multiservice qos-enabled mac for optical burst switching · multiservice qos-enabled mac for...

530 J. OPT. COMMUN. NETW./VOL. 2, NO. 8 /AUGUST 2010 Triay et al.

Multiservice QoS-Enabled MAC forOptical Burst Switching

Joan Triay, Georgios S. Zervas, Cristina Cervelló-Pastor, and Dimitra Simeonidou

tHwe(trdatpt

tOaccfewrmtrr

cittwuloo[Q(alwpMO

Abstract—The emergence of a broad range ofnetwork-driven applications (e.g., multimedia, onlinegaming) brings in the need for a network environ-ment able to provide multiservice capabilities withdiverse quality-of-service (QoS) guarantees. In thispaper, a medium access control protocol is proposedto support multiple services and QoS levels in opticalburst-switched mesh networks without wavelengthconversion. The protocol provides two different ac-cess mechanisms, queue-arbitrated and prearbitratedfor connectionless and connection-oriented bursttransport, respectively. It has been evaluated throughextensive simulations and its simplistic form makes itvery promising for implementation and deployment.Results indicate that the protocol can clearly providea relative quality differentiation for connectionlesstraffic and guarantee null (or negligible, and thus ac-ceptable) burst loss probability for a wide range ofnetwork (or offered) load while ensuring low accessdelay for the higher-priority traffic. Furthermore, inthe multiservice scenario mixing connectionless andconnection-oriented burst transmissions, three dif-ferent prearbitrated slot scheduling algorithms areevaluated, each one providing a different perfor-mance in terms of connection blocking probability.The overall results demonstrate the suitability of thisarchitecture for future integrated multiservice opti-cal networks.

Index Terms—Optical fiber communication; Me-dium access control; Optical burst switching.

I. INTRODUCTION

T riple-play services (i.e., data, voice, and video)and the new deployment of web-based multime-

dia applications have increased the amount of burstytraffic on the Internet. Such services may benefit from

Manuscript received December 17, 2009; revised April 29, 2010;accepted May 28, 2010; published July 14, 2010 �Doc. ID 119755�.

Joan Triay (e-mail: [email protected]) and CristinaCervelló-Pastor are with the Department of Telematics Engineeringat Universitat Politècnica de Catalunya (UPC), Esteve Terradas 7,08860, Castelldefels, Spain.

Georgios S. Zervas and Dimitra Simeonidou are with the School ofComputer Science and Electronic Engineering at University ofEssex, Wivenhoe Park, CO4 3SQ, Colchester, United Kingdom.

Digital Object Identifier 10.1364/JOCN.2.000530

1943-0620/10/080530-15/$15.00 ©

he utilization of packet/burst-switched networks.owever, certain applications such as IPTV, togetherith grid and cloud computing (e.g., PC virtualization,tc.) can benefit from time division multiplexingTDM) connection-oriented transport networks. Theraffic diversity created by such applications and theapid advance of optical technologies has driven theevelopment of new optical network architecturesble to provide flexible and dynamic resource alloca-ion [1]. In this sense, a desirable requirement is torovide and guarantee a great variety of services overhe same optical network infrastructure.

A promising technological option for the future op-ical Internet is optical burst switching (OBS) [2].BS can satisfy the future bandwidth requirementsvoiding the inefficient resource utilization of opticalircuit switching (OCS) and the requirements of opti-al packet switching (OPS) in terms of optical buffers,ast processing, and implementation complexity. Nev-rtheless, optical burst-switched networks withoutavelength conversion may exhibit a high burst loss

ate in the core of the network, especially if no othereans of contention resolution are present. In view of

his, we are interested in pursuing new contentionesolution (or avoiding) methods to keep losses to aeasonably low level.

Medium access control (MAC) protocols can effi-iently manage the huge optical bandwidth by provid-ng contention avoidance schemes to further improvehe packet/burst delivery on the network. This charac-eristic becomes of special interest for all-optical net-orks with limited contention resolution. However, itsse in OBS networks has not been intensively ana-

yzed, and most current MAC solutions have focusedn metro-ring architectures, leaving out of their scopether common topologies, i.e., mesh. In this sense, in3] an adaptation of the IEEE 802.6 Distributedueue Dual Bus (DQDB) [4] for OBS networks

DAOBS) was introduced. DQDB defines a queue-rbitrated access to the channel guaranteeing zeroosses for transmitted frames. The original protocolas enhanced by adapting it over mesh network to-ologies and integrating it as a wavelength-awareAC for wavelength continuity constraint slottedBS networks. In DAOBS, a set of counters keep

2010 Optical Society of America

ttpctrl

wsutptpabr

pspbmr

spiptfihi

ictcq[tPtsNsP

hi[crb

Triay et al. VOL. 2, NO. 8 /AUGUST 2010/J. OPT. COMMUN. NETW. 531

track of the requests for free slots and their availabil-ity using request and burst control packets. With thisinformation, bursts are only transmitted when thereare free resources; hence overlapped burst losses areavoided.

QoS differentiation in OBS networks is also an im-portant issue. Certain types of QoS techniques appliedin traditional store-and-forward electronic networks,such as active queue management, are no longer thebest way to provide service differentiation in OBS un-less we accept the loss of the optical data transpar-ency. Thus, other types of techniques need to be ana-lyzed.

In this paper we give insight into the performanceof a lossless (in the core of the network) enhancedmultiservice OBS MAC protocol with QoS. A contribu-tion of this paper with respect to past approaches isthe use of the MAC on more complex topologies suchas mesh. The protocol supports a great number of ser-vices by means of a dual access scheme: a queue-arbitrated (QA) and a prearbitrated (PA) burst trans-mission method, for connectionless and connection-oriented TDM OBS services, respectively. Differenti-ation among connectionless classes is provided by amultiqueue priority system along with a distributedqueue-arbitrated channel access module. This systempermits higher-priority bursts from a node to preemptlower-priority bursts from other nodes on the net-work and be placed and transmitted ahead in time.Connection-oriented services use the prearbitratedchannel access, which can incorporate different slot-scheduling algorithms. The proposed algorithms de-liver diverse results depending on the network stateand traffic distributions; hence a dynamic decision-making protocol may enhance the scheduler accordingto the type of service request, the present network uti-lization, and the current traffic distribution on thenetwork. The protocol is named DAOBS, which standsfor distributed access for optical burst switching.

The remainder of the paper is as follows. Section IIreviews existent QoS schemes, recent multiservice ar-chitectures, and MAC protocols for OBS networks. InSection III, the proposed MAC protocol is introducedand its QoS enhancements and access modes are de-scribed in detail. Results through simulations areanalyzed in Section IV, and finally Section V concludesthis paper with the main contributions and results.

II. BACKGROUND ON QOS, MULTISERVICE

ARCHITECTURES, AND MAC FOR OBS

In OBS, QoS differentiation can be achieved by us-ing many different strategies in order to guaranteedifferent quality parameters, although most of themare based on a per-class approach [5]. Additionally, in

he per-class method, QoS parameters can be differen-iated as absolute (a fixed quality in terms of a certainarameter) or relative guarantees (the quality of eachlass is qualitatively or proportionally guaranteed be-ween classes). Offset-based, preemption-based, andestriction-based schemes are three of the main burstoss differentiation QoS strategies for OBS.

Offset-based schemes [6] rely on the fact that burstsith a greater offset time should have more time to

earch for free resources on the network and be sched-led for switching with higher probabilities thanhose bursts with a smaller offset time. Thus, high-riority bursts are given a greater offset. Althoughhis scheme increases the delivery rate of high-riority bursts, it also increases the latency of them,nd for this reason, it is not a feasible scheme whenoth delay and loss rate need to be guaranteed at aeasonable level.

In the preemption-based QoS scheme [7], high-riority bursts are able to take over (preempt) the re-ources taken by low-priority bursts, whereas low-riority bursts can never preempt high-priorityursts. Hence, on average, high-priority bursts seeore available resources, which results in a lower loss

ate.

Resource restriction-based schemes exclusively re-erve a subset of the available resources for high-riority traffic only. An example is wavelength group-ng [8], which prereserves wavelengths for high-riority bursts that can only be used by them even ifhe wavelengths are available for lower-priority traf-c. Intuitively, the more wavelengths are reserved forigh-priority traffic, the smaller its blocking probabil-

ty will be.

Regarding the provisioning of multiple services us-ng a common optical networking infrastructure, re-ent works have proposed the use of polymorphous op-ical burst switching (POBS) as an efficient way toope with the high-speed and optical transparency re-uirements of future optical networks. For instance,9] proposes a common and integrated signaling sys-em to provide an extensive number of services usingOBS, whereas [10] gives an exhaustive analysis of

he best-effort burst-based service performance whenome channel capacity is allocated to TDM services.evertheless, neither of them provides a detailed and

pecific architectural protocol analysis to supportOBS.

In terms of MAC protocols for OBS networks, theseave been given little attention, and almost all exist-

ng studies are focused on OBS metro-ring networks11]. For instance, [12] proposes a simple MAC proto-ol called beforehand bandwidth reservation for OBSing networks to reserve the empty slots in the nextig-slot cycle. Likewise, in [13] a loss-free OBS metro

gccft

iipcbstvccsadvummme

h(sbdHvanBltt

Fne


ring architecture (CORNet) is designed along with adistributed MAC protocol to integrate the support ofdifferentiated service and fairness access. The pro-posed QoS provisioning mechanism adopts abandwidth-reservation approach that combines real-time transmission establishment and terminationroutines. Furthermore, a credit-based fairness controlscheme is defined to guarantee the transmission fair-ness of best-effort traffic. Recently, [14] has proposed amultiple-token-based MAC protocol for OBS ring net-works using a tunable transmitter and one tunable re-ceiver (TT-TR). In order to avoid or reduce the receivercontentions, tokens manipulate the wavelength acces-sibility and the destination queues decide on the burstscheduling.

The lack of MAC protocols on mesh network topolo-gies and the need to provide integrated optical net-work architectures able to address the QoS and mul-tiservice demands for the future optical Internetmotivate the present work. Next, a detailed descrip-tion of the MAC protocol and its operation follows.

III. MULTISERVICE QOS-ENABLED MAC FOR OBS

In this paper we conceive an optical burst-switchednetwork without wavelength conversion. Further-more, the network data channel is time sliced, e.g.,slotted OBS [15], and a constant-based offset schemeis used as in [16]. This last approach allows us to usefixed offsets by means of input delay lines at each in-put burst data port of a length equivalent to the maxi-mum delay incurred in the processing of the burstcontrol packet (BCP).

One of the main differences of DAOBS comparedwith other MAC-based schemes for OBS networks isthat, in the former case, the protocol runs on meshnetwork topologies. To realize this, the network islogically partitioned in monocolor light-trees. Concep-tually, the DAOBS optical tree has some similaritiesto a light-trail [17]. Both propose network architec-tures that can span over ring and mesh topologies us-ing optical buses/lightpaths enabling the participatingnodes to share the capacity of the channel in a spatialreuse time-shared basis. Moreover, both also definethe figure of a controller, the head of bus (HoB) inDAOBS, and the arbitrator in the light-trail. However,DAOBS is also a burst MAC protocol with collisionavoidance in the core of the network that ensures thedelivery of the burst without using electronic buffers,as in [17]. Furthermore, the request-granting processoperates differently. In the DAOBS QA access modethis process works as a very simple algorithm basedon a counting and monitoring process, whereas in thelight-trail solution an explicit request grant is estab-lished between the client node (the node requestingslots) and the arbitrator, which decides what node is

ranted in a per-slot basis. Finally, DAOBS also in-ludes a wavelength assignment module since a nodean belong to more than a single light-tree, and there-ore multiple path/wavelength options are available tohe nodes while transmitting data.

Despite that the analysis of the different methods tonstantiate light-trees is out of the scope of this paper,t is worth mentioning that for the evaluation of theresent protocol a greedy metaheuristic graph-oloring algorithm has been used. This algorithm isased on the coloring of the graph using minimum-panning trees and the remaining uncolored links, al-ogether conforming a superset of the set of light-pathirtual topologies [18] able to improve the utilizationapacity of the network. One of the advantages of thisoloring scheme is that through this tree-based in-tantiation, collisions from two different input portsre avoided, which is especially useful in the case ofealing with OBS networks without wavelength con-ersion. Furthermore, light-tree-based topologies areseful for delivering multicast or groupcast services orerely for multimedia broadcasting. Besides, opticalusticasting can be more efficient than electroniculticasting since “splitting light” is conceptually

asier than copying a packet in an electronic buffer.

The root or topmost node of the light-tree is theead-of-bus node, and all leaf nodes are tail-of-bus

ToB) nodes (see Fig. 1). Under this scenario, we con-ider two unidirectional control channels, which cane in-fiber (i.e., using a specific wavelength): theownstream or forward channel, which goes from theoB node to the ToB nodes, and the upstream or re-

erse channel, on which QA access request packetsre forwarded from the ToBs to the HoB. The HoBode is responsible for generating and forwarding theCPs in the DAOBS light-tree at every time slot. The

ight-tree is normally composed of other nodes be-ween the HoB and ToB. According to the operation ofhe different MAC access modes, all the nodes can

To B

BC P

R C P

T o B

BCP

BCP

BCP

RCP

RCP

RCP

Forward control channel

Reverse control channel

ig. 1. (Color online) Example of a DAOBS light-tree. If node N1eeds to transmit to any ToB, it requests for free slots to the HoB orventually waits for a free slot.

uoPesct

mMaaslam

A

csfsfdspsstm


transmit bursts to the rest of the downstream nodesmaking use of the multiplexing capabilities within thelight-tree.

As has previously been introduced, every node canbelong to manifold light-trees. As a result, a node mayhave to manage multiple DAOBS entities, one foreach accessible light-tree. In the proposed architec-ture, a DAOBS entity is identified by its input port+output port+wavelength, and a DAOBS light-tree byits HoB+wavelength on the network.

DAOBS provides two channel access mechanisms:queue arbitrated and prearbitrated. Each accessscheme uses a different slot type. The former is de-vised for connectionless burst transport services, thatis, burst transmissions that have not been explicitlyacknowledged about the availability of a certain chan-nel capacity. On the contrary, PA is used by services orapplications that need a guaranteed reserved band-width. For a more detailed application usage examplewe refer to Table III in the results section. Figure 2shows an example where bursts of different servicetypes are transported on three wavelength channels.As seen in the figure, connectionless bursts are trans-mitted in QA slots, sharing the available channel ca-pacity together with connection-oriented applicationsmaking use of PA slots. The number of PA reservedslots depends on the application bandwidth require-ments. For instance, on channel �3 two consecutiveslots (every six) are used by a single connection.

In order to support such an architecture of services,the burst control packet has a 1 bit field (PA/QA) toannounce the upcoming slot type as shown in Fig. 3. Ifthis field is equal to 1, the network node currently pro-cessing the BCP runs the PA access mode. Otherwise,the node executes the QA access mode. BCPs are cre-ated at the HoB nodes every time slot, transmittedover the control channel and processed by the corenodes in advance to the reception of the upcoming slot(e.g., eventually taken by a burst). The PA/QA type isassigned according to the flow diagram from Fig. 4.The HoB always (i.e., at every slot) creates QA slots

Fig. 2. (Color online) Data channel PA/QA slot example.

nless a PA slot has been reserved by the service layern the slot supercycle window. In such a case, a newA slot is created by setting the PA/QA field to 1. Oth-rwise, a QA slot is forwarded to the next node down-tream on the light-tree. In both cases, the HoBhecks whether it is indeed able to fill the next slot byransmitting one of its enqueued bursts.

Figure 5 shows a high-level diagram of theultiservice-aware DAOBS architecture. The OBSAC is composed of four main submodules: the QA

nd PA burst access modules; the slot scheduling (SS)lgorithm module (only in the HoB); and the burst as-embly, classification, and scheduling module. Thisast module includes the burst assembly algorithmsnd classifies the burst between the QA and the PAodules according to its type.

. Queue-Arbitrated Access

In the QA access, the two unidirectional controlhannels transport the BCPs and RCPs: on the down-tream or forward channel, the BCPs are forwardedrom the HoB node to the ToB nodes, and on the up-tream or reverse channel, the RCPs are forwardedrom the ToBs to the HoB. As has already been intro-uced, both types of control packet are sent every timelot. This is necessary in order to allow all the nodesarticipating in the light-tree to check whether freelots are coming and to permit them to request freelots periodically. Since the bursts are much longerhan the packets, the electronic processing require-ents can easily be met.

Fig. 3. BCP and RCP packet formats.

Fig. 4. Flow diagram at the HoB slot processing.

oc

wtqolqo(lapiTgDpcsb

iDmfteatnDnue

Rf


All the nodes in the light-tree can transmit bursts todownstream nodes according to the operation of theQA access, that is, a node requests free slots to the up-stream and HoB nodes using the REQ bits of the RCPto subsequently transmit bursts by taking the upcom-ing free slots on the downstream direction. A set ofcounters for each priority keeps track of the requestsand free slots coming from downstream and upstreamnodes, respectively. Connectionless burst losses onlyhappen at the edge of the network due to queue block-ing when more load is offered than the acceptable.

BCPs have a BUSY bit for announcing whether theupcoming slot is taken up, whereas RCPs have a REQbit for each burst priority class. Figure 3 shows an ex-ample of the packet format for both the BCP and RCPwith three priority classes. Recall that a QA accessslot is announced by the PA/QA field set to 0. In thecase that the BUSY bit is equal to 1, then the rest ofthe fields in the BCP (BurstID, Source, Dest., etc.) aremeaningful since the upcoming slot is taken.

The QA access module of the DAOBS MAC entity ina node is composed of the following components foreach priority class i: a distributed access state ma-chine (DASM), a request control machine (RQM), a lo-cal queue (LQ), and a distributed queue (DQ). Figure6 depicts the relation between all these elements for aDAOBS entity on wavelength �k with three burst pri-

A p p l i c a t i o n

S e r v i c e L a y e r S i g n a l i n g

O B S M A C P r o t o c o l

B C P B C P

R C P

F o r w a r d c o n t r o l c h a n n e lR e v e r s e c o n t r o l c h a n n e l

R C P

H o B T o B







Q AB u r s t A c c e s s

P AB u r s tA c c e s s

C l a s s i f i c a t i o n , b u r s t a s s e m b l y a n d s c h e d u l i n g

N o d e






S l o t S c h e d u l e r


Fig. 5. (Color online) DAOBS protocol and network architecture.

F o r w a r d C o n t r o l C h a n n e l

R e v e r s e C o n t r o l C h a n n e l

Q A M o d u l e

B C P

R C P

L Q λ k D Q λ k

R Q 0 C D 0

D A S M 0

R Q 1 C D 1

D A S M 1

R Q 2 C D 2

D A S M 2

B C P

p r o c e s s o r

R C P

p r o c e s s o r

p 1

p 0

p 2

p 0

p 1

p 2

E n t i t y λ k

E n t i t y λ l

E n t i t y λ w

A g g r .

q u e u e s

Wavelength/Entity assignment

Input packets

R Q M 0

R Q M 1

R Q M 2

O p t i c a l

c h a n n e l

s e n t e v e r y s l o t

s e n t e v e r y

s l o t

Fig. 6. (Color online) QA entity with three burst priorities onwavelength � .
k
rity classes, and the interconnection between theontrol, state machines, and packet processors.

The LQ temporarily stores the bursts from theavelength/entity assignment module while waiting

o gain access to the channel. The DQ is a one-positionueue that stores the next burst to transmit. The setf DQs from the nodes that belong to the DAOBSight-tree forms the so-called virtual distributedueue. Thus, when a burst at a certain node gets intone of its DQs, it is like accessing a first in, first outFIFO) queue distributed among the nodes from theight-tree. Furthermore, at each priority there is also

RQM that monitors the requests triggered for thatriority. Finally, the DASM is responsible for monitor-ng and managing the counting process of the protocol.here is a DASM for each priority, and in the exampleiven in Fig. 6 these are DASM0, DASM1, andASM2, where the greater the number, the higher itsriority. As shown in the figure, high-priority DASMan signal and preempt lower-priority DASMs to en-ure that higher-class connectionless bursts receive aetter service level.

Each DASM can be in two different states as shownn Fig. 7. In the idle state, the OBS node, for a certainAOBS light-tree and priority, has nothing to trans-it, whereas in the active state, the node has success-

ully made a request for a free slot and is waiting toransmit the burst using that slot. Apart from this,ach DASM has two counters: a request counter (RQ)nd a countdown counter (CD). While the RQ moni-ors the number of requests made by downstreamodes on the optical light-tree and by higher-priorityASMs in its own DAOBS entity, the CD counts theumber of free slots the current node is not allowed tose before being given access to transmit a burst. Allntity counters are initialized to 0.

While a DASM at priority i is idle, it monitors theCPs on the reverse control channel and BCPs on the

orward control channel and increases [see Fig. 7 step

Fig. 7. QA DASM flow at priority i.

wf

lmFtc

B

rasctltcst

snut

Fa


(a1)] or decreases (a3) the RQi for every REQj=1 inthe RCP of a priority j� i and for every BUSY=0 inthe BCP, respectively. Similarly, if the DASMi receivesa SELF_REQj signal from a higher-priority DASMj�j� i� (a2), then the RQi is also increased.

As soon as a burst is stored in the LQ of a certainDAOBS entity and priority i, if the DASMi is idle itswitches to the active state following the transition(a4) shown in Fig. 7. The same happens if after re-turning from a successful burst transmission thereare more bursts to transmit in the LQ. This statetransition (a4) triggers the following events: first, aSELF_REQi signal is sent to the remaining DASMj�j� i� in the entity; then the value of the RQi isdumped to the CDi �RQi←CDi�, and the RQi is thenreset to 0; and finally a REQi signal is sent to theRQM of that same priority in order to set the REQ bitof that priority to 1 in the next upcoming free RCP re-ceived from the downstream node.

In the active state, the DASMi continues monitoringthe BCPs and RCPs. For any REQ bit of priority j� i(b2) or SELF_REQj signal from a DASMj with j� i(b4), CDi is increased by 1. These two steps let higher-priority bursts on the DAOBS light-tree be placedahead of and thus transmitted before, even betweenbursts from two different nodes of the same tree. In(b3), for every REQ bit of priority j= i, the RQi is in-creased by 1. Likewise, for every empty slot on the for-ward control channel (b5), and provided that the CDi�0, the CDi is decreased by 1 (down to 0). Finally, thetransmission of the burst from the DQ at priority ihappens when CDi reaches 0 and an empty slot comesfrom the upstream link. This last step involves thetransition (b1) of DASMi from active to idle.

The wavelength assignment is responsible for as-signing burst transmissions to a specific DAOBSlight-tree. As stated before, each DAOBS tree is iden-tified by its HoB and wavelength. Therefore, thewavelength selection is in fact a DAOBS entity assign-ment (see Fig. 6). In the QA access, the wavelength as-signment is controlled by an algorithm that takes intoaccount the values of the counters RQ and CD and thenumber of bursts ahead in the LQ. As we have previ-ously introduced, the value of these counters deter-mines the position of the node in the distributed FIFOqueue at a certain priority class. The number of burstsin the LQ gives information about the number oftransmissions not processed yet. Depending on thesevalues, the node will take longer to transmit a burston that specific light-tree, thus increasing or decreas-ing the channel access delay, and consequently, theend-to-end delay of the burst. In this process, theDAOBS light-tree that is expected to provide the low-est access delay is always selected. That is, for priorityi,

min�j�Wnm�RQi,�j

+ CDi,�j+ size�LQi,�j

��, �1�

here Wnm is the wavelength/DAOBS candidate list

or transmitting a burst from node n to m.

One of the advantages of the proposed protocol is itsow complexity. In the QoS-enabled DAOBS QA access

ode, the hardware complexity is easily bearable.urthermore, the scheduling complexity is again vir-

ually null as the QA mode just works as a set ofounters that can easily be monitored and updated.

. Prearbitrated Access

In the PA access, downstream (core) nodes explicitlyequest the connection setup to the light-tree HoB forguaranteed channel bandwidth (periodic transmis-

ion of bursts). In turn, the HoB acknowledges theore node about the success of the connection. Al-hough this feature is supported by a higher serviceayer that monitors applications’ service requests, it ishe MAC layer that implements the scheduling, allo-ation, and differentiation of slots. Service layer mes-ages are transmitted and forwarded through the con-rol channel together with the BCPs and RCPs.

Figure 8 shows an example of an explicit multilayeretup and release of a TDM connection-oriented chan-el and the slot allocation at the MAC level. Let aser/application request a connection channel be-ween node N1 and a downstream node on the light-

H o B N 1

0

1

1

1

0

0

0

1

0

0

0

1

0

1

1

1

0

0

0

1

0

0

0

0

0

s e t u p

a c k

r e l e a s e

a c k

N 2 T o B

0

0

0 1 1

0

0

0

Q A s l o tP A s l o t

n o t b u s y

P A s l o t

b u s y

R

R

R

R

R

R

R

R

R

R

R

R

R

R

R R C P

R

S e r v i c e L a y e r ( N o d e N 1 )

U s e r / a p p l i c a t i o n L a y e r ( N o d e N 1 )

( 1 )

( 2 )

MAC Layer

( 3 )

( 4 )

( 5 )

( 7 )

( 8 )

( 6 )

( 9 )

( 1 0 )

ig. 8. (Color online) Example of a multilayer PA service setup andllocation of slots in the MAC layer.

tmdstbc1td

FTfiTaosbwwrjs

pb�dmcsfd


tree (1) (e.g., the ToB). The service layer computes theconnection requirements of the application call (2).For instance, let the service layer at node N1 map theconnection to use 1 out of 2 slots, as shown in Fig. 8.At this step, the service layer also sorts the list of pos-sible HoBs that can serve this service. Using this al-gorithm, the setup message is transmitted to the cho-sen HoB that is in charge of scheduling all the PAslots for this DAOBS light-tree (3). Upon receiving thesetup message, the HoB allocates, if possible, the slotsfor the connection from N1. It is worth noting thatmany different scheduling algorithms can be imple-mented in order to satisfy a variety of service require-ments, such as, jitter, delay, bandwidth, call blockingprobability, etc. After processing the slot scheduling,the HoB acknowledges the node whether the connec-tion has been scheduled or not by means of an ac-knowledgment message (4). The service layer at N1 isnotified about the availability and successful setup ofthe connection (5), in which case, the application canstart transmitting its data packets. The transmissionof bursts will be made according to the PA reservedslots allocated to node N1. When the service finishes,the service layer initiates the disconnection (6) bysending a release message to the HoB node (7), whichliberates the allocated slots for the connection. An ac-knowledgement message is then transmitted back tonode N1 (8) to inform the service layer of the success-ful release of the connection (9). An explicit acknowl-edgment signal can also be sent to the applicationlayer (10).

One of the advantages of the DAOBS PA slot alloca-tion is that even using an explicit release, eventualupcoming PA slots that would remain unused can bechanged by the designated core node and announcedas a QA slot to the rest of the downstream nodes. Forinstance, in Fig. 8, the node N1 initiates the release ofthe connection sending a message to the HoB. Mean-while, a PA slot has already been created and sched-uled at the HoB following the information stored inthe connection table, which is not updated yet due tothe message propagation delay. In such a case, if thePA slot belongs to a connection from N1, and the con-nection is in the release state, the node can change thetype of slot to be used by other downstream nodes oreven reuse it to transmit one of its bursts through theQA access.

In order to reduce the connection establishment de-lay, a cross-layer algorithm in the service module sortsthe list of HoBs available to serve the TDM connectionso that light-trees for whose own service request nodeis actually a HoB are given preference. As a result, thesetup can be processed between the service and MAClayers in the node itself avoiding a two-way connec-tion establishment, hence reducing the connectionsetup delay.

Along with the PA channel access, the MAC layer inhe HoB nodes is enhanced with a slot schedulinganager for allocating the connections for all its

ownstream nodes, including itself. The allocation oflots is made along a supercycle slot window of a sizehat depends on the minimum and maximum possibleandwidth request. For instance, with a 10 Gbpshannel and a minimum connection request of55 Mbps, equivalent to an OC-3 of a synchronous op-ical network (SONET), the size of the supercycle win-ow will be

W = � 10 � 109

155 � 106 � = 64 slots . �2�

With regard to the specific scheduling algorithms,ig. 9 shows three different algorithms for a 4-slotDM connection: (1) a request for 4-slot service overrst-fit consecutive slots (FF); (2) a request for 4-slotDM over a periodic number of slots (SPFF); or (3) anllocation of a 4-slot bandwidth guaranteed servicever a number of slots not necessarily periodic or con-ecutive, but random (NCR). FF-based algorithms cane considered in order to support not only applicationsith guaranteed bandwidth requirements, but alsoith minimum delay variation between supercycle pe-

iods. Periodic scheduling is also necessary to provideitter-controlled characteristics for other multimediaervices with periodicity within the supercycle.

The periodic scheme is implemented by the slidingeriodic first-fit (SPFF) algorithm. In SPFF, the num-er of requested slots N must obey that size�W�0�mod N�, where size�W� is the size of the slot win-

ow W (in number of slots). Equation (3) representsathematically the set of slots that guarantee the

onstraints imposed by the SPFF algorithm. Thecheduling is successful if a group of slots Si can beound such that each slot si is free �v�si�=0� in the win-ow fulfilling the periodicity constraint, t=W /N:

Si = �si � W�v�si� = 0 ∀ si+1 = si + t,t =W

N,1 � i � N� .

�3�

Fig. 9. (Color online) PA slot scheduling algorithms.

pwaaswlwsia

crcptbt

mbt

tcwQsana

A

mqt1flb5ssrse�oFtcpt


Figure 10 shows in detail the SPFF algorithm. Inthe algorithm pseudocode, the input Sw stands for thepredefined amount of sliding slots for every executionof the algorithm; S0 is the last reserved slot from theprevious algorithm invocation; and sp, ss, and sj areauxiliary pointers used by the algorithm to check thefree slots that fulfill the constraints from Eq. (3). Line(6) captures the last reserved slot used in the previousinvocation of the algorithm and adds to it a number ofslots �Sw� so that the initial slot to be checked movesahead (slot sliding). From that point, the availabilityof free slots is checked for the number of slots Npassed as a parameter to the algorithm until thegroup of slots, Si, is found. Depending on N, the num-ber of iterations �t� in the while loop changes, hencethe smaller the number of requested slots N, thegreater the possibilities to find Si.

One of the benefits of this algorithm is that, not onlydoes it accomplish the periodicity for the slot schedul-ing within the slot window, but it also enhances thesuccess rate for n-slot connections. Now the connec-tions demanding very few slots, m (where n�m), areallocated along the slot window reducing the collisionwith the required space demanded by other n-slot con-nections.

IV. RESULTS

This section analyzes the performance of the multi-service QoS-enabled MAC protocol proposed in this

Fig. 10. SPFF scheduling algorithm.

aper. To this end, simulations are conducted on theell-known NSFNET network composed of 14 nodesnd 21 bidirectional links. In such a scenario, therere 16 wavelengths with 10 Gbps per channel. We as-ume in all the examples that the network neither hasavelength converters nor FDLs for contention reso-

ution, hence burst transmissions are subject to theavelength continuity constraint. Regarding the

etup of hardware devices, the control packet process-ng time and the nonblocking matrix switching timere set to 10 and 5 �s, respectively.

With respect to the traffic characteristics, bursts arereated at each node by using a volume-based algo-ithm [19] with an input packet arrival Poisson pro-ess and fixed size per burst of 100,000 bytes. For sim-licity, the burst destination is uniformly distributedo all the remaining nodes, so that the probability of aurst being sent to any other node in the network ishe same.

Results are gathered using the batch meansethod. 95% confidence intervals were also obtained,

ut since they are quite narrow, they have been omit-ed in order to improve the readability of the graphs.

The results section is divided into four main subsec-ions. First, insight into the performance of the QA ac-ess mode as a function of the local queue size andithout QoS is provided. Second, an analysis of theA access mode with QoS is presented. The next sub-

ection evaluates the performance of the whole MACrchitecture with both connection-oriented and con-ectionless burst services. Finally, a critical analysisbout the proposed service classification is presented.

. Results of the QA Access Mode Without QoS

The initial results deal with the protocol perfor-ance of the QA access mode when different local

ueue (LQ) sizes are used and only one connectionlessraffic class is transmitted on the network. Figure1(a) shows the burst blocking probability (BBP) as aunction of the total offered load to the network in er-angs per wavelength (Er/wl). The LQ sizes (in num-er of bursts) used throughout the simulations are 2,, 100, and 1000 bursts. Intuitively, the smaller theize of the LQ, the sooner the blocking probabilitytarts rising. For sizes between 2 and 100 bursts, theesults at high loads asymptotically converge to theame value. Only for the case in which the LQ size isquivalent to 1000 bursts can we see an improvement50%� of the blocking probability but at the expensef increasing its size by nearly 2 orders. Furthermore,ig. 11(b) represents the mean access delay (in ms) for

he same group of LQ lengths. At the expense of de-reasing the mean blocking probability, the delay ex-erienced when the LQ size is of 1000 bursts rises upo nearly 30 ms at very high loads. Thus, a trade-off

tsfibahahhcpsBt

ftdwt1aWsc2etfi

eocatmbctdg

bshNdf


between BBP and access delay comes up. At low-to-medium offered loads, the BBP improvement forlarger LQ is considerable. However, at high loads, anaverage access delay 15 times greater may not justifya little BBP improvement of about 50%.

B. Results of the QA Access Mode With QoS

In this section, we give insight into the results ofthe DAOBS QA access mode when a number of differ-ent connectionless burst traffic classes are transmit-ted on the network. In this experiment the LQ size islimited to 5 bursts. The following notation is used inthe graphs: burst priority classes are numbered from0 to 2, with class 2 being the highest-priority traffic.To demonstrate the performance of the protocol on dif-ferent traffic scenarios, two traffic class distributionshave been tested, as shown in Table I.

Figure 12(a) shows the BBP as a function of the of-fered load in the two traffic distributions. Class 2blocking probability is not plotted on the graph be-cause in both cases it is null for the whole load range.The rest of classes get different results when changing

0.0001

0.001

0.01

0.1

1

0 2 4 6 8 10 12 14

Bur

stbl

ocki

ngpr

obab

ility

Offered load (Er/wl)

2 er/wl

50% impr.

LQ 2 burstsLQ 5 burstsLQ 100 burstsLQ 1000 bursts

(a)

0

5

10

15

20

25

30

0 2 4 6 8 10 12 14

Mea

nac

cess

dela

y(m

s)


LQ 2 burstsLQ 5 burstsLQ 100 burstsLQ 1000 bursts

(b)

Fig. 11. (Color online) QA access mode without QoS for differentLQ lengths: (a) burst blocking probability, and (b) mean accessdelay.

he traffic configurations. Besides, we can see in theecond distribution that when the higher-priority traf-c load is decreased with respect to the total, theurst blocking probability decreases for both class 0nd class 1. Intuitively, the lower the class 2 (i.e., theighest priority) traffic load, the more resources avail-ble for the rest of the classes. Consequently, lessigh-priority traffic preempts lower-priority traffic;ence the low-priority traffic loss probability de-reases. Nevertheless, at high load the mean BBPlots of both traffic distributions tend to converge toimilar values, which ensures a predictable averageBP performance of the protocol whatever traffic dis-

ribution is present on the network.

Figure 12(b) compares the mean access delay as aunction of the offered load. In both traffic configura-ions, class 2 bursts not only have the lowest accesselay, but also get a delay nearly constant for thehole load range. Class 0 bursts have a similar delay

rend for both traffic configurations, and, finally, classbursts (i.e., the intermediate class) get a different

ccess delay depending on the traffic distribution.hen the class 2 proportion is 10% and class 1 repre-

ents 20% of the total traffic on the network, the ac-ess delay for class 1 resembles more the delay of class. Thus, it can be concluded that the delay experi-nced by a burst traffic class depends on the aggregateraffic between itself and all its higher-priority traf-cs.

Figure 12(c) shows a different performance param-ter. In the graph, the probability that a certain classf traffic is being transmitted from a HoB node isounted. Clearly, we can observe that class 2 burstsre mainly transmitted from HoBs, whereas theransmission rate for class 0 and class 1 decreases al-ost linearly as the load is increased. High-priority

urst traffic is more often transmitted from a HoB be-ause this node tends to see the capacity channel ofhe network as more available while the requests fromownstream nodes are not yet received due to propa-ation delays.

Figure 12(d) shows the mean route length (in num-er of hops) for the three burst priorities. As can beeen in the graph, the highest-priority bursts almostave a constant route length for the entire load range.onetheless, the lowest-priority bursts experience arop in the path length that is directly related to theact pointed out in Fig. 12(c). Moving class 0 burst

TABLE ITRAFFIC DISTRIBUTION CONFIGURATIONS

Distribution Class 0 Class 1 Class 2

1 50% 30% 20%2 70% 20% 10%

balFasfila


transmissions from the HoB to inside the DAOBSlight-tree brings the origin node closer to the destina-tion, which at most can be one of the ToBs, and there-fore the route is shortened.

With respect to the burst end-to-end delay (in ms),Fig. 12(e) shows a comparison between the threeclasses. In this specific scenario, an average end-to-end delay between 1.5 ms is kept between the threetraffic classes still providing a clear differentiation of

0.0001

0.001

0.01

0.1

1

0 2 4 6 8 10 12 14

Bur

stbl

ocki

ngpr

obab

ility


Class 0 dist. 1Class 1 dist. 1Mean dist. 1Class 0 dist. 2Class 1 dist. 2Mean dist. 2

(a)

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12 14

HoB

Tx

prob

abili

ty


Class 0 dist. 1Class 1 dist. 1Class 2 dist. 1

(c)

11.5

12

12.5

13

13.5

0 2 4

Mea

nen

d-to

-end

(ms)

Offe

Class 0 disClass 1 disClass 2 dis

(e)

Fig. 12. (Color online) QA access mode with QoS and LQ=5 bursdistributions, (b) access delay, (c) HoB transmission probability distdelay distribution 1.

urst loss probability. Specifically, class 0 bursts seen increase in their end-to-end delay as the offeredoad rises, according to the access delay increase inig. 12(b). However, we can see a trend change fromn offered load of 6 Er/wl onwards where the delaytarts falling. Following the reasoning from previousgures, this fact can be explained as follows: at high

oads, class 0 bursts get a higher blocking probabilitynd those that are able to get to a destination tend to

0

0.5

1

1.5

2

2.5

0 2 4 6 8 10 12 14

Mea

nac

cess

dela

y(m

s)




b)

1.6

1.8

2

2.2

2.4

0 2 4 6 8 10 12 14

Mea

nnu

mbe

rho

ps



d)

8 10 12 14

load (Er/wl)

(a) burst blocking probability comparison between the two traffiction 1, (d) mean number of hops distribution 1, and (e) end-to-end

(

(

6

red

t. 1t. 1t. 1

ts:ribu

astgttls1aTsne2sc

wcmbtdtiw

bTiIfcsaar

I

I

V

L


follow a shorter path, i.e., fewer number of hops,hence decreasing the mean delay even though the ac-cess delay goes up as shown in Fig. 12(b). The rest ofthe traffic classes (class 1 and class 2) experience anend-to-end delay only affected by the access delay in-crease since both of them show almost a constantmean route length within the load range under con-sideration [see Fig. 12(d)].

In order to evaluate the performance not only interms of the offered traffic load, but also as a functionof the carried traffic load, Fig. 13 shows the blockingprobability as a function of the average network loadratio �=i=1

n Mi /i=1n Ci, where Mi is the traffic carried

by link i, Ci is the capacity of link i, and n is the num-ber of links (bidirectional) on the network. The net-work load is represented between 0 and 0.5 and com-puted using the link utilization results from the samesimulation runs. As in Fig. 12(a), class 2 is lossless forthe considered network load; hence it is not repre-sented in the graph. For the rest, the protocol providesdifferentiation of up to three orders of magnitude be-tween class 0 and class 1 at a load of 0.3, and an im-provement of 2 orders for class 1 traffic between thetwo distributions at load 0.45. Based on the QoS re-quirements shown in Table II and providing that thepacket loss rate (PLR) can be approximated by theBBP for fixed size bursts, the protocol can guaranteethe QoS of a diverse number of applications. High orvery high loss sensitive traffic (e.g., grid applicationsor live video broadcasting) can be mapped as class 1for a great network load range (up to 45% for the sec-ond traffic distribution) or even be mapped as class 2traffic, which is loss-free.

C. Results With Mixed Connectionless andConnection-Oriented Burst Services

This subsection analyzes the network performanceof the proposed OBS MAC protocol using the three

1e-005

0.0001

0.001

0.01

0.1

1

0 0.1 0.2 0.3 0.4 0.5

1e-4

3e-3

5e-2

Bur

stbl

ocki

ngpr

obab

ility

Network load

1e-3

1e-2

Class 0 dist. 1Class 1 dist. 1Class 0 dist. 2Class 1 dist. 2

Fig. 13. (Color online) QA access mode BBP as a function of thenetwork load.

forementioned slot scheduling algorithms in a multi-ervice scenario where connectionless bursts sharehe capacity of the channel with connection-orienteduaranteed bandwidth services (TDM). Regarding theraffic characteristics, bursts are of fixed size for bothypes of service (connection-oriented and connection-ess) and equal to the slot size �100 kB�. For the TDMervices, we consider two different connection types of55 Mbps (S-155) and 622 Mbps (S-622), equivalent tobandwidth capacity of OC-3 and OC-12, respectively.he former is mapped to use 1 slot every 64, and theecond to use 4 out of 64 slots. In both cases, the con-ection call arrivals follow a Poisson process with anxponentially distributed holding time of 100 and00 ms, respectively (intentionally short to speed upimulations). The average load generated by theonnection-oriented calls is calculated as follows:

ACO =1

Ci=1

n

�i

1

�ibi, �4�

here C denotes the channel capacity in bps, �i is theall arrival rate of connection service i, 1 /�i is theean holding time, and, finally, bi is the demanded

andwidth of service i in bps. Finally, two differentraffic distributions have also been considered; first, aistribution in which TDM services represent 20% ofhe offered load and connectionless bursts the remain-ng 80% (Dist. 20%–80%), and second, a distributionith 40% and 60% (Dist. 40%–60%), respectively.

Figures 14(a) and 14(b) show the mean burst/calllocking probability of both service types (bursts andDM) as a function of the offered load to the network

n Er/wl and distributed by the scheduling algorithm.n Fig. 14(a) we can see that the BBP is very similaror the three simulated scheduling algorithms, espe-ially at high loads, where the three converge to theame values. Only at low loads do SPFF and NCR getslight improvement with respect to the remaining

lgorithm. In the NCR, the allocation of PA slots isandom within the supercycle slot window, so that in

TABLE IIAPPLICATIONS’ QOS REQUIREMENTS [20]

Application Delay JitterLoss sensitivity

(PLR)

nteractiveaudio/video

�150 ms �75 ms High ��1e−3�

nteractivetransactiondata

�50 ms �10 ms High (game �5e−2)to very high(grid �1e−4)

ideo/audiostreaming

�2 s �40 ms High ��3e−3� to veryhigh (live video�1e−4)

egacyapplications

Notspecified

Notspecified

Low

cc

lttcdtarsSrntunvaccs

ect


between them connectionless bursts can be transmit-ted with lower waiting delay, thus decreasing theblocking probability. Regarding the call blocking prob-ability (CBP) of the TDM services, this always re-mains below the BBP because the MAC protocol givespreference to the allocation of PA slots. Moreover, themean CBP of SPFF is the greatest among the three al-gorithms, especially influenced by the call blocking ofthe n-slots connections due to the periodicity restric-tions imposed by Eq. (3). Although the SPFF perfor-mance is not especially high, this type of algorithmprovides extra functionalities for services that requirerestricted jitter and delay variations.

Furthermore, comparing both figures, we can seethat the greater the connection-oriented offered traf-fic, the higher the BBP for the connectionless bursttraffic and the higher the CBP for the TDM services.For instance, increasing the TDM load from 20% to40% reduces the blocking probability difference be-tween connectionless and connection-oriented calls toless than 1 order at high loads. This is due to the con-siderable increase of the CBP (by 1 order of magni-tude). As a result, doubling the TDM traffic load in-

0.0001

0.001

0.01

0.1

1

0 2 4 6 8 10 12 14

Bur

st/C

allb

lock

ing

prob

abili

ty


1 order

2 orders

Dist. 20-80%Bursts FFCalls FFBursts NCRCalls NCRBursts SPFFCalls SPFF

(a)

0

0.5

1

1.5

2

2.5

3

3.5

4

0 2 4 6 8 10 12 14

Bur

stac

cess

/Cal

lset

upde

lay

(ms)


Dist. 20-80%Bursts FFBursts NCRBursts SPFFS-155 FFS-155 NCRS-155 SPFFS-622 FFS-622 NCRS-622 SPFF

(c)

Fig. 14. (Color online) Dual PA/QA access modes. (a) Mean burst/with traffic distribution 40%–60%. (c) Mean access delay for connect20%–80%. (d) HoB connectionless burst transmission and HoB conn

reases both BBP and CBP; hence a decrease of thearried traffic load is also expected.

In Fig. 14(c) the mean access delay for connection-ess bursts and the mean connection setup delay forhe connection-oriented services in the first traffic dis-ribution scenario (20%–80%) is shown. In general,onnectionless bursts see an increase in their accesselay due to the decrease of channel capacity left byhe allocation of TDM services. Although the resultsre very similar using the three scheduling algo-ithms, a slightly shorter delay in the NCR can be ob-erved. The connectionless burst access delay in thePFF is slightly greater than the other two algo-ithms because in this case a greater number of con-ections are required by the core nodes in the light-ree; hence less network capacity is left to otherpstream nodes between the HoB and the origin coreodes. With respect to the connection-oriented ser-ices, as expected, S-622 connections have a higherverage call setup delay because a greater number ofonnections are initiated by the core nodes due torankback. In such a case, the more restrictive thecheduling constraints, the longer the setup delay. For

0.0001

0.001

0.01

0.1

1

0 2 4 6 8 10 12 14

Bur

st/C

allb

lock

ing

prob

abili

ty


> 1 order

< 1 order

Dist. 40-60%

Bursts FFCalls FF

Bursts NCRCalls NCR

Bursts SPFFCalls SPFF

b)

0.4

0.5

0.6

0.7

0.8

0.9

1

0 2 4 6 8 10 12 14

HoB

tx/r

eser

vera

te


Dist. 20-80%

Bursts FFBursts NCRBursts SPFFS-155 FFS-155 NCRS-155 SPFFS-622 FFS-622 NCRS-622 SPFF

d)

blocking probability (BP) on traffic distribution 20%–80%. (b) BPless bursts and mean connection setup delay on traffic distributionion setup rates on traffic distribution 20%–80%.

(

(

callion

NnbdlSsn

uddrtwamaotwq

D

spacanmoolbw

Fc


instance, because the SPFF algorithm constraints arethe most restrictive, the setup delay for the n-slot con-nections is the largest. Unlike SPFF, the NCR algo-rithm allows the bandwidth connection to be set uprandomly using the void slots on the slot window;hence it is less restrictive and therefore easier for then-slot connections to be allocated.

Figure 14(d) shows the HoB burst transmission ratefor the connectionless services and the connectionsetup rate when the origin node of the connection-oriented services is the HoB of the light-tree. This isan interesting performance parameter because it ex-plains many of the results from previous paragraphs.In general, the higher the offered load to the network,the lower the HoB connectionless burst transmissionrate. Besides, we can also see that the reservationrate of S-622 connections using the SPFF algorithm isthe lowest in comparison with the other two algo-rithms, which corroborates the performance of themean call setup delay seen in Fig. 14(c).

Finally, Fig. 15 shows in more detail the CBP dis-tributed by algorithm and connection type, but now asa function of the mean network load. The CBP ofS-622 connections using the SPFF gets the worst per-formance. This is because, even when enhancing thescheduling algorithm with slot sliding, the fact thatthe scheduling of slots must be done in accordancewith Eq. (3) considerably reduces the slot allocationsearch space for n-slot connections. On the contrary,the S-155 connections that only require 1 slot can eas-ily be allocated, and in fact, the blocking probability ofS-155 turns out to be null for the whole load range.Moreover, the allocation of 1-slot connections decre-ments even more the search space for the n-slot con-nections, hence diminishing the setup successfulnessratio of these second connections. Regarding the othertwo scheduling algorithms, the CBP of S-622 connec-tions using FF is higher than NCR because the slot al-location is continuous in the FF service type, whereas

0.0001

0.001

0.01

0.1

0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Cal

lblo

ckin

gpr

obab

ility

Network load

Dist. 20-80%

S-155 FFS-155 NCR

S-155 SPFFS-622 FF

S-622 NCRS-622 SPFF

Fig. 15. (Color online) BP for different connection types on trafficdistribution 20%–80%.

CR takes advantage of randomly scheduling the con-ections among the void slots so that a greater num-er of options are available given the same slot win-ow size. However, S-155 connections get a slightlyower CBP using the FF, merely due to the fact that-622 connections get a higher CBP; hence more freepace is available to allocate the remaining 1-slot con-ections.

Each algorithm delivers its optimum performancender specific traffic conditions. For this reason, intro-ucing a scheduling machine able to make dynamicecisions and autonomously select the optimum algo-ithm may easily improve the overall PA slot reserva-ion performance. The selection of the algorithmould need to take into account the type of servicend the network status, e.g., resource utilization, toake sure that the slot reservations can be efficiently

ccommodated in the spare channel capacity. More-ver, in order not to disrupt the service of live connec-ions (e.g., already set up), the algorithm optimizationould need to be processed only for new connection re-uests.

. PA/QA Access Mode Transmission Decision

In view of the results from previous analysis, in thisubsection, a service and access mode assignment isroposed using the two DAOBS access schemes (PAnd QA) and the QoS differentiation within the QA ac-ess mode itself. First of all, it is of interest to evalu-te the burst blocking performance under similar sce-arios. Figure 16 gives the BBP as a function of theean offered load for the connectionless bursts (80%

f the load) when these are transmitted together withther connection-oriented TDM services (20% of theoad). Likewise, the mean BBP of the connectionlessurst with only QA access supporting multiple QoSith the first traffic distribution from Table I is also

1e-005

0.0001

0.001

0.01

0.1

1

0 0.1 0.2 0.3 0.4 0.5

Bur

stbl

ocki

ngpr

obab

ility

Network load

with20%

TDMtra

ffic

with20%

class

2 traffic

PA/QA sim. dist. 20-80% (NCR)QA with QoS sim. dist. 1

ig. 16. (Color online) BBP comparison between standalone QA ac-ess with QoS and PA/QA dual access modes.

awdrTafhitrihemtmamatblmcgdCisterti

TSETEs


represented. We have taken these two plots because inboth cases the highest-priority traffic (connection-oriented traffic with PA access in the first case, andclass 2 traffic in the second case) represents 20% ofthe offered load. As we can see, at high loads, the twoplots lay within the same value range, with veryslight differences between them. Therefore, it is ex-pected that the class differentiation between the twolowest-priority traffics in the QA access mode will per-form similarly even in the case that the channel isalso shared with connection-oriented traffic using PAaccess, as long as the load of this traffic class is simi-lar to class 2 traffic.

Turning to the performance of the highest-prioritytraffic in the circumstances mentioned in Fig. 16, thedecision making of which traffic should be assigned asclass 2 connectionless or as connection-oriented, de-pends specifically on the type of traffic and service re-quired. In Subsection IV.B, the QoS differentiationprovided by the QA access mode has been analyzedconcluding that for both traffic distributions class 2 islossless for the whole load range. On the contrary,from Subsection IV.C the connection-oriented traffic,which is given preference using the PA access mode,gets a nonnull CBP starting from medium offeredloads but is still lower than the blocking probability ofconnectionless bursts. In view of this, class 2 QA traf-fic is envisioned for very high-priority connectionlesstraffic, such as control traffic without explicit jittercontrol guarantees, whereas PA access mode can bemore suitable for services demanding a guaranteedbandwidth for a long period (in comparison with theburst size) that under some circumstances may alsodemand jitter and delay guarantees. As a concludingremark, Table III resumes the usage of each class ofservice provided by the multiservice DAOBS architec-ture.

TABLE IIIDAOBS CLASSES OF SERVICEa

Mode CoS Type Use Application

QA Class0

CL Best-effort trafficwithout lossguarantees or delay

Transactionaldata

Class1

CL Priority traffic thattolerates some losseswithout guaranteeddelay

Interactiveaudio/video

Class2

CL Very high-prioritytraffic that does nottolerate losses at all

Interactivedata,control

PA Class3

CO Long-life TDM trafficwith guaranteedbandwidth and delay

IPTV, cloudcomputing,grid

aCL, connectionless; CO, connection-oriented.

V. CONCLUSION

This paper has introduced and extensively analyzednovel multiservice MAC protocol for OBS mesh net-orks. An advantage of this protocol is its integratedesign that permits it to potentially serve a broadange of applications with diverse QoS requirements.he MAC provides two main access methods: queuerbitrated for connectionless bursts and prearbitratedor TDM connection-oriented services. On the oneand, the queue-arbitrated access is based on a count-

ng and monitoring process of burst and request con-rol packets traveling in opposite control channel di-ections and a distributed preemption-based schemen a multiqueue access priority system. On the otherand, the prearbitrated mode is based on the preres-rvation of slots supported by an upper service layerodule. Apart from its simplicity and low implemen-

ation complexity, the protocol also offers great perfor-ance regarding the relative differentiation of QoS

mong different burst traffic classes in the QA accessode for both QoS parameters, burst blocking prob-

bility, and end-to-end delay. Results evaluatedhrough simulations show that the highest-priorityursts are guaranteed zero losses and very low accessatencies in the QA access mode. Even for the inter-

ediate traffic class, the protocol can guarantee an ac-eptable BBP for a diverse number of applications. Re-arding the PA access mode, results show thatoubling the offered TDM traffic load increases theBP by more than 1 order, slightly affecting the block-

ng of connectionless bursts. Moreover, three differentlot scheduling algorithms for allocating TDM connec-ions along with connectionless bursts have also beenvaluated, providing diverse results depending on theequested bandwidth. The overall results demonstratehe suitability of the proposed architecture for futurentegrated multiservice optical networks.

ACKNOWLEDGMENTS

his work has been supported by the Government ofpain through project TEC2009-13901-C02-01, theC-FP7 Euro-NF, the i2CAT Foundation’s projectARIFA, and by the Government of Catalonia and theuropean Social Fund through a predoctoral scholar-hip (FI) and a BE fellowship.

REFERENCES

[1] G. Zervas, Y. Qin, R. Nejabati, D. Simeonidou, F. Callegati, A.Campi, and W. Cerroni, “SIP-enabled optical burst switchingarchitectures and protocols for application-aware optical net-works,” Comput. Netw., vol. 52, no. 10, pp. 2065–2076, July2008.

[2] Y. Chen, C. Qiao, and X. Yu, “Optical burst switching: a newarea in optical networking research,” IEEE Network, vol. 18,no. 3, pp. 16–23, May 2004.

[3] J. Triay, J. Perelló, C. Cervelló-Pastor, and S. Spadaro, “On

Eftv

afrwtGa(

si(aph

DwavcSWdivHcad


avoiding-minimizing burst collisions in optical burst-switchednetworks without wavelength conversion,” Proc. of 11th Int.Conf. on Transparent Optical Networks (ICTON), Azores, Por-tugal, 2009, paper Mo.D3.4.

[4] IEEE Standards Board, “Distributed queue dual bus (DQDB)subnetwork of a metropolitan network—802.6,” Dec. 1990.

[5] H. Overby, “Quality of service differentation: teletraffic analy-sis and network layer packet redundancy in optical packetswitched networks,” Ph.D. dissertation, Dept. of Telematics,Norwegian University of Science and Technology, Trondheim,Norway, 2005.

[6] M. Yoo, C. Qiao, and S. Dixit, “QoS performance of optical burstswitching in IP over WDM networks,” IEEE J. Sel. Areas Com-mun., vol. 18, no. 10, pp. 2062–2071, Oct. 2000.

[7] H. Overby, N. Stol, and M. Nord, “Evaluation of QoS differen-tiation mechanisms in asynchronous bufferless optical packet-switched networks,” IEEE Commun. Mag., vol. 44, no. 8, pp.52–57, Aug. 2006.

[8] Q. Zhang, V. M. Vokkarane, J. P. Jue, and B. Chen, “AbsoluteQoS differentation in optical burst-switched networks,” IEEEJ. Sel. Areas Commun., vol. 22, no. 9, pp. 1781–1795, 2004.

[9] X. Liu, C. Qiao, W. Wei, and T. Wang, “A universal signaling,switching and reservation framework for future optical net-works,” J. Lightwave Technol., vol. 27, no. 12, pp. 1806–1815,June 2009.

[10] J. A. Hernández, P. Reviriego, J. L. García-Dorado, V. López, D.Larrabeiti, and J. Aracil, “Performance evaluation and designof polymorphous OBS networks with guaranteed TDM ser-vices,” J. Lightwave Technol., vol. 27, no. 13, pp. 2495–2505,July 2009.

[11] L. Xu, H. G. Perros, and G. N. Rouskas, “A simulation study ofoptical burst switching and access protocols for WDM ring net-works,” Comput. Netw., vol. 41, no. 2, pp. 143–160, Feb. 2003.

[12] W. P. Chen, W. F. Wang, and W. S. Hwang, “A novel and simplebeforehand bandwidth reservation (BBR) MAC protocol forOBS metro ring networks,” J. High Speed Networks, vol. 17,no. 1, pp. 59–72, Jan. 2008.

[13] H. T. Lin and W. R. Chang, “CORnet: an OBS metro ring net-work with QoS support and fairness control,” Comput. Netw.,vol. 52, no. 10, pp. 2045–2064, July 2008.

[14] L. M. Peng and Y. C. Kim, “Investigation of the design of MACprotocols for TT-TR-based WDM burst-switched ring net-works,” J. Opt. Commun. Netw., vol. 1, no. 2, pp. 25–34, July2009.

[15] Z. Zhang, L. Liu, and Y. Yang, “Slotted optical burst switching(SOBS) networks,” Comput. Commun., vol. 30, no. 18, pp.3471–3479, Dec. 2007.

[16] M. Klinkowski, D. Careglio, and J. Solé-Pareta, “Comparisonof conventional and offset time-emulated optical burst switch-ing,” Proc. 8th Int. Conf. on Transparent Optical Networks (IC-TON), Nothingham, UK, 2006, pp. 47–50.

[17] A. Gumaste, N. Ghani, P. Bafna, A. Lodha, A. Agrawal, T. Das,J. Wang, and S.-Q. Zheng, “DynaSPOT: dynamic services pro-visioned optical transport test-bed—achieving multirate multi-service dynamic provisioning using strongly connected light-trail (SLiT) technology,” J. Lightwave Technol., vol. 26, no. 1,pp. 183–195, Jan. 2008.

[18] B. Mukherjee, Optical WDM Networks, 1st ed. Springer, 2006.[19] A. Rostami and A. Wolisz, “Modelling and synthesis of traffic in

optical burst-switched networks,” J. Lightwave Technol., vol.25, no. 10, pp. 2942–2952, Oct. 2007.

[20] IST IP Nobel Phase 2: Next Generation Optical Networks forBroadband European Leadership Phase 2, Project deliverableD1.1: “Architectural Vision of Network Evolution,” Aug. 2006.

Joan Triay received a B.Eng. and a M.Eng.in telecommunications engineering and aM.Sc. in telematics engineering, in 2004,2006, and 2007, respectively, all from Uni-versitat Politècnica de Catalunya (UPC),Spain. In 2007, he was awarded with a4-year predoctoral FI scholarship from theGovernment of Catalonia, and since then hehas been a Ph.D. candidate in the Depart-ment of Telematics Engineering at UPC. Hewas a visiting fellow at the University of

ssex (UK) from June 2009 to February 2010 thanks to a BE-DGRellowship. His research interests include, but are not limited to, fu-ure optical network architectures and the provisioning of multiser-ice capabilities on high-speed networks.

Georgios S. Zervas was awarded theM.Eng. degree in electronic and telecommu-nication systems engineering with distinc-tion and the Ph.D. degree at the Universityof Essex (UK) in 2003 and 2008, respec-tively. He is a Senior Research Officer in theHigh Performance Networks Group at theUniversity of Essex involved in current andpast EC-funded projects MAINS, STRON-GEST, GEANT3, BONE, Phosphorus MU-FINS, and e-Photon/One. He is author

nd co-author of over 50 papers in international journals and con-erences. His research interests include high-speed optoelectronicouter design, subwavelength networks (e.g., OBS), GMPLS net-orks, and grid networks. He is also involved in standardization ac-

ivities in the Open Grid Forum (OGF) as an active member of therid High Performance Networking Research Group (GHPN-RG)nd co-founder of the Network Service Interface Working GroupNSI-WG).

Cristina Cervelló-Pastor received herM.Sc. and Ph.D. in telecommunication engi-neering in 1989 and 1998, respectively, bothfrom the Universitat Politècnica de Catalu-nya (UPC), Barcelona, Spain. She is cur-rently an Associate Professor in the Depart-ment of Telematics Engineering at UPC,which she joined in 1989, and a leader ofthe optical networks research group withinBAMPLA. Her research trajectory has beencentered on the field of routing in high-

peed networks and the development of new protocols and servicesn OBS/OPS, taking part in diverse national and European projectsFEDERICA, ATDMA, A@DAN, Euro-NGI, Euro-FGI, EURO-NF)nd being responsible of various public and private funding R&Drojects, some of them with the i2CAT Foundation. In parallel sheas presented several patent proposals about OBS networks.

imitra Simeonidou is the Head of the High Performance Net-orks Group, which includes the Photonics Networks Laboratorynd the newly established Network Media Laboratory, at the Uni-ersity of Essex, UK. She joined Essex in 1998 (previously with Al-atel Submarine Networks). While at Alcatel, she held the post ofenior Principle Engineer and contributed to the introduction ofDM in long-haul submarine links and pioneered the design and

eployment of optical add–drop multiplexers. At Essex, she is lead-ng a group of 30 researchers and Ph.D. students and she is in-olved in numerous national and international research projects.er research is focusing in the fields of optical networks, grid and

loud computing, and the future Internet. She is author and co-uthor of over 350 papers, 11 patents, and several standardizationocuments.

multiservice qos-enabled mac for optical burst switching · multiservice qos-enabled mac for...

Documents