Computer Networks 56 (2012) 244–259

Contents lists available at SciVerse ScienceDirect

Computer Networks

journal homepage: www.elsevier .com/locate /comnet

Traffic-prediction-assisted dynamic bandwidth assignment for hybridoptical wireless networks

Maysam Mirahmadi, Abdallah Shami ⇑Department of Electrical and Computer Engineering, The University of Western Ontario, Canada

a r t i c l e i n f o

Article history:Received 24 January 2011Received in revised form 16 June 2011Accepted 12 August 2011Available online 16 September 2011

Keywords:Access networksDynamic bandwidth allocationOptical/wireless access networks

1389-1286/$ - see front matter � 2011 Elsevier B.Vdoi:10.1016/j.comnet.2011.08.018

⇑ Corresponding author.E-mail addresses: mmirahma@uwo.ca (M. Mir

g.uwo.ca (A. Shami).

a b s t r a c t

Hybrid optical-wireless networks provide the inexpensive broadband bandwidth, vital formodern applications, as well as mobility, and scalability required for an access network.However, in order to provide satisfactory Quality of Service (QoS) on such a non-homogeneous network, innovative designs are required.

This paper proposes a novel scheduling mechanism to significantly improve the delayguarantee, while maintaining high-level throughput, by predicting the incoming trafficto optical network units (ONU). The proposed scheduler managed to exploit the availableinformation in hybrid optical-wireless networks, to enhance the ONU scheduler. Thisresults in accurate prediction of incoming traffic, which leads to intelligent and traffic-aware, scheduling and dynamic bandwidth assignment (DBA).

Based on the proposed architecture, two DBA algorithms are proposed and their perfor-mance is evaluated by extensive simulations. Moreover, the maximum throughput of suchnetwork is analyzed. The results show that by using the proposed algorithms, the delaybound of delay-sensitive traffic classes can be decreased by a factor of two, without anyadverse effect on the throughput.

� 2011 Elsevier B.V. All rights reserved.

1. Introduction

In recent years, numerous bandwidth consuming appli-cations have emerged. These applications like Internet Pro-tocol TeleVision (IPTV) and Video-on-demand (VoD),require considerable bandwidth and are attracting wide-spread attention among consumers. Moreover, service pro-viders intend to offer voice, video, data and even wirelessservices to their customers based on an all-IP shared net-work infrastructure. As a result, a broadband access net-work capable of ensuring Quality of Service (QoS) fordifferent service types is required.

Optical fiber networks are capable of transferring mas-sive amounts of data, satisfactory for the everyday growingdemand of recent applications. Modern passive optical net-works, such as Gigabit Ethernet Passive Optical Network

. All rights reserved.

ahmadi), ashami@en

(GEPON) or Wavelength Division Multiplexed Passive Opti-cal Network (WDM-PON) can transfer several gigabits ofdata per second and yet they have less maintenance costin comparison to old optical networks, like SynchronousOptical Network (SONET). Nevertheless, it is still quiteexpensive to lay fiber to every user’s premise and build apurely optical network. It is impractical to build a trueoptical network, such as Fiber To The Home (FTTH), in acongested and built-up urban area, where considerablenumber of broadband users are.

On the other hand, wireless networks are relativelyinexpensive and have the unique feature of delivering datato mobile users. The drawbacks are that they offer muchless bandwidth and they are error-prone. Another disad-vantage of wireless networks is that wireless spectrum isshared among many users, further limiting the bandwidthoffered to each user.

The integration of optical and wireless networks offersan inexpensive broadband solution. In a hybrid network,

Fig. 1. Integrated EPON and WiMAX architectures [27] [moved accordingly].

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 245

such as the one shown in Fig. 1, the optical component isused to connect several wireless base stations as well asbroadband customers to the central office. The abundantbandwidth of optical network is divided among the wire-less base stations to serve their wireless users. In addition,hybrid optical-wireless networks offer several value-addedfeatures, such as redundancy, mobile access and coopera-tive communication which make it more appealing.

The hybrid optical-wireless networks seem to be thetechnology of future but some challenging issues still needto be addressed. Here, we focus on ensuring the QoS over ahybrid network, which is vital to offer a satisfactory ser-vice. The choice of how the scheduler assigns bandwidthto different flows has a tremendous impact on the QoSparameters. The scheduler typically relies on a dynamicbandwidth assignment (DBA) algorithm to divide the avail-able bandwidth between different flows.

In this work, we propose a method to enhance the DBAalgorithms in the optical part of hybrid optical-wirelessnetworks. A prediction method, which uses the informa-tion in the wireless domain, helps decrease the delaybound of delay-sensitive flows, and hence results in im-proved QoS. The extensive simulation-based studies notonly shed light on the factors, affecting the proposed algo-rithm, but also compare fixed and adaptive cycle lengthoptical DBA algorithms side by side.

The content of this paper addresses the integration ofWorldwide Interoperability for Microwave Access (Wi-MAX) and Ethernet Passive Optical Network (EPON) as suc-cessful technologies of their kind. However, most of thediscussions can be applied to any hybrid optical-wirelessnetwork.

The paper is organized as follows. Section 2 introducesthe system model. Section 3 compares advantages and dis-advantages of integration in hybrid optical-wireless net-works. The proposed algorithm is presented in Section 4and the maximum throughput of optical-wireless networkis analytically studied in Section 5. The traffic predictionassisted algorithms, used in simulation experiments, are

explained in Section 6, while Section 7 defines thesimulation platform. Section 8 analyzes the simulation re-sults and finally Section 9 concludes the paper.

2. System model

The topology of hybrid optical-wireless network is com-monly considered to be tree [2,5,14,16,27,30–32].

The root of the tree is the Optical Line Terminal (OLT)which is located at the central office (CO) and is directlyconnected to the core network. Each optical network unit(ONU), also called gateway [16], is connected to the OLTthrough one or several splitters as shown in Fig. 1. WiMAXBase Stations (BS) are connected to an ONU via a standardinterface, like Ethernet, or even implemented in the samebox. We refer to the latter case as ONU-BS. In practice,the WiMAX BSs share the optical bandwidth with otherusers, such as residential or corporate local area networks.

The WiMAX BSs provide wireless connectivity to theirsubscriber stations (SS) through either a single hop legacywireless network or a relay-based wireless network withmesh topology. In the latter case a routing algorithm, likethe one presented in [16], is necessary to effectively deliverthe packets.

Due to the intrinsic property of passive optical networks,downstream data, transmitted from OLT to ONU, is broad-cast to all ONUs. By contrast, upstream data, transmittedfrom the ONUs can only be received at the OLT. Thus, thedownstream and upstream links are considered as Point-to-MultiPoint (PMP) and point-to-point, respectively.

In cases where the wireless base station and the OLT arecollocated, the wireless and optical parts can be integrated,resulting in an intelligent network. The key point in theintegrated architecture is that the BS and the ONU canshare their internal data, such as detailed informationabout bandwidth requests, bandwidth allocation and pack-et scheduling, which can be used to improve the overallperformance. An intelligent scheduler at the integrated

246 M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259

ONU-BS should leverage this extensive information toachieve better performance and QoS guarantee.

3. Integration concept and literature review

Researchers have proposed a number of methods tointegrate wireless networks, such as WiMAX or LTE, withpassive optical networks. The idea of using information al-ready available at the wireless BS to assist the optical partis first introduced in [27]. But no specific methods or per-formance evaluation has been done. To the best of ourknowledge, none of the researchers devised a comprehen-sive method to use the information to improve the perfor-mance of the optical domain, and hence improve theoverall performance.

Some addressed hybrid optical-wireless network byproposing a single MAC frame and QoS structure on bothdomains [6,11] while others translated one technology’sQoS parameters into another’s [2,5,12,14,29,32]. Due tothe differences between optical and wireless systems, suchas their scalability requirements, it is too difficult to pro-pose a single QoS structure for both technologies withoutsacrificing scalability.

Sarkar et al. presented an architecture called Wireless-Optical Broadband Access Network (WOBAN) [17–24];This work proposed several approaches on the placementof ONU-BS nodes. The work presented in [6] addressedthe integration of WiMAX and Optical Burst Switching(OBS) and evaluated the performance for BE and UGS traf-fic flows. Another architecture, namely MARIN, that incor-porates multiple OLTs in the network was proposed in [26].

Ou et al. [14] devised a method to dynamically map theWiMAX classes to the Gigabit PON (GPON) classes. Thedesign was implemented using commercially availableequipments to deliver video on demand service to testusers. The work was later enhanced in [32] and an ExcessDistribution DBA scheme was adopted. Yan et al. proposeda DBA algorithm that takes channel condition into account[30]. Another channel aware algorithm was presented in[2]. The proposed scheduling algorithm considers wirelesscondition of the cell and head of line delay of rtPS queue aswell as queue length. The authors of [5] proposed a two-step mapping that combines BE and nrtPS into a priorityqueue and rtPS and ertPS into another one and then mapsthem to EPON classes of traffic. Obele et al. [11,12] pro-posed an architecture based on a unified QoS infrastructurewhich is similar to IEEE 802.16a. The performance of theirproposed architecture was analyzed under self-similar andlong-reach-dependent traffic loads. Wang et al. [29] andLuo et al. [9,10] also used a similar unified QoS schemethroughout the network. The work described in [31] usedthe time of next EPON polling to control the admission ofdelay-sensitive traffics.

Amongst the various challenges that ensuring QoS onEPON poses, the dynamic bandwidth assignment of the up-load channel attracts more research attention. Unlike thedownload channel where the bandwidth assigner has fullknowledge of the bandwidth needs, the upload bandwidthassigner, which is located at the OLT is not aware of realtime bandwidth requirements of each ONU. Therefore, a

mechanism is required to report real time bandwidthrequirements of each individual ONU to the OLT.

To exchange the bandwidth requirement information, arequest/grant mechanism, called Multi-Point Control Pro-tocol (MPCP), is used. In MPCP, ONUs send informationregarding their bandwidth requirements to the OLT. Forexample, in a widely used method, ONUs send their queuelength to the OLT at the end of the grant slot. OLT processesthe results and sends the bandwidth assignments back toONUs by grant packets. The problem is that MPCP intro-duces latency which degrades the performance of thebandwidth assigner, i.e., the bandwidth assigner in OLTcan only know the ONU bandwidth requirements withsome delay. The MPCP request is sent once in each EPONcycle. As a consequence, the packets that are received aftersending the request remain in ONU buffer for an extraEPON cycle.

Using information about the incoming traffic can allevi-ate the degradation of the extra delay. Integrated WiMAX/EPON provides us with a tool to give this information toOLT. Prior to sending upstream data, the user requeststhe required wireless bandwidth from the WiMAX BS.The BS scheduler assigns the available bandwidth basedon these requests and then sends the grant in UL map fieldof WiMAX frame. Therefore, the BS has full knowledge ofthe incoming data for at least the next coming frame. If thisinformation becomes accessible to the OLT scheduler, itcan be used to predict the incoming traffic which would re-sult in a more effective bandwidth scheduling. However, inorder to achieve optimal performance, an intelligent de-sign that considers the EPON cycle and WiMAX framelength as well as the traffic load and its classes is required.

3.1. Effect of sharing knowledge of incoming traffic on thedelay

Predicting the incoming traffic makes it possible to re-quest the required bandwidth for transmitting a packetthat has not yet arrived. At the next EPON cycle, due topre-assignment of the required bandwidth, the packetsthat have just arrived at the ONU, are transmitted to theOLT right away without experiencing any additional delayfor exchanging MPCP request/grant messages.

Since the WiMAX BS assigns bandwidth to the MobileSubscribers (MSs) at the beginning of each frame, it isaware of the quantity of data to be received during the fol-lowing frame. Fig. 2 compares the process of sending amarked packet in normal and integrated scheduling. As itcan be seen in the Fig. 2, predicting the packet reducesits queuing delay. This helps ensure QoS for delay sensitiveapplications.

In order to use knowledge of incoming traffic to opti-mize the bandwidth assignment in hybrid optical-wirelessnetworks, the length of the WiMAX frame and EPON cyclesshould be carefully adjusted. The essential condition forpredicting incoming traffic is that the WiMAX frame lengthshould be longer than the EPON cycle length. For betterillustration, consider Fig. 3(a) in which the WiMAX frameis considerably shorter than EPON cycles. Since the BSknows only the traffic of the current frame, it is not possi-ble to accurately predict the amount of traffic that accumu-

Fig. 2. Timing diagram of sending a marked packet (a) normal scheduling (b) integrated scheduling [moved accordingly].

Fig. 3. EPON cycle in comparison with WiMAX frames.

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 247

lates until the next EPON cycle. Therefore, only a fraction ofincoming traffic can be predicted. Hence, the performanceimprovement that could be gained from full knowledge re-cedes. By contrast, Fig. 3(b) illustrates the case in whichWiMAX frame is longer than EPON cycle. Hence, theONU-BS knows the quantity of data to be sent to the OLTat the next EPON cycle.

The key part of the proposed algorithm is estimating theamount of traffic that an ONU-BS will have to relay to theOLT in the next grant. Without lack of generality, we as-sume that all received data at the ONU-BS should be re-layed to the OLT, i.e., there is no local traffic at ONU-BS.Given that the amount of incoming traffic is known,ONU-BS requests OLT for the amount of bandwidth, de-noted by Br below.

Br ¼ Q þ BpðTnxÞ; ð1Þ

where Q denotes the current length of the queue andBpðTnxÞ is the predicted incoming traffic that will be re-ceived up to beginning of the next grant, which is repre-sented by Tnx.

It is essential to know the time of next transmission toestimate the bandwidth requirement. If EPON uses fixedlength cycles, the next time of transmission is simply cal-culated as,

Tnx ¼ Tg þ C; ð2Þ

where Tg is the grant window beginning time and C is thecycle length. It is worth mentioning that this method doesnot account for the variations that occur because of differ-ent ONU grant sizes in each EPON cycle. A more accurateapproach is to calculate Tnx based on the internal data ofOLT scheduler. This approach is applicable to both fixed-length cycle and adaptive-length cycle algorithms.

4. Algorithm description

The proposed algorithm consists of two phases. Thecore of the algorithm is the inter-ONU scheduling whichis performed at OLT scheduler. This part of the algorithmdoes not consider different classes of traffic and assignsthe bandwidth to each ONU based on the available band-width, ONU requests and traffic prediction. The details ofinter-ONU scheduling is described in Section 4.1.

The second part of the algorithm, namely intra-ONUscheduling, is performed at each ONU to assign the grantedbandwidth to each traffic class. The detail of this part of thealgorithm is explained in Section 4.2.

The algorithm is independent of the WiMAX BS sched-uler. Therefore, it does not impose any requirement onthe WiMAX scheduler. The only modification needed atWiMAX BS is that its upload bandwidth assignments, i.e.,UL map, should be accessible to the ONU scheduler.

Fig. 4. Relationship between F(t) and WiMAX frame.

Fig. 5. A representation of FðtÞ: (a) General case, (b) simplified.

248 M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259

Although choosing the WiMAX scheduler may affect theoverall performance, the study of WiMAX scheduling algo-rithms and their impact is beyond the scope of this work. Acomprehensive survey of WiMAX scheduling algorithmscan be found in [28].

4.1. Inter-ONU scheduling

Information about incoming traffic in the next WiMAXframe is abstracted in the function FðtÞ, which is definedas the aggregated traffic at time t. FðtÞ predicts the ONUqueue length at time t. It is an increasing function of timethat yields the amount of traffic that has been received upto time t. The relation between WiMAX frame and FðtÞ isillustrated in Fig. 4.

Every time a UL map is generated at the BS, FðtÞ is de-fined for the next WiMAX frame. Note that it models theworst case scenario and depending on the type of thescheduler used in BS, the real incoming traffic can beslightly less.

The summarized information, i.e., FðtÞ, is sent to the OLTalong with bandwidth requests to be used in the schedul-ing process. In order to limit message exchanges and keepthe OLT scheduler as simple as possible, the informationregarding the incoming traffic class is not transmitted tothe OLT. However, this can be simply implemented. FðtÞcan be modeled by a linear piecewise function of time.

FðtÞ ¼

a0t þ b0; t0 < t 6 t1;

a1t þ b1; t1 < t 6 t2;

a2t þ b2; t2 < t 6 t3;

..

.

ant þ bn; tn < t 6 tnþ1;

8>>>>>>><>>>>>>>:

ð3Þ

where ai and bi; i ¼ 1; . . . ; n are coefficients. To simplify thecalculations and prohibit time wrapping, time is typicallymeasured from a common origin such as beginning ofthe cycle. Note that FðtÞ is only defined in the predictionrange ðt0; tnþ1� which is discussed more in the nextparagraphs.

The slope of FðtÞ depends on the user’s transmissionmodulation and coding. In scenarios where all of the MSsuse the same modulation and coding technique and TDDduplexing is used, FðtÞ could be simplified to

FðtÞ ¼b0; t 6 t1;

a1t þ b1; t1 < t 6 t2;

b2; t2 < t 6 t3:

8><>: ð4Þ

This case is not unreal and it happens if the network oper-ates at maximum capacity, where all users transmit at themaximum modulation and coding rate. An example of FðtÞis shown in Fig. 5.

The scheduler uses FðtÞ to estimate the incoming trafficand grant the appropriate amount of bandwidth. Since it isassumed that TDD is used as the duplexing method, thereare active and idle periods. Data packets are arriving duringthe active periods, where SSs are allowed to send. In WiMAX,the active period for upstream is the part of the UL subframethat is granted to the SSs. During idle periods users are notallowed to send and hence FðtÞ remains constant.

FðtÞ ¼

b0; t0 < t 6 t1s ;

a1t þ b1; t1s < t 6 t1

e ;

a1t1e þ b1; t1

e < t 6 t2s ;

..

.

antne þ bn; tn

e < t 6 tnþ1s ;

8>>>>>>><>>>>>>>:

ð5Þ

where tis and ti

e are the beginning and end of ith activityperiod, respectively. From WiMAX TDD point of view,these are equivalent to beginning and end of the UL slot.Normally, the WiMAX scheduler puts together the grantslot of all users that transmit with the same modulationand coding. It also packs the grant at the beginning of theslot [13]. Therefore, it is not expected to have more thanone activity period in the prediction range. It also makesthe transmission of abstract data easier.

Considering the WiMAX rate, time division duplexingand request scheme, FðtÞ is,

FðtÞ ¼Q ; tr < t 6 t1

s ;

RW t þ Q � RW �minðt1s ; trÞ; t1

s < t 6 t1e ;

RW t1e þ Q � RW �minðt1

s ; trÞ; t1e < t < t2

s ;

8><>: ð6Þ

where RW is the WiMAX transmission rate, Q is the requestor equivalently, ONU’s queue length at the time of transmit-ting request, and tr is the time of request transmission. FðtÞ is

Fig. 6. Illustration of FðtÞ and request corrections method.

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 249

illustrated in blue in Fig. 6. At the time of sending the re-quest, FðtÞ is initialized to Q. Then it increases with the Wi-MAX transmission rate when the UL period begins. Theincrease continues until the end of UL period and then re-mains constant.

The prediction information is updated regularly in away that tr < t2

s , i.e., the updated prediction range overlapswith the previous one. Thus makes continuous predictionpossible.

Since the WiMAX frame is supposed to be longer thanthe EPON cycle, it is possible that receiving data continuesbetween the grant periods. The ideal solution is to considerthese newly arrived packets and extend the grant to trans-mit them, if possible. The grant can be extended until thebuffer becomes empty and all of the newly arrived packetsare transmitted.

In our proposed algorithm, the newly arrived data ispredicted with FðtÞ. In order to consider this prediction inthe dynamic bandwidth assignment algorithm, it is re-quired to update the requests, fed into the DBA, with thepredictions. Since the correction depends on the DBA re-sults, it is done iteratively. This way the DBA bandwidthassignments and request corrections are calculated jointly.

The DBA outcome without applying the request correc-tions is considered as the initial bandwidth assignment. Tocalculate the request correction, the granted bandwidth ismodeled with the same approach as incoming data. Thismodel, which we denote by GðtÞ, on the other hand, de-scribes transmitted bits. It is modeled as a line segmentwhich starts from tg with a slope of RE. tg represents the starttime of the granted window and RE represents the EPONtransmission rate. These models are illustrated in Fig. 6.

GðtÞ ¼0; t P tg ;

RE � ðt � tgÞ; t > tg :

�ð7Þ

The reception prediction function ðFðtÞÞ and transmis-sion function ðGðtÞÞ coincide with each other at time tc .This is the time that ONU queue is expected to becomeempty if enough bandwidth is available.

FðtcÞ ¼ GðtcÞ: ð8Þ

By substituting Fð�Þ and Gð�Þ from (6) and (7), respec-tively and solving the equation for tc , we have,

tc¼

tgþ QRE; tg6 ts� Q

RE;

RW ts�Q�REtg

RW�RE; ts� Q

RE< tg < te� 1

REðQþRW te�RW tsÞ;

tgþQþRW te�RW tsRE

; tg P te� 1REðQþRW te�RW t1Þ;

8>><>>:

ð9Þ

where ts and te denote the beginning and the end of thenext grant slot, respectively. The calculation should be donefor all ONUs. Then the request correction of each ONU is,

Rknew ¼ RE � ðtk

c � tkgÞ; ð10Þ

where Rknew is the new request for ONU k. tk

c and tkg denote tc

and tg for ONU k, respectively.The correction process is repeated for each ONU until

the granted bandwidths stall or a predetermined maxi-mum iteration is reached. Then the granted bandwidth issent to the ONUs along with the grant start times and pre-dicted incoming traffic.

Algorithm 1 (Proposed Inter-ONU DBA pseudo-code).

Require: QðiÞ; i ¼ 1; . . . ;Nfor i :¼ 1 to N do

Rði;1Þ QðiÞend forfor k :¼ 1 to Itemax do

Rtex 0

Btex 0

for i :¼ 1 to N doif Rði; kÞ > Wmax then

RexðiÞ Rði; kÞ �Wmax

Btex Bt

ex þ BexðiÞelse

Btex Bt

ex þWmax � ðRði; kÞÞend if

end forfor i :¼ 1 to N do

if Btex > Rt

ex thenGðiÞ Rði; kÞ þ Bt

ex�Rtex

N

elseif Rði; kÞ < Wmax then

GðiÞ Rði; kÞelse

GðiÞ Wmax þ RexðiÞRt

ex� Bt

ex

end ifend if

end forUpdate Tsch

Update Pði;G; Tsch; FiðtÞÞRði; kþ 1Þ Rði; kÞ þ Pði;G; TschÞ

end forUpdate Wmax

The pseudo-code of an inter-ONU DBA scheme withprediction is shown in Algorithm 1. The pseudo-code dem-onstrates how the prediction could be applied to a DBAalgorithm.

Note that the time frame in each ONU and OLT is differ-ent due to propagation delay. For the sake of simplicity,this effect is not considered in the pseudo-code, but ithas to be considered in simulations.

The notations used in the pseudo-code are explained inTable 1.

4.2. Intra-ONU scheduling

After the available bandwidth is divided between theONUs, the intra scheduler at ONU has to divide the granted

Table 1Notations used in algorithm.

Notation Description

N Number of ONUsRði; kÞ Request of ONU i in iteration kQðiÞ Queue length at the beginning of grantItemax Maximum iteration for solving the bandwidth

allocationRt

exTotal extra requested bandwidth

Btex

Total extra bandwidth available

GðiÞ Granted bandwidth to ONU iPði;G; TschÞ Prediction of incoming traffic for ONU i according to

granted time slots. Calculated according to Eq. 9

250 M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259

bandwidth between different traffic classes. Here, we con-sider three different classes of traffic, namely expeditedforwarding (EF), assured forwarding (AF) and best effort(BE). Here, we adopt strict priority in the ONU scheduler.Therefore, the ONU scheduler serves EF, the highest prior-ity class first. Then it serves AF and after that BE.

When strict priority is applied in the ONU scheduler,high priority packets that are recently received can con-sume the granted bandwidth that was originally requestedby the lower priority packets. It causes the delay of lowerpriority packets to decrease significantly in light loadONUs. The phenomenon is called light load penalty. A rem-edy for this problem is suggested in [8]. The suggestion isto move the contents of queues in order of their priorityto a separate transmit buffer at the time of sending the re-quest. This buffer will be transmitted at the next grantregardless of the higher priority packets that may arriveafter that. After the transmit buffer is emptied, otherqueues are served in order of their priority. In other words,the transmit buffer acts as the highest priority queue. Thisalgorithm is called two-stage DBA.

We propose a variant of two-stage DBA that does notrequire another buffer. At the time of sending the request,the ONU records the length of each queue. The recorded re-quest is then used to determine how many packets can besent from each queue during the next grant.

The ONU scheduler works in two rounds. In the firstround, it serves the queues in order of their priority up tothe saved request for each queue. Since the queues areFirst-In-First-Out (FIFO), it means that only the packetsthat were in the queue at the time of request ðtrÞ, areserved. After that, the scheduler starts transmitting thenext priority queue. The second round starts after allqueues are swept once and all packets that were in thequeues at request time are transmitted. In this cycle theONU scheduler performs normal strict priority schedulingto serve the newly arrived packets. The pseudo-code ofthe algorithm is shown in Algorithm 2.

When DBA is enhanced with the prediction, it basicallyassigns bandwidth to incoming packets. But as it was de-scribed in the previous section, the class of the incomingtraffic is not differentiated in order to keep the algorithmsimple. To be able to use the bandwidth that is assignedto incoming packets, the accounted predicted traffic is alsoreceived along with the grant size. This predicted trafficgranted bandwidth is represented with Gp. The scheduleradds Gp to the highest priority traffic request. Since the

incoming traffic at the highest priority is taken into ac-count, the remaining of Gp is used to transmit the newly re-ceived packets of the next priority queue. This continues tolower priority queues. After the first round, the secondround is performed without any modification.

The pseudo-code of the algorithm is shown in Algo-rithm 2. In the pseudo-code, RðiÞ denotes the length ofthe queue i at time of request, QðiÞ denotes the length ofqueue at the beginning of grant, G is the granted band-width and P is the predicted coming traffic. GcðiÞ representsthe granted bandwidth for class i;Nc is the number of clas-ses and Brem represents the remaining unused bandwidth.

Algorithm 2 (The intra-ONU scheduling algorithm pseudocode).

Require: RðiÞ; i ¼ 1; . . . ;Nc

Require: QðiÞ; i ¼ 1; . . . ;Nc

Require: G; PFirst round:Rð1Þ Rð1Þ þ Gp

for i :¼ 1 to Nc do

GcðiÞ min RðiÞ;G�Pi�1

k¼1GcðkÞ� �

end forSecond round:for i :¼ 1 to Nc do

Brem min 0;G�PNc

k¼1GcðkÞ� �

GcðiÞ GcðiÞ þminðQðiÞ � RðiÞ;BremÞend for

5. Maximum throughput analysis

The maximum throughput is achieved when the wholegranted bandwidth is used. Therefore in order to calculatethe maximum throughput, it is assumed that all ONUs arebacklogged. A direct result is consuming the whole grantedbandwidth. In each EPON cycle, which is denoted by C, thefactors that are explained in following paragraphs, wastethe bandwidth.

The required gap between ONU’s granted slots, knownas inter-ONU guard time ðG1Þ, is a contributing factor tobandwidth waste. In each cycle, there are Nonu gaps be-tween the grants; where Nonu is the number of active ONUsin the network. Another contributing factor is the gap be-tween two consecutive cycles, which we represent as G2.The bandwidth that is used for sending the report mes-sages also reduces user bandwidth. Again, there are Nonu

report messages in each EPON cycle.Adding up all of the waste factors, it is possible to work

out the maximum throughput. It is assumed that the DBAalgorithm that is used and the round trip time of ONUs, letthe grant slots to be assigned without any gap in betweenother than the described ones. The maximum throughputis,

Rmax ¼ 1�Nonu � G1 � G2 � Nonu � Lrep

RE

C; ð11Þ

where Rmax is the maximum throughput and Lrep is thelength of the report message in bits.

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 251

Some DBA algorithms such as the one presented in [3],require all ONU requests to process the bandwidth assign-ment. In these algorithms, the walk time between frameswastes a considerable amount of bandwidth. The walktime is basically composed of three different components;waiting time for reception of the last ONU request, the pro-cessing time of scheduling algorithm and the round triptime of first ONU. The last factor comes from the fact thatthe link remains idle for the duration of transmitting thefirst ONU’s grant and receiving data from it. Supposingthe processing time is negligible, it is equal to the firstscheduled ONU’s round trip time. In summary, walk timeis defined as the time between reception of last reportmessage and reception of the first data bit of the firstONU. In this period, the link is idle and the correspondingbandwidth is wasted. Taking the walk time into accountthe maximum throughput is worked out as,

Rmax ¼ 1� 1C� Nonu � G1 � G2 � Nonu �

Lrep

RE� Trtt � Tproc

� �;

ð12Þ

where Trtt is the round trip time of the ONU that its reportis the last report and Tproc is the processing time of DBAalgorithm.

6. Traffic-prediction-assisted algorithms

The proposed traffic prediction mechanism can be ap-plied to any type of DBA algorithm. Since it provides theDBA algorithms with real-time information of traffic, itgenerally improves the delay performance. There are twoapproaches among DBA algorithms to EPON cycle lengthwhich is a significant factor in the performance of the pro-posed method. Some DBA algorithms, such as the well-known Interleaved Polling with Adaptive Cycle Timing(IPACT) [7], use dynamic cycle length while others, suchas Excess Distribution [3], adapt fixed cycle length. In orderto investigate the proposed method’s performance in bothcategories, the proposed method is applied to a DBA algo-rithm from each category.

The first DBA algorithm is Excess Distribution which isfirst proposed in [3] and later improved in [4]. This DBAscheme first divides the available bandwidth equallyamong the ONUs to define the threshold window. All ofthe ONU’s requests that are less than the threshold aregranted. Then, the unused remaining bandwidth is distrib-uted equally between the unsatisfied ONUs. The variantused in this study adapts a two-stage scheduling algo-rithm, as described in Section 4.2, to allocate bandwidthto each of the traffic classes. This algorithm is referred toas Excess Distribution or ED, while the traffic-prediction-assisted version of it, is referred to as Integrated ED. Thepsuedo-code of the integrated ED is shown in Algorithm1 as an example of applying the proposed traffic predictionon a DBA algorithm.

Unlike the first algorithm that maintains the fixed EPONcycle length, the second algorithm, namely CB-IPACT, ad-justs EPON cycle length according to the introduced trafficload. Another difference is that CB-IPACT grants the re-quests on-the-fly, i.e., it processes each request indepen-

dent of others. CB-IPACT grants the request if it is lessthan a predefined maximum grant. Otherwise, it grantsthe maximum grant.

Grant ¼ minfRequest; MaxGrantg: ð13Þ

Then the two-stage approach, explained in Section 4.2, isemployed to distribute the granted bandwidth to differenttraffic classes. Employing the two-stage approach helps sat-isfy the QoS requirement for each of the classes. Applyingtraffic prediction to this algorithm by the method, ex-plained in Section 4.1, is fairly simple. The traffic-predic-tion-assisted CB-IPACT is referred to as Integrated-CBIPACT.

7. Simulation platform

The performance of the proposed algorithm is exten-sively studied using simulations setup under different sce-narios. Simulation experiments are conducted with theOPnet simulation package for wireless internetworkingV.15 [1]. Custom models for OLT, ONU-BS as well as linkare developed to ensure utmost similarity to the realequipment.

WiMAX standard supports several physical layers. Thisincludes Orthogonal Frequency-Division Multiple Access(OFDMA), Orthogonal Frequency-Division Multiplexing(OFDM) and Single-Carrier Time Division Multiple Access(SC-TDMA); each supports Time Division Duplexing(TDD) or Frequency Division Duplexing (FDD). In thisstudy, without lack of generality, we consider SC-TDMAwith TDD.

The simulation model is depicted in Fig. 7. Unless men-tioned otherwise, the scenario given in Fig. 7 is used as thesimulation topology. This simulation model consists of oneOLT which is connected to 32 ONU-BSs. The performanceof the system is analyzed by simulating it under differentDBA algorithms.

The proposed traffic prediction assisted bandwidthassignment method is applied to CB-IPACT and Excess Dis-tribution. The performance of the enhanced algorithms aswell as the original algorithms is studied extensivelythrough simulations.

The Ethernet and WiMAX workstations are responsiblefor generating different classes of traffic. Since we focus onthe performance of uplink, all of the traffic load is destinedtoward a server located near the OLT. The default simulationparameters are shown in Table 2, unless otherwise stated.

To simulate the WiMAX part of the network, a modifiedversion of standard OPnet WiMAX models is used. Theadded modifications enable these models to share someof their internal information with the developed EPONmodels, which is vital to implement the proposed DBAalgorithms for hybrid optical-wireless networks.

The EPON part of the network is custom designed toimplement the point to multipoint aspects of the EPONas well as its Ethernet frame architecture. The EPON mod-els consist of OLT MAC layer, ONU MAC layer and link.These models are accessible from [1].

Optical networks typically performs as the backboneinfrastructure. In real implementations, they connect cor-porate networks as well as wireless base stations to the

Fig. 7. Simulation scenario [replaced by a more meaningful diagram].

Table 2Simulation parameters.

Simulationparameters

Value

EPON cycle length 2 msWiMAX phy TDD

SCWiMAX frame size 5 msInter ONU guard time 1 lsPropagation delay 10 ls

Table 3Traffic loads in realistic scenario [added].

Low load (Mbps) High load (Mbps)

ONU 12.5 150ONUBS 4.5 50

252 M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259

operator’s core. This realistic scenario has been studiedSection 8.4.

In the realistic scenario, half of the ONUBSs have beenreplaced by the legacy ONUs, that serve the corporateusers, to study their effect on the performance of the pro-posed algorithm. The legacy ONUs cannot predict the traf-fic and has to request for bandwidth.

In addition, in a real network the real-time load on theusers are not equal. To model this effect, in the scenario,80% of the total traffic load is generated in 25% high loadONUs and ONUBSs. The remaining load is then generatedby low load ONUs and ONUBSs. The traffic load on theONUs and ONUBSs are shown in Table 3.

7.1. Traffic types

Three traffic types are considered in the simulationexperiments. The highest priority traffic is the ExpeditedForwarding (EF) which models the voice or video traffic.The Quality of Service requirements of this class are de-fined by maximum tolerable delay, maximum jitter, maxi-mum loss and throughput.

The second priority traffic is the Assured Forwarding(AF). Multimedia streaming traffic falls into this category.This class of traffic is delay-constrained and typically itsdelivery is assured as long as its throughput remains with-in predefined limits.

The third, and most basic traffic class is the Best Effort(BE) traffic that models file transfers, email and web traffic.This class has no specific requirement in terms of maxi-mum delay or jitter. However, the throughput plays animportant role for this class and hence, some researchersdefine a minimum throughput requirement to prohibitBE traffic from being cut-off from the network in highlyloaded conditions.

7.2. Traffic generation

WiMAX users (Fig. 7) are responsible for generatingtraffic with the required distribution. The EF traffic loadis generated by a Poisson process while AF and BE trafficclasses are generated by self-similar models to better mod-el real traffic in the network. To model the traffic morerealistically, EF traffic is composed of packets with fixedlength equal to 100 bytes, while AF and BE packet lengthis uniformly distributed between 100 and 1500 bytes.The total traffic load is composed of BE, AF and EF; each

Fig. 8. Normalized throughput versus normalized load.

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 253

of them shares 0.4, 0.4 and 0.2 of the total traffic, respec-tively [25].

A simple way to generate self-similar traffic is to multi-plex several traffic flows with a distribution that has long-range dependence (LRD) property, like Pareto distribution[33]. The density function of Pareto distribution is,

f ðxÞ ¼ ak

kx

� �aþ1

; x > k; a > 0: ð14Þ

Since a traffic with heavy-tailed distribution is desired, theshape parameter is bounded ð1 < a < 2Þ. The degree ofself-similarity which is normally expressed as Hurstparameter (H) is related to shape parameter by,

H ¼ 3� a2

; 1 < a < 2: ð15Þ

The model used in simulation experiments, is a super-position of Fractal Point Processes (Sup-FRP). The Hurstparameter is set to 0.8 which is suitable for web traffic aswell as file transfer [15]. The fractal onset time scale isset to 0.1 s.

The simulation scenario consists of three users per eachONU-BS. Each user is responsible of generating one class oftraffic. This gives us the flexibility of modifying each trafficclass and its corresponding parameters without affectingother classes of traffic.

Fig. 9. Delay of EF flow in different loads.

8. Simulation results

In this section, the performance of the proposed algo-rithm is investigated under different conditions. Unlessstated otherwise, the cycle length and the frame lengthare equal to 2 ms and 5 ms, respectively. Moreover, theload is sufficient enough to saturate the network.

The remainder of this section compares the performanceof the CB-IPACT, Excess Distribution, Integrated-CBIPACTand Integrated ED, under different load conditions. For thesake of comparison, some of the performance criteria arealso measured for static scheduling. The effect of modifyingkey parameters is also investigated.

8.1. Load analysis

In the following paragraphs, the behavior of the pro-posed algorithms under different loads is investigated. Un-der light and moderate loads, the network throughput,shown in Fig. 8, is the same for all algorithms. This is anobvious result of introducing less load than network capac-ity which all of the DBA algorithms are capable of deliver-ing. The throughput performance curves separate whenthe introduced load approaches the link capacity.

Fig. 9 compares the delay of the EF flow for differentalgorithms as the load increases. As expected, an increasein load results in more delay. Under lower loads, the grad-ual increase in the delay of CB-IPACT is mainly because ofthe increase in cycle length. The cycle length of CB-IPACTextends with load which in turn results in 1 ms increasein average delay. The probability distribution function ofdelay of EF flows (Fig. 11) clearly shows that increasingthe load results in longer delays. In full load conditions, cy-

cle length saturates to its maximum and the EF delay is in-creased by more than the maximum cycle length. The samediscussion applies for AF flows with the minor differencethat AF flows experience a faster increment due to theirlower priority.

Since incoming traffic prediction assumes that the traf-fic is continuous, the scheduler is not aware of the band-width requirement of each individual packet. Therefore,the scheduler may assign a portion of the required band-width to send the packet. Since fragmentation of packetsis not permitted in EPON, this granted bandwidth wouldbe wasted. The inaccuracy caused by partial bandwidthassignment slightly deteriorates the performance of theintegrated algorithms and causes the delay to increaseslightly. However, the reward of bearing a slightly longerdelay in light loads is a significant delay decrease in highload, as shown in Figs. 9 and 10.

Both Excess Distribution and integrated ED algorithmsfix a cycle length, equal to maximum cycle length of CB-IPACT and integrated-CBIPACT. As a result, their cyclelength is longer than the average cycle length of CB-IPACT

Fig. 10. Delay of AF flow in different loads.

Fig. 11. Probability density function of EF delay with CB-IPACT andintegrated-CBIPACT.

Fig. 12. Probability density function of EF delay with ED and integratedED.

254 M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259

and integrated-CBIPACT in light and moderate loads.Therefore, their delay is generally greater than that ofCB-IPACT and integrated-CBIPACT. ED can keep constantdelay when load is up to 75% of link capacity for EF trafficand up to 50% of link capacity for AF traffic. This is becausehigher priority traffic can take advantage of the excessbandwidth which is distributed between ONUs, to sendtheir packets. Again, Since AF has lower priority, it is moresusceptible to increasing load and its effect can be seensooner. The delay of both flows increases significantly atfull load. However, if the proposed traffic prediction algo-rithm is applied, it counteracts the load effect and signifi-cantly decreases the delay. The effect of prediction can bebest demonstrated with the probability distributed func-tion of the delay that is shown in Fig. 12.

Fig. 9 shows an interesting phenomenon. The averagedelay of the EF flow decreases as load increases from 0.8to 1. In moderate loads, there is an amount of excess band-

width that is distributed among the ONUs. This excessbandwidth varies significantly for each cycle and as a re-sult, the granted bandwidth also significantly varies.Therefore, the position of the granted slot in the cyclechanges. This might increase the cycle length that is mea-sured in some ONUs and as a result, increase their averagedelay. Not all of the ONUs are affected by the phenomenon.Since both ED and CB-IPACT serve ONUs in a round robinfashion, the variation in grants does not affect first and lastONUs. However, the full effect of variation can be seen atthe ONUs that are granted in the middle of the cycle. Whenthe load increases to full load, less excess bandwidth re-mains and the granted slots become equal. Therefore, thegranted slot does not change its position and the phenom-enon vanishes, which is shown in Fig. 9 as a slight decreasein average EF delay.

From QoS point of view, the throughput is almost thesame for all four algorithms. The delay bound of the algo-rithms is the same in light lights. But in heavy loads, inte-grated algorithms, due to their unique prediction method,allow considerably lower delay bound, at the expense of aninsignificant delay increase in light loads. A comparison offixed and adaptive cycle length algorithms reveals thatalthough both categories have the same delay bound, theadaptive cycle length algorithms perform better in termsof delay, in light load conditions. This holds true for inte-grated algorithms.

8.2. EPON cycle length analysis

Since the ability to predict incoming traffic depends onthe ratio between the lengths of the EPON cycle and Wi-MAX frame, as discussed in Section 3 and 4, changing thecycle length has a significant impact on the performanceof the proposed algorithm. In saturated conditions, whereall of the ONUs are backlogged, almost all DBAs assignthe available bandwidth equally to the ONUs. Therefore,the difference in performance is mainly due to trafficprediction. By analyzing the full load behavior of the

Fig. 14. Normalized analytic throughput versus cycle length.

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 255

algorithm, it is possible to investigate the prediction per-formance in nearly isolated environment.

Fig. 13 compares the network throughput in differentcycle lengths. The frame length is set to 5 ms in all of thesimulations. The throughput is normalized to nominalcapacity of the link which is 1 Gbps in EPON. The normal-ized throughput varies between 0.92 and 0.99. CB-IPACTassigns the bandwidth on the fly, i.e., as the scheduler re-ceives each bandwidth request, it processes the requestand sends the grant. On the other hand, ED processes all re-quests at the same time. Hence, walk time causes a reduc-tion in ED’s maximum throughput in comparison with CB-IPACT. Static bandwidth assignment does not perform theassigning based on requests and its maximum throughputis the same as CB-IPACT.

Maximum throughput increases with longer cyclelengths. This is mainly because there is less frequent re-quest/grant exchanges which directly translates in reducedwalk time. Therefore, the bandwidth consumed for the re-quest/grant pairs and also the bandwidth wasted in walktimes can be conserved.

Fig. 13 also indicates an insignificant throughput reduc-tion in integrated versions of CB-IPACT and ED in compar-ison with the original algorithms. This is due to theoverhead caused by extra message exchange to share thetraffic information with the OLT in integrated versions.

The simulated throughput as well as the analyticalthroughput are shown in Fig. 14. The minor difference be-tween simulation and analytic throughput is caused by thefragmentation effect, which we did not consider in analy-sis. Since fragmentation is not allowed in EPON, the as-signed bandwidth is wasted if the packet is larger thanthe remaining granted bandwidth. This slightly decreasesthe throughput in simulation experiments. For shorterEPON cycles, this can happen more frequently and thusthe effect is larger.

Fig. 15 shows the average delay of the EF stream with dif-ferent cycle length. Generally, by increasing the cyclelength, the average delay is also increased because thereare less frequent grant slots available to the ONU. A packetmay remain in queue longer in longer cycles. This justifiesthe linear increase in the delay shown in Fig. 15. By knowing

Fig. 13. Normalized throughput versus cycle length.

the incoming traffic in advance, the proposed algorithmscan save the time that is normally spent on exchanging re-quest/grant. This results in almost a cycle length reductionin delay when full knowledge of incoming traffic exists.For example, assuming the frame length is 5 ms, the ONUhas full knowledge of incoming traffic when the cycle lengthis equal to 1 ms or 2 ms (refer to Section 3 for details). It isbacked by 1 ms and 2 ms delay reduction when cycle lengthis set to 1 ms and 2 ms, respectively.

By contrast, given that the WiMAX frame length is 5 ms,in 5 ms and 10 ms EPON cycle lengths, ONU has only partialknowledge on incoming traffic and hence a part of incom-ing traffic is handled in conventional manner, i.e., request/grant mechanism. Therefore it results in only a partialreduction in delay. This phenomenon can be best under-stood through careful investigation of PDF of the EF delay.The EF delay probability density function of ED and Inte-grated ED is shown in Fig. 16. When EPON cycle length isset to 2 ms, adding prediction to the algorithm clearly shiftsthe PDF leftward and decreases the delay. However, whenEPON cycle length is 5 ms, it is not always possible to sendthe required information to the prediction module in time

Fig. 15. EF stream delay versus EPON cycle length.

Fig. 18. AF stream delay versus EPON cycle length.

256 M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259

and information may be received when it is expired. This isdemonstrated in Fig. 17. The required information for theprediction of incoming traffic which is extracted from theDL/UL map is transmitted to the OLT scheduler along withthe request. In case of BS1 the information is updated aftertransmission of request and therefore, the incoming packethas to wait for the next request to be sent if there is no extragranted bandwidth available for sending it. On the otherhand, BS2 updates the information before sending requestand hence the incoming packet is predicted in the sched-uler and the required bandwidth is assigned for transmit-ting it. The result is the PDF shown in Fig. 16.

The same effect can be seen in Fig. 18 for the AFstreams. However, because the prediction module doesnot differentiate between traffic types, the bandwidth as-signed to the predicted traffic is assigned to the highestpriority flow, which is EF flow. This does not affect the per-formance when ONU has full knowledge of the comingtraffic. But in case of partial knowledge, i.e., EPON cyclelength is 5 ms or 10 ms, the unpredicted EF packets cantake advantage of the bandwidth, originally assigned toAF packets. This deteriorates the performance of lower pri-ority packets in case of partial knowledge.

In summary, throughput is not affected by augmentingthe algorithms with the prediction mechanism. Howeverthe delay of high priority classes can be decreased by a fac-

Fig. 16. Comparison of EF delay PDF for different EPON cycle length.

Fig. 17. Timing diagram for partial k

tor of two. Although the method always decreases the de-lay, it performs best with full knowledge, i.e., when theEPON cycle length is less than WiMAX frame length.

8.3. WiMAX frame length analysis

Changing the length of the WiMAX frame affects theperformance metrics of the proposed algorithms. The onlyperformance metric which is not affected is systemthroughput which is shown in Fig. 19. The EPON cyclelength is fixed to 2 ms in all simulation experiments. Likethe previous discussion, the ED’s throughput is slightly lessthan CB-IPACT due to walk time.

The delay of EF and AF streams for different frame lengthsis shown in Figs. 20 and 21, respectively. Although, changingthe frame length does not affect the amount of traffic load, itchanges the frequency and shape of the incoming traffic; i.e.,when the frame length is extended, the frequency of incom-ing traffic decreases, however the amount of traffic deliv-ered instantly to optical network increases.

The delay of the EF stream remains fixed for WiMAXframe lengths up to 5 ms and then it slightly increases,while the delay of AF streams is only fixed for WiMAXframe lengths up to 2 ms and then it increases dramaticallyafterward. Because WiMAX TDD is used as the physicallayer of the wireless part of the network, the incoming traf-fic is bursty in nature. This does not affect the delay, if

nowledge of incoming traffic.

Fig. 19. Maximum throughput versus frame length.Fig. 21. AF stream delay versus frame length.

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 257

WiMAX frame length and EPON cycle length are set in away that EPON sees continuous traffic, i.e., the amount oftraffic introduced to EPON is almost the same in each EPONcycle. But as WiMAX frame length grows to more than thatof the EPON cycle, the traffic becomes bursty and pulse-shaped. The amount of the traffic in each pulse also in-creases. In long frames, the amount of traffic delivered tothe ONU is more than the maximum grant size. Thus it isdivided among several EPON cycles, resulting in increasedqueuing delay. Since, the prediction algorithm does notprovide the information about the amount of traffic in eachclass, the biggest share of the predicted bandwidth is allo-cated to the EF stream and hence it exhibits the increasingtrend later than AF stream does. The phenomenon can beseen clearly in Figs. 20 and 21.

Extending the frame length leads to extending therange of the prediction, and in turn, decreases the meanEF delay. However, the decreasing trend stops at framelength equal to 5 ms and increases thereafter. As the Wi-MAX frame is extended, the amount of traffic aggregatedin each frame grows. The aggregated traffic is delivered al-most instantaneously to the ONU for transmission. In the

Fig. 20. EF stream delay versus frame length.

short frames, the amount of traffic is little enough to besent in a single EPON cycle. The lack of distinction betweenthe traffic classes in prediction also acts in favor of highpriority traffic. Since EF traffic can also use the grantedbandwidth that is originally assigned to AF flows. But whenthe aggregated EF traffic is increased to more than thegranted bandwidth, it spreads over two or more EPON cy-cles, which results in increased delay.

In summary, increasing the WiMAX frame length incomparison to EPON cycle length helps the prediction algo-rithm predict traffic for a longer range. On the other handaggregating traffic, caused by longer WiMAX frames, canresult in the loss of statistical multiplexing gain. Long Wi-MAX frame lengths also increases the queuing delay whichcontracts the delay reduction gained by traffic prediction.Therefore, it is vital to intelligently choose the EPON cyclelength as well as the WiMAX frame length to take advan-tage of the prediction.

8.4. A realistic scenario

The delay of EF and AF traffic classes are shown inFig. 22. It shows the delay of high and low load ONU and

Fig. 22. The delay of EF and AF traffic classes in realistic scenario [added].

Fig. 23. ONU/ONUBS throughput in realistic scenario [added].

258 M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259

ONUBSs when various DBA algorithm are employed. As canbe seen in the figure, applying the proposed prediction-based DBA algorithm does not affect the delay of legacyONUs. However, it lowers the delay of ONUBSs’ delay-sen-sitive packets dramatically. This is due the fact that incases where the network as a whole is working with fullcapacity, the total requested bandwidth is greater thanthe total available bandwidth. Thus, the ONUs and ONUBSsare granted with the requested bandwidth or less whichresults in an extra delay for requesting bandwidth. Sincethe proposed scheme pre-grant the packets from ONUBSs,the algorithm benefits both of the high and low loadONUBSs. In summary, the proposed DBA algorithm is ben-eficial when the network is working in full capacity regard-less of the traffic distribution.

Fig. 23 shows the throughput of ONU/ONUBSs. In thisfigure, High load ONU and ONUBS are represented by ONU_Hand ONUBS_H, whereas ONU_L and ONUBS_L represent lowload ONUs and ONUBSs. It is concluded from the figure thatthe delay decrement in delay-sensitive classes of trafficcomes with a slight decrease of BE throughput in high loadONUs. On the other hand, the high load ONUBSs are grantedmore bandwidth to enable them to better serve the delay-sensitive classes. It is worth mentioning that the through-put of low load ONUs and ONUBSs are not affected. Also,applying the proposed algorithm does not harm the EFand AF throughput of high load ONUs.

9. Conclusion

Hybrid optical-wireless network is a promising technol-ogy that is relatively inexpensive and can address the hugebandwidth need of modern applications. In this work, a

mechanism was proposed to enhance the performance ofthe optical scheduler algorithm by augmenting it with a no-vel incoming traffic prediction, which is extracted from theinternal information of the wireless scheduler. The predic-tion method was described and investigated in detail. Basedon the prediction method, two DBA algorithms were pro-posed. Extensive simulation experiments proved the per-formance of the algorithms and showed that due to theprediction method applied, the delay of both algorithmsin high loads can be decreased by a factor of two, withoutaffecting the throughput. The effects of changing the mainparameters were also investigated. In addition to that, theperformance of the proposed algorithm was proved in a realworld scenario which consists of ONUs and ONUBSs.

The results showed that predicting the traffic removesthe significant delay increase which normally occurs inhigh load conditions. In fact, the delay of high priority traf-fic remains the same for high and light load conditions.This enables the service providers to establish a lower de-lay bound and a better Quality of Service for consumers,even in high load conditions.

Since the proposed algorithm is superior to conven-tional algorithms, only in the high load conditions, and alsobecause the main part of the algorithms runs in the OLT, itis very simple to further tailor the algorithm to be acti-vated just in high loads, where it is useful the most, andto remain idle otherwise. This results in a simpler DBA inlight load conditions and a more efficient one in high loadconditions. In future works, we will establish a proper loadthreshold to activate the proposed DBA algorithm.

References

[1] Opnet simulation software. [Online]. <www.opnet.com>.

M. Mirahmadi, A. Shami / Computer Networks 56 (2012) 244–259 259

[2] F. Alsolami, P. Ho, A novel two-level hierarchical scheduling schemefor integrated WiMAX and EPON, in: ICS, 2009.

[3] C. Assi, Y. Ye, S. Dixit, M. Ali, Dynamic bandwidth allocation forquality-of-service over ethernet pons, IEEE Journal on Selected Areasin Communications 21 (9) (2003) 1467–1477.

[4] A. Dhaini, C. Assi, A. Shami, Dynamic bandwidth allocation schemesin hybrid tdm/wdm passive optical networks, in: 3rd IEEE ConsumerCommunications and Networking Conference, 2006, CCNC 2006, vol.1, 2006, pp. 30–34.

[5] I.-S. Hwang, J.-Y. Lee, C.-W. Huang, Z.-D. Shyu, Advanced dynamicbandwidth allocation and scheduling scheme for the integratedarchitecture of epon and wimax, IEEE International Conference onMobile Data Management 0 (2009) 655–660.

[6] K. Katrinis, A. Tzanakaki, S. Dweikat, S. Vassilaras, R. Nejabati, D.Simeonidou, G. Zervas, Backhauling wireless broadband traffic overan optical aggregation network: Wimax over obs, in: SixthInternational Conference on Broadband Communications,Networks, and Systems, 2009, BROADNETS 2009, 2009, pp. 1–4.

[7] G. Kramer, B. Mukherjee, G. Pesavento, Ipact: a dynamic protocol foran ethernet pon (epon), IEEE Communications Magazine 40 (2)(2002) 74–80.

[8] G. Kramer, B. Mukherjee, S. Dixit, Y. Ye, R. Hirth, Supportingdifferentiated classes of service in ethernet passive opticalnetworks, The Journal of Optical Networking 1 (8) (2002) 280–298.

[9] Y. Luo, S. Yin, T. Wang, Y. Suemura, S. Nakamura, N. Ansari, M.Cvijetic, QoS-aware scheduling over hybrid optical wirelessnetworks, in: National Fiber Optic Engineers Conference, 2007.

[10] Y. Luo, T. Wang, S. Weinstein, M. Cvijetic, S. Nakamura, Integratingoptical and wireless services in the access network, in: Optical FiberCommunication Conference and Exposition and The National FiberOptic Engineers Conference, Optical Society of America, 2006. PaperNThG1.

[11] B.O. Obele, M. Iftikhar, S. Manipornsut, M. Kang, Analysis of thebehavior of self-similar traffic in a qos-aware architecture forintegrating wimax and gepon, Journal of Optical Communicationsand Networking 1 (4) (2009) 259–273.

[12] B.O. Obele, M. Kang, Fixed mobile convergence: a self-aware qosarchitecture for converging wimax and gepon access networks,International Conference on Next Generation Mobile Applications,Services and Technologies 0 (2008) 411–418.

[13] F. Ohrtman, WiMAX Handbook: Building 802.16 Wireless Networks,ser. McGraw-Hill Communications, McGraw-Hill, 2005.

[14] S. Ou, K. Yang, M. Farrera, C. Okonkwo, K. Guild, A control bridge toautomate the convergence of passive optical networks and ieee802.16 (wimax) wireless networks, in: Broadband Communications,Networks and Systems, 2008, BROADNETS 2008, September 2008,pp. 514–521.

[15] J. Potemans, B.V. den Broeck, Y. Guan, J. Theunis, E.V. Lil, A.V. deCapelle, Implementation of an advanced traffic model in opnetmodeler, in: OPNETWORK 2003, Washington DC, USA, 2003.

[16] A. Reaz, V. Ramamurthi, S. Sarkar, D. Ghosal, S. Dixit, B. Mukherjee,CaDAR: an efficient routing algorithm for a wireless-opticalbroadband access network (WOBAN), IEEE/OSA Journal of OpticalCommunications and Networking 1 (5) (2009) 392–403.

[17] S. Sarkar, S. Dixit, B. Mukherjee, Hybrid wireless-optical broadband-access network (woban): a review of relevant challenges, Journal ofLightwave Technology 25 (11) (2007) 3329–3340.

[18] S. Sarkar, Y. Hong-Hsu, S. Dixit, B. Mukherjee, Hybrid wireless-optical broadband access network (woban): network planning usingLagrangean relaxation, IEEE/ACM Transactions on Networking 17 (4)(2009) 1094–1105.

[19] S. Sarkar, B. Mukherjee, S. Dixit, Optimum placement of multipleoptical network units (onus) in optical-wireless hybrid accessnetworks, in: Optical Fiber Communication Conference, 2006 andthe 2006 National Fiber Optic Engineers Conference, OFC 2006,March 2006, p. 3.

[20] S. Sarkar, H.-H. Yen, S. Dixit, B. Mukherjee, Dara: delay-awarerouting algorithm in a hybrid wireless-optical broadband accessnetwork (woban), in: IEEE International Conference onCommunications, 2007, ICC’07, June 2007, pp. 2480–2484.

[21] S. Sarkar, H.-H. Yen, S. Dixit, B. Mukherjee, A mixed integerprogramming model for optimum placement of base stations andoptical network units in a hybrid wireless-optical broadband accessnetwork (woban), in: Wireless Communications and NetworkingConference, March 2007, WCNC 2007, IEEE, 2007, pp. 3907–3911.

[22] S. Sarkar, H.-H. Yen, S. Dixit, B. Mukherjee, Hybrid wireless-opticalbroadband access network (woban): network planning and setup, IEEEJournal on Selected Areas in Communications 26 (6) (2008) 12–21.

[23] S. Sarkar, H.-H. Yen, S. Dixit, B. Mukherjee, A novel delay-awarerouting algorithm (dara) for a hybrid wireless-optical broadbandaccess network (woban), IEEE Network 22 (3) (2008) 20–28.

[24] S. Sarkar, H.-H. Yen, S. Dixit, B. Mukherjee, Radar: Risk-and-delayaware routing algorithm in a hybrid wireless-optical broadbandaccess network (woban), in: Conference on Optical FiberCommunication and the National Fiber Optic EngineersConference, 2007, OFC/NFOEC 2007, March 2007, pp. 1–3.

[25] A. Shami, X. Bai, N. Ghani, C. Assi, H. Mouftah, Qos control schemesfor two-stage ethernet passive optical access networks, IEEE Journalon Selected Areas in Communications 23 (8) (2005) 1467–1478.

[26] W.-T. Shaw, S.-W. Wong, N. Cheng, L. Kazovsky, Marin hybridoptical-wireless access network, in: Conference on Optical FiberCommunication and the National Fiber Optic Engineers Conference,2007, OFC/NFOEC 2007, March 2007, pp. 1–3.

[27] G. Shen, R. Tucker, C.-J. Chae, Fixed mobile convergencearchitectures for broadband access: integration of epon and wimax[topics in optical communications], IEEE Communications Magazine45 (8) (2007) 44–50.

[28] C. So-In, R. Jain, A.-K. Tamimi, Scheduling in ieee 802.16e mobilewimax networks: key issues and a survey, IEEE Journal on SelectedAreas in Communications 27 (2) (2009) 156–171.

[29] S. Wang, H. Li, Y. Ji, Distribution qos scheme for a novel of hybridoptical wireless network, in: J. Wang, G.-K. Chang, Y. Itaya, H. Zech(Eds.), Proc. SPIE, vol. 6784, no. 1, SPIE, 2007, p. 67842N.

[30] Y. Yan, H. Yu, L. Dittmann, Wireless channel condition awarescheduling algorithm for hybrid optical/wireless networks, in:Access Nets: Third International Conference on Access Networks,AccessNets 2008, October 15–17, 2008, Springer, Las Vegas, NV, USA,2009, p. 397. Revised papers.

[31] Y. Yan, H. Yu, H. Wang, L. Dittmann, Integration of epon and wimaxnetworks: uplink scheduler design, in: SPIE, W. Hu, S.-K. Liu, K. ichiSato, L. Wosinska (Eds.), vol. 7137, no. 1, 2008, p. 713704.

[32] K. Yang, S. Ou, K. Guild, H.-H. Chen, Convergence of ethernet pon andieee 802.16 broadband access networks and its qos-aware dynamicbandwidth allocation scheme, IEEE Journal on Selected Areas inCommunications 27 (2) (2009) 101–116.

[33] L. Ying, K. Feng-ju, Z. Lian-jiong, W. Zhi-Guang, A self-similar trafficgeneration method and application in the simulation of mobile ad-hoc network, in: International Colloquium on Computing,Communication, Control, and Management, 2009, CCCM 2009,ISECS, vol. 4, 8–9 2009, pp. 229–233.

Maysam Mirahmadi received the B.Sc. andM.Sc. Degree in Electrical Engineering fromAmirkabir University of Technology (TehranPolytechnic), Tehran, Iran, in 2004 and 2007,respectively. From 2007 to 2009, he was aSenior Designer at Kavoshkam R&D Group,Tehran, Iran. Currently he is working towardhis Ph.D. degree in Electrical Engineering inthe Department of Electrical and ComputerEngineering, University of Western Ontario,London, ON, Canada. His current researchinterests include wireless broadband access

networks, quality of service in heterogeneous networks and wireless/optical integrated networks.

Abdallah Shami received the B.E. degree inElectrical and Computer Engineering from theLebanese University, Beirut, Lebanon in 1997,and the Ph.D. Degree in Electrical Engineeringfrom the Graduate School and UniversityCenter, City University of New York, NewYork, NY in September 2002. In September2002, he joined the Department of ElectricalEngineering at Lakehead University, ThunderBay, ON, Canada as an Assistant Professor.Since July 2004, he has been with The Uni-versity of Western Ontario, London, ON, Can-

ada where he is currently an Associate Professor in the Department ofElectrical and Computer Engineering. His current research interests are inthe area of wireless/optical networking.