centralized managementmodel to optimize the utilization of the networkinformation.

2. OBS, issues and related work

Fig. 1 shows the basic procedure of sending one burstfrom an ingress node to an egress node in an OBSnetwork. At the ingress nodes of an OBS network, allTCP/IP packets are assembled into bursts. The ingressnode sends out a control (or setup) packet beforesending out the data burst. There is an offset timebetween the CP and the data burst to give the

intermediate OBS nodes enough time to configure theirswitching fabrics and reserve channel for the followingdata burst. The CPs are sent out on one or morededicated control channels (e.g. wavelengths) and gothrough O/E/O conversion at each intermediate node toprovide information about the coming burst. However,the data burst will go through each intermediate node inthe optical domain without any O/E/O conversion.There are many interesting issues in OBS, such as burstscheduling, burst assembly, offset time setting andcontention resolution [1,2]. Currently, how to efficientlyassemble IP packets to bursts in an OBS network is stillan open issue. Although fiber delay lines (FDLs) are notmandatory in OBS architecture, the system performancecan be significantly improved by employing them. Thus,it is of interest to evaluate the system dynamicsconsidering FDLs. Study of OBS with FDLs is achallenging task due to the unique behavior of FDLs.

Assume that the maximum delay that can be providedby an FDL is t seconds. Unlike conventional electronicbuffers, where a packet can stay in the buffer for anindefinite amount of time, the amount of time that anoptical burst can stay is constrained to be less than t.This is known as buffering with bounded delay. Inaddition, unlike electronic buffers, where a packet canuse a buffer as long as it is available, an optical burst canoccupy an FDL only if the FDL is idle and the requesteddelay is less than t. There are a few papers in theliterature dedicated to the study of OBS with FDLs.Turner [2] applied the M/M/k/D queuing model to studythe performance of OBS. Yoo et al. [4] used the M/M/k/D queuing model to find lower and upper bounds on theperformance and indicated unique behavior of FDLs.The basic OBS architecture is based on the premise thatdata are aggregated into bursts and transported from aningress point to an egress point in the network by settingup a short-life light path in the network in such a waythat the burst finds the path configured when it crossesthe network nodes. This light path is set in such a waythat it maximizes the utilization of the network’sresources. If the light path is to be explicitly destroyedthen either the ingress or the egress node will issue acontrol packet with a message that will remove theconfigured status for that data channel in each of thenodes; otherwise, in the implicit release scenario, eachnode will compute (in the case of estimated release) orassume (in the case of reservation for a fixed duration)

ARTICLE IN PRESSA.K. Garg, R.S. Kaler / Optik 121 (2010) 1412–14171414

the time after which the channel is again free andavailable. There are a number of variations on theoriginal OBS architecture, the most relevant of whichare: use of fiber delay lines as buffers, implementation ofburst segmentation, burst and path grooming, inte-grated burst assembly and disassembly with electronicbuffering and burst add drop at nodes, etc. Labeledoptical burst switching (LOBS) as proposed by Qiao andStaley [6] is viewed as a natural extension of the multi-protocol label switching (MPLS) framework for OBS[5]. In this architecture, the MPLS functionality servesas an integration layer between IP and the WDM. LOBSprovides path provisioning, traffic and resource engi-neering, network survivability and several other featuresrelated to the MPLS framework. In [6] LOBS nodearchitecture is proposed, in which incoming bursts maybe locally disassembled, and again assembled andreinserted into the network. MPLS messages are usedto control burst switching, bandwidth reservation andmainly serve to reduce the complexities associated withdefining and maintaining a separate optical (burstswitching) layer. Additionally, [6,7] also suggests thatfrom LOBS-based networks, migration and inter-networking with OPS will be easier. Dynamic wave-length-routed optical burst switched network architec-ture (DWR–OBS) was proposed in 2003 by Zapata andBayvel [8]. It is a compromise between the TAG–OBSand TAW–OBS, as it proposes a node whose function isto act as a reservation request broker to the network.DWR–OBS elects one node in the network to evaluatethe resource reservation requests from the edge nodes.This node, called central node, then issues backacknowledgment or rejection to the requesting nodes,thus managing all the network resources. Analysisperformed in [8] shows that this architecture can copeup to 115 nodes, thus making it suitable for medium sizenetworks. This limitation rises because of the computa-tional load posed on the central node, which mustprocess all the requests from all the nodes. Anotherlimitation of this network is the burst assembly time(AT), which must be long enough so as to allow therequest to travel from the ingress node to the centralnode and back. In [9] a pre-booking mechanism ispresented, derived from the DWR–OBS architecture.Authors claim that for a 90ms end-to-end delay with a10�4 bit loss tolerance, the pre-booking mechanismyields approximately twice as much traffic as theDWR–OBS. The ‘‘Lightnet Architecture’’ as proposedin [10] may also be viewed as a signaling protocol, sinceit can be interpreted as a no-reservation protocol (incomplement to TAG and TAW). This architectureimplements light paths using the availability of WDMdata channels, trying to maximize the wavelengthcontinuity constraint along the source–destinationpaths, in order to minimize switching and processingeffort inside the network.

3. Proposed efficient optical burst switched

network architecture

The proposed efficient OBS network architecturecomprises of the following:

�

Analytical OBS model

�

Dynamic burst assembly algorithm

�

Dynamic route calculation technique

3.1. Analytical OBS model

�

Consider an output queuing N�N optical burstswitch architecture, with k wavelengths in each port.Each input port has f FDLs for storing the bursts.The maximum delay offered by the FDL is t seconds. � Full range wavelength conversion capability. An

incoming burst can be directed from any wavelengthin the input port to any output port.

� A just-enough-time-type signaling protocol is used.

The latest available unscheduled channel (LAUC)scheduling algorithm is used for the classless trafficwhile the latest available void channel (LAVC)scheduling algorithm is used for the prioritizedtraffic.

� Optical bursts arrive according to a Poisson process

with a total intensity of l bursts per second on eachport. All burst lengths are exponentially distributedwith an average length of 1/m seconds. The burstdestinations are uniformly distributed. The systemutilization is r ¼ k/m.

� All bursts will contend for the wavelength in the

destined output port first. If a burst is blocked it willtry to reserve one of the available FDLs in the inputport. The burst is blocked if there is no FDLavailable or the FDL cannot provide enough delay.FIFO service discipline has been assumed for eachoutput port.

� The base offset time has been ignored, since it is

common for all bursts.

3.1.1. Classless and prioritized traffic models

Assuming output queuing switch architecture and auniform traffic distribution, it is appropriate to studythe system behavior at one typical output port. Since thefiber link contains k wavelengths, each output port is ak-server queuing system. Under the assumption ofPoisson burst arrivals and exponential burst duration,each output port can be modeled as a M/M/k queue. Inthe case of an electronic switch, a packet occupies thebuffer space until it is served. As a result, the bufferspace capacity is independent of the duration that apacket can stay in the buffer and a conventional k-serverqueuing model is sufficient. The maximum number of

ARTICLE IN PRESS

Fig. 2. Classless traffic model.

Fig. 3. Prioritized traffic model.

Fig. 4. A novel burst assembly algorithm [11].

A.K. Garg, R.S. Kaler / Optik 121 (2010) 1412–1417 1415

packets that can be held by the system equals thenumber of packet buffers. An incoming packet will bedropped whenever there is no buffer space to hold thispacket. In the case of an optical buffer, assume that themaximum delay provided by a FDL is t seconds andeach FDL can hold multiple bursts. With the assump-tion of an exponential distribution for the burstduration, the number of bursts contained in the FDLis actually being unbounded (as shown in Figs. 2 and 3).In this paper, the state of the system has been taken tobe the number of bursts in the system.

3.2. Dynamic burst assembly algorithm

It is assumed that the ingress node has one dedicatedburst assembly queue for each egress node. All incoming

packets will be forwarded to the corresponding queueaccording to their destinations. When the queue sizereaches a threshold or the waiting time of the packets inthe queue reaches a threshold, the packets in this queueare sent out as a burst. In this paper, a novel dynamicburst assembly algorithm has been proposed (it candynamically change the value of assembly time of anyqueue at every ingress node (e.g. ATqueue) according tothe length of burst recently sent).

So, in the proposed algorithm, a CP is generated whena burst exceeds a minimum burst length (MBL) or whenthe assembly times out, whichever comes first. These twoparameters (maximum assembly period (MAT) andMBL) can be set such that the MBL is smaller than theaverage burst length (obtained using Eq. (3)) and themaximum AT is approximately minf(RTO�RTT) (as inEq. (2), Fig. 4).

3.3. Dynamic route calculation technique

The dynamic route calculation can be based on manydifferent metrics such as the physical distance, numberof hops, congestion information and link utilization.The routes are recomputed every t units of time. Theweight w(i, j) is based on a single metric or acombination of metrics. One option is to set the weightfunction equal to the congestion metric, resulting in theleast congested path. The issue with the above metric is

ARTICLE IN PRESS

Traffic Load Per Wavelength0.45

Bur

st L

oss

Pro

babi

lity

1e-7

1e-6

1e-5

1e-4

1e-3

1e-2Conventional OBS Architectures with k =120Proposed OBS Architecture with k =120

0.50 0.55 0.60 0.65 0.70 0.75 0.80

Fig. 5. Burst-loss probability performance of OBS architec-

tures.

Load (Erlang)2

Thro

ughp

ut (B

urst

s)

0

20

40

60

80

100

4 6 8 10 12 14

Fig. 6. Throughput (bursts) vs load.

A.K. Garg, R.S. Kaler / Optik 121 (2010) 1412–14171416

that some of the resultant routes will have a highnumber of hops. Therefore, while sending the burst onthe least congested route results in low packet lossprobability at lower loads, under higher loads, longerpaths will result in higher overall network loads, therebyincreasing the probability of contention. In order toavoid this situation, a weighted function has beenconsidered based on congestion as well as hop distance:

wði; jÞ ¼ rði; jÞ þ 1 (6)

where r(i, j) is the offered load on the link (i, j)Another option is to define the weight function based

on congestion as well as physical distance:

wði; jÞ ¼ rði; jÞ þ dði; jÞ=dmax (7)

where d(i, j) is the physical distance of the link (i, j) anddmax is the maximum physical distance of any link in thenetwork. In general, the hop-based metric (Eq. (6))results in better performance in terms of loss, sinceminimal number of nodes are selected in a path, therebyreducing the probability of contention. On the otherhand, the distance-based metric (Eq. (7)) results in betterperformance in terms of delay, since minimal linkdistances are selected in a path, thereby reducing thepropagation delay. Thus, in this paper Eq. (7) has beenconsidered to have reduced propagation delay in theproposed OBS architecture.

4. Performance analysis

In this simulation, a National Science FoundationNetwork (NSFNET) topology [12] with 12 nodes hasbeen considered. It is also assumed that each single fiberlink is bidirectional and has the same number ofwavelengths, each operating at 2.5Gbps. Each node inthe network can route, generate and receive traffic. Thefollowing performance metric, the burst (packet) lossprobability, which is the main metric in OBS networksand throughput (bursts) has been considered.

4.1. Simulation parameters

Average burst length ¼ 90 ms; control burst proces-sing time ¼ 2.5 ms; switching time ¼ 12 ms; propagationdelay on a link ¼ 0.2ms; the data bursts are notretransmitted; bit errors in transmission are ignored;the size of the electrical buffers in the edge nodes isinfinite; the simulation experiments were run for asufficiently long time and were repeated several times;b ¼ 12 as the assembly factor (based primarily on thetopology and the shortest path routing for every nodepair); assembly time (AT) ¼ 0.02 s; the 95% confidenceinterval range is within 3% of the values plotted.

4.2. Numerical results

Fig. 5 compares the performance of conventional andproposed OBS architecture with k ¼ 120 wavelengths.For such a large system, it is seen that loss rates of 10�6

has been achieved at quite high utilizations, around 0.47for conventional OBS and 0.53 for the proposed OBSarchitecture. In addition to this, Fig. 6 shows that withan appropriate assembly algorithm and route selectiontechnique, the proposed (simulated) OBS networkarchitecture provides an accurate fit with analyticaltraffic models for the performance of the highest trafficclass and lower bounds for the other traffic classes thatare tighter than earlier known results.

5. Conclusion

Based on the properties of optical burst flows andbandwidths provisioning in the networks, an efficientOBS network architecture employing a novel burstassembly algorithm and fiber delay lines along withdynamic route selection technique has been proposed toimprove system performance. The simulation resultshave indicated that the proposed OBS architecture canreduce the simultaneous contention in the core networkand make the traffic smoother and hence improve theperformance in terms of burst-loss rate and throughput.Simple closed-form expressions are obtained for the

ARTICLE IN PRESSA.K. Garg, R.S. Kaler / Optik 121 (2010) 1412–1417 1417

burst-loss probability of both classless and prioritizedtraffic. Simulation results show that the proposedarchitecture provides an accurate fit for the performanceof the highest traffic class and lower bounds for theother traffic classes that are tighter than earlier knownresults.

References

[1] C. Qiao, M. Yoo, Optical burst switching (OBS) – a new

paradigm for an optical internet, J. High Speed Networks

8 (1999) 69–84.

[2] J. Turner, Terabit burst switching, J. High Speed

Networks 9 (1999) 69–84.

[3] Y. Xiong, M. Vandenhoute, H. Cankaya, Control

architecture in optical burst-switched WDM networks,

IEEE J. Sel. Areas Commun. 18 (2000) 1838–1851.

[4] M. Yoo, C. Qiao, S. Dixit, QoS performance of optical

burst switching in IP over WDM networks, IEEE J. Sel.

Areas Commun. 18 (2000) 2062–2071.

[5] C. Qiao, Labeled optical burst switching for IP-over-

WDM integration, IEEE Commun. Mag. 38 (9) (2000)

104–114.

[6] C. Qiao, J. Staley, Labeled analog burst switching for any

high layer-over-signal channel integration, patent filed in

USA, US 2003/0067919 A1, filed in 10 April 2003.

[7] E. Mannie (Ed.), Generalized Multi-Protocol Label

Switching (GMPLS) Architecture, RFC 3945, Internet

Engineering Task Force (IETF), 2004.

[8] A. Zapata, P. Bayvel, Dynamic wavelength-routed optical

burst switched networks: scalability analysis and compar-

ison with static wavelength-routed optical networks, in:

Proceedings of the Optical Fiber Communication Con-

ference and Exposition, OFC 2003, vol. 1, Atlanta, GA,

USA, 2003, pp. 212–213.

[9] H. Kong, C. Philips, Pre-booking reservation mechanism

for next-generation optical networks, IEEE J. Sel. Top.

Quantum Electron. 12 (4) (2006) 645–652.

[10] I. Chlamtac, A. Ganz, G. Karmi, Lightpath communica-

tions: an approach to high bandwidth optical WANs,

IEEE Trans. Commun. 40 (7) (1992) 1171–1182.

[11] P. Fan, C. Feng, Y. Wang, N. Ge, Investigation of the

time-offset-based QoS support with optical burst switch-

ing in WDM networks, in: Proceedings of ICC’02, 2002,

pp. 2682–2686.

[12] Network Simulator, NS-2, available at /http://www.

isi.edu/nsnam/nsS.