ieee transactions on mobile computing 1 self … ieee papers/ns2/ns2... · self-organizing resource...

1536-1233 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI10.1109/TMC.2014.2338302, IEEE Transactions on Mobile Computing

IEEE TRANSACTIONS ON MOBILE COMPUTING 1

Self-Organizing Resource ManagementFramework in OFDMA Femtocells

Jongwon Yoon, Mustafa Y. Arslan, Member, IEEE, Karthikeyan Sundaresan, Senior Member, IEEE,Srikanth V. Krishnamurthy, Fellow, IEEE, and Suman Banerjee, Member, IEEE

AbstractNext generation wireless networks (i.e., WiMAX, LTE) provide higher bandwidth and spectrum efficiency leveragingsmaller (femto) cells with orthogonal frequency division multiple access (OFDMA). The uncoordinated, dense deployments offemtocells however, pose several unique challenges relating to interference and resource management in OFDMA femtocell networks.Towards addressing these challenges, we propose RADION, a distributed resource management framework that effectively managesinterference across femtocells. RADIONs core building blocks enable femtocells to opportunistically determine the available resourcesin a completely distributed and efficient manner. Further, RADIONs modular nature paves the way for different resource managementsolutions to be incorporated in the framework. We implement RADION on a real WiMAX femtocell testbed deployed in a typical indoorsetting. Two distributed solutions are enabled through RADION and their performance is studied to highlight their quick self-organizationinto efficient resource allocations.

Index TermsDistributed, femtocells, OFDMA, resource management, WiMAX

F

1 INTRODUCTION

RECENT industry reports predict that mobile data usagewill grow by a factor of 18 by 2016 with around 40%of the mobile data will originate indoors (i.e., homes andenterprises) [1]. The consequent demand for higher capacityhas led to smaller cells, called femtocells, that operate usingthe same OFDMA-based technology as the macro cells [2],[3]. Femtocells offer improved coverage and a cheap andenergy-efficient deployment option for indoor usage [4], [5],[6]. They provide numerous advantages to both mobile broad-band customers and service providers. For example, customerequipment (e.g., 4G-capable smart phone, laptop) can saveenergy on the uplink (while enjoying high throughput), sinceit does not have to transmit to a distant macro base station. Thedownlink throughput is also increased due to shorter ranges[7]. For service providers, femtocells offer (i) a cost-effectiveway of providing coverage, (ii) increased system capacityvia spatial reuse and (iii) reduced operational expenses andsubscriber churn. Since users can seamlessly migrate betweenfemtocells and macrocells on a licensed spectrum, femtocellsoffer an attractive alternative to the unlicensed and often con-gested WiFi systems. Recent surveys indicate that consumersfind femtocells appealing and consider installing one in theirhomes [8].

To preserve compatibility with the macrocell network, cer-tain mobility related signaling of femtocells (e.g., paging,

J. Yoon and S. Banerjee are with the University of Wisconsin-Madison,USA. E-mail: {yoonj, suman}@cs.wisc.edu

M. Y. Arslan and K. Sundaresan are with the NEC Labs America, NewJersey, USA. E-mail: {marslan, karthiks}@nec-labs.com

S. V. Krishnamurthy is with the University of California Riverside, USA.E-mail: [email protected]

Manuscript received 7 Jun. 2013; revised 4 Feb. 2014; accepted 30 Jun. 2014.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-2013-06-0284.

attachment, handover, idle mode etc.) are controlled by thecellular operator using the femtocell backhaul (DSL at home)to the core network. On the other hand, operators do notcontrol either the number or the location of femtocells [9]to keep their operational expenses at a certain level. Asa result, femtocells get typically deployed in an unplannedfashion within residential areas (e.g., apartments) by end usersand can be moved or switched on/off at any time. Hence,as femtocell deployments continue to grow, interference willinevitably be a performance limiting factor in a manner similarto that experienced by residential WiFi networks [10]. Theunplanned deployment and large number of femtocells makeinterference management complicated and prohibit the useof centralized network management techniques. To overcomesuch difficulties, the authors in [11] point out that femtocellsshould be self-configurable and self-optimizing and shouldmitigate interference in a distributed fashion.

There are two potential sources of such interference forfemtocells: (i) interference between femtocells and the macro-cell, and (ii) interference between femtocells. This paperfocuses on the problem of interference mitigation betweenmultiple femtocells (in range) that stems from their unplanneddeployment where the installed femtocells cannot be expectedto cooperate (within and across service providers), centralizedsolutions cannot be applied and new challenges arise.1 More-over, efficient resource management solution is necessary tomaximize the benefits of operating multiple femtocells. Forsuch settings, distributed self-organizing solutions that relaxthe assumption of femtocell cooperation are essential.Inapplicability of solutions used in macrocells or in WiFinetworks: The problem of interference mitigation throughefficient resource management is not new and has been exten-

1. Other work has focused on the macro-femto interference problem [12],and is beyond the scope of this paper.




sively studied in different settings. For instance, this problemarises even in the design of macrocell topologies by cellularoperators. However, the planned nature of such macrocelldeployments leads to very efficient frequency planning andother coordination-based solutions (e.g., fractional frequencyre-use), that cannot be applied to unplanned femtocell deploy-ments. WiFi network designs have also needed to addressthe problem for unlicensed spectrum. The basic toolkit forWiFi is to address interference problems include randomaccess channel contention strategies that de-synchronize com-peting transmitters in time through methods of carrier sensingand back-off. Unfortunately, femtocells cannot leverage suchstrategies since their MAC-PHY protocols are required tofollow the same synchronous channel access methods as theirmacrocell counterparts. More specifically, standard OFDMAfemtocells do not employ carrier sensing based deferral.Hence, they cannot sense the spectrum occupied by othercells to tune themselves onto orthogonal frequencies to resolveinterference. Finally, unlike in WiFi, a single OFDMA framecarries data for multiple clients, which poses unique challengesfor resource management solutions.RADION Distributed femtocell resource management:We propose RADION, a framework for distributed manage-ment of time-frequency resources in OFDMA-based femtocellnetworks. This framework is designed for the scenario wherenearby femtocells cannot explicitly coordinate or interact witheach other, and hence is suitable for unplanned residentialdeployments. Our specific solution requires each femto basestation (BS) to intelligently probe availability of resources anduse them in an opportunistic and distributed manner. We con-trast this design to a centralized resource management solution,called FERMI, that was recently proposed for femtocells [13].More specifically, FERMI focused on femtocell deploymentsin enterprise settings, where multiple femtocells were assumedto cooperate towards a network-wide objective. In particular,the design in FERMI required the centralized collection of theglobal network view, constructed through explicit informationexchange between the various femtocells. Such cooperationis not realistic for residential settings. Hence the solution inRADION is stylized such that each residential femto BS willtry to optimize for its own local objective, without explicitlyexchanging information with neighboring femtocell BSs.

Several challenges arise in designing a distributed resourcemanagement framework for OFDMA femtocells: (i) OFDMAschedules multiple clients in the same frame. The clients ex-perience different levels of interference and hence, a resourcemanagement framework has to account for the characteristicsof each client. Specifically, clients with strong interferenceneed to operate on orthogonal resources (i.e., frequencyisolation), while clients with weak interference can operateon all frequency resources (i.e., reuse). For efficient use ofresources in a frames, we need to first accurately identifyeach clients needs with respect to frequency isolation. (ii)Given that multiple clients share the frame resources, the nextchallenge is how to accommodate multiple clients of variousclasses in the same frame?. The frame resources have to becarefully managed for various clients considering their inter-ference levels and demands. In particular, each cell needs to

determine how much resources can be reused without causinginterference to the neighboring cells and how much frequencyresources need to be isolated to mitigate the interference fromother cells. The frame structure can impact the network wideresource reuse as well, where multiple contention domainsare involved. (iii) The resources allocated to one femtocelldirectly impact the resources for the interfering cells. We needto determine the time and frequency resources of operation forthe clients in each cell while accounting for the resources usedby neighboring cells. Resource allocation for each femtocellshould be adaptive to the network changes, but without explicitcoordination due to lack of central component in the system.

In RADION, we address the above challenges through thespecific design of the following three building blocks:

Client categorization: Active probing is used by the BSto categorize clients into two classes, those that requireresource isolation to mitigate the interference and thosethat can reuse the spectrum, based on interference levels.

Resource decoupling: Once the clients are categorized,RADION is capable of scheduling multiple clients in thesame frame by employing a three-zone frame structure.This new frame structure enables frame resource reuse,and hence promotes more efficient use of spectrum. Italso decouples resources across contention domains topromote better network-wide reuse.

Two-phase adaptation and allocation: The resources usedby each BS are determined by a fast and iterative mech-anism that is completely distributed. RADION employsa two-phase adaptation for determining jointly the timeand frequency resources for each femtocell.

Contributions: We design and implement RADION on aWiMAX testbed. We evaluate the various components ofRADION to highlight the functionality and accuracy of each.RADION maintains standards compatibility; it is immediatelydeployable with commercial WiMAX clients. Further, RA-DIONs components and modules are applicable to the otherOFDMA technologies, LTE and LTE-A. We wish to pointout that RADION is modular. In particular, depending on thecontext and the objectives, different resource allocations canbe easily incorporated within the RADION framework.

The rest of the paper is organized as follows. In section 2,we provide background on WiMAX and discuss related work.Section 3 discusses the challenges in managing resourcesin OFDMA femtocell networks and briefly describes howRADION addresses them. In section 4, the functional blocks inRADION are described. We evaluate RADION using testbedand simulations in section 5. Section 6 concludes the paper.

2 BACKGROUND AND RELATED WORKWiMAX preliminaries: While RADION applies to OFDMAsystems in general, our prototype is implemented on a Wi-MAX femtocell testbed. In WiMAX, the spectrum is dividedinto multiple sub-carriers and several sub-carriers are groupedto form a sub-channel.

WiMAX has a two-dimensional frame (see Fig.1) that car-ries data across time (symbols) and frequency (sub-channels).A combination of a symbol and a sub-channel constitutes a




DL Burst1

DL Burst2 ULBurst2

ULBurst1

FCH

Downlink Uplink

F

req

ue

nc

y

(Su

b-c

ha

nn

els

)

Pre

am

ble

DL

-MA

P

UL

-MA

P

Tiles

Time (Symbols)

MCS0: BPSK 1/2MCS1: QPSK 3/4MCS2: 16QAM 1/2MCS3: 16QAM 3/4MCS4: 64QAM 1/2MCS5: 64QAM 2/3MCS6: 64QAM 3/4MCS7: 64QAM 5/6

Fig. 1. WiMAX frame structure and MCS.

tile, which is the basic unit of resource allocation. Data tomultiple mobile stations (MSs) are scheduled as rectangularbursts of tiles in a frame and frames are sent out periodi-cally (every 5ms). In multi-cell OFDMA systems, frames aresynchronized both between the BS and MSs and across BSs.Mainly, frames consist of three parts; control (preamble, FCHand maps), downlink (from BS to the MS) and uplink (fromthe MS to the BS) data bursts. While the preamble allows a MSto associate with the BS, the control consists of FCH (framecontrol header) and DL/UL-MAPs. The DL-MAP indicateswhere each burst is placed in the frame, which MS it is sentto, and what modulation and coding scheme (MCS) should beused to decode it. The different modulation levels in 802.16eare shown in Fig. 1. While data bursts can be modulated usingany MCS, the control part is always transmitted using QPSK.The preamble is transmitted with a higher power compared tothe other parts of the frame.Related work: While OFDMA standards (e.g., WiMAX,LTE, and LTE-A) have been drafted recently, related researchhas existed for some time [11], [14]. Efforts that focus onthe macro-femtocells interference [15] and the interferenceto cell-edge users in OFDMA macrocells exist. The local-ized (cell edge) interference and planned cell layouts haveaided various interference management solutions [16] for bothdownlink ([17], [18]) and uplink ([19], [20]). These alsoinclude fractional frequency partitioning (FFR) approaches[21], where the spectrum is partitioned into pre-determinedstatic sets that require different levels of coordination betweenthe macrocells. Unlike macrocells, femtocells are deployed inan unplanned manner without coordinated operations and arehence, vulnerable to interference [22]. This necessitates novelinterference mitigation solutions that incorporate dynamic re-source partitioning strategies. There have been recent studies[15], [23], [24] in this direction but are restricted to theorywith several simplifying assumptions that restrict their scopeand deployment in practice.

Recently in [13], FERMI, a centralized resource man-agement solution for enterprise femtocell deployments wasproposed. However, RADION is a completely distributedsolution targeting unplanned deployment settings (such asresidences), where cooperation among femtocells (implicitlyassumed in [13]) is not realistic. The concepts of zoningand client categorization were also used in FERMI; however,access to global knowledge at a central controller made iteasy to design these functions. RADION allows each femtocellto intelligently determine its zones in a conservative manner.Unlike in FERMI, to account for inaccuracies in distributedzone determination, a transition zone is introduced (details

later). This in turn makes the client categorization in RADIONalso different from that in FERMI; clients that access themedium in the transition zone have to be chosen with care.Due to lack of coordination, accurate client categorization isall the more critical for efficient use of resources.

Diverging from the cellular context, recent efforts showthe benefits of OFDMA in WiFi context by building systemsthat enable dynamic spectrum fragmentation [25] and adaptivechannel width [26]. In WiFi, APs use one among multiple 20MHz frequency chunks (channels) and several conventionaldistributed channel selection algorithms [10] can be used toconfigure APs on different channels to avoid interference.However, in femtocells, the entire spectral chunk (say 20MHz)is available to all the cells. They need to operate on mutuallyorthogonal subsets of frame resources (sub-channels and timesymbols) to avoid interference. Thus, resource allocation hasto adapt to network dynamics (such as traffic, load etc.). Here,we seek to distributively and quickly determine the resourcesfor each cell. Further, resource allocation constitutes just onecomponent in our broader goal of designing a framework fordistributed resource management for femtocells.

3 RESOURCE MANAGEMENT CHALLENGESThe problem of resource management is for each femtocellto distributively determine the frame resources (tiles) that itcan use to schedule its clients. Since the resource allocationdecisions of one cell impact multiple other cells, efficient adap-tive and iterative mechanisms are needed to quickly convergeto a network-wide resource allocation. We first discuss thechallenges in achieving this objective and briefly describe howRADION addresses each of them.Client categorization: OFDMA schedules multiple clients inthe same frame. Different clients may experience differentlevels of interference from neighboring femtocells; clientssubject to strong interference need to operate on orthogonalsub-channels partitioned across cells (i.e., frequency isolation),while clients with weak interference can use the entire spec-trum (i.e., reuse) and still tolerate interference via link (rate)adaptation. Unlike in WiFi, interference avoidance comesat the cost of a reduced set of tiles per femtocell, whichin turn leads to under-utilization if not properly exercised.Specifically, not all clients in a femtocell require resourceisolation; link adaptation may suffice to cope with interferencefor clients that are in close proximity to the BS. Differentiatingthe clients is key in realizing good spectral efficiencies. Aclient categorization is included within RADION to accu-rately differentiate between such clients (which can reuse thespectrum reuse clients) and those that need spectral isolation(isolation clients). Since sensing the medium is not possiblewith standard OFDMA femtocells, RADION uses an intelli-gent active probing technique to achieve client categorizationwith high accuracy. Note that the accuracy of categorizationhas a direct impact on how efficiently the tiles are utilized.Resource decoupling among the clients: Once the clients arecategorized, next question is how to accommodate multipleclients of different categories (isolation and reuse clients)in the same frame?. Frame resources have to be carefully




5

15

ReuseTransition Isolation

5 15

c

R

R:T: I:

RI

I

21

ControlC:

Time

Frequency

0 27

c RI

c

5

T

21

BS1

BS2

Interference from BS1

21

5

c RI

210 27

(a) Introducing transition zone

0

3

6

9

12

0 20 40 60 80 100 120 0

0.2

0.4

0.6

0.8

1

Thro

ughp

ut (M

bps)

Bur

st D

eliv

ery

Rat

io

Time (Sec)

ThroughputBDR

(b) BDR and throughput measurement

RESOURCEISOLATION

F

req

uen

cy

FR

EE

RE

US

E

Time

OC

CU

PIE

D

(c) Measurement zones

Fig. 2. (a) BS2 sets transition zone to mitigate interference from BS1s reuse zone. (b) Burst Delivery Ratio (BDR)provides an accurate representation of throughput. (c) Free and Occupied zones are used for client categorization.

assigned to both the reuse and isolation clients; else, it leadsto under-utilization of network resources. Each frame can besegmented into two zones, the reuse and the isolation zone,to accommodate the two types of clients (BS1 in Fig.2(a));clients in the reuse zone will receive data encoded on the entirespectrum, while clients in the isolation zone will receive on asubset of frequencies (determined by an allocation algorithm).The static FFR approaches [21], using pre-determined sizesof reuse and isolation zone across cells, will however notwork in femtocell deployments where interference is pervasive(i.e., not localized). Further, the load of a femtocell as well asinterference from other BSs also need to be taken into accountin adapting the zone sizes of a particular cell. RADION usesa novel three-zone frame structure to address these issues aswill be described in section 4.3; in a nutshell the additionalthird zone is a transition zone that prevents resource couplingacross cells (discussed next).Resource decoupling across the femtocells: When twointerfering femtocells have different reuse zone sizes (basedon their loads), the larger reuse zone will interfere with theisolation zone of the other cell (BS2 in Fig.2(a)). Having acommon reuse zone for the two cells is essential to avoidthis. However, irrespective of whether the maximum or theminimum of the reuse zone sizes is chosen as this commonzone size, it is easy to see that there is under-utilization inone of the cells (either part of the reuse zone or part ofthe isolation zone is not utilized). While this coupling andresulting resource under-utilization is inevitable to alleviateinterference, the challenge is to avoid this coupling frompropagating to the entire network, which would result insignificant under-utilization that is unwarranted. RADIONstransition zone intelligently localizes this resource couplingand prevents such propagation.Resource allocation: Each femtocell has to determine itszone sizes and resource usage in a completely distributedmanner. RADION uses an iterative, joint time (zone sizes)and frequency (sub-channels in the isolation zone) resourceallocation algorithm that converges to efficient allocations andadapts to network dynamics quickly and efficiently.

4 RADION AND ITS COMPONENTSWe next describe RADION in detail by elaborating on thefunctionalities of its building blocks. We explain how building

blocks work and present experimental results of the compo-nents to validate their efficacy.

4.1 Client Throughput EstimationThe interference from the neighboring BS can be measuredonly from the client by accounting the throughput. While afemto BS does not have direct access to the throughput ata client, it receives feedback about the reception of each databurst via ACKs on the uplink frame. We define Burst DeliveryRatio (BDR) to be the ratio of successfully delivered bursts tothe total number of transmitted bursts. The BS can estimateBDR by taking the ratio of the number of ACKs receivedto the total number of feedbacks received for a given client.We perform experiments to understand if the BDR estimateat the BS can provide an understanding of the throughput atthe client. While transmitting data burst to a client, we movethe location of the client to sample varied SNRs. Fig.2(b)shows that indeed the BS can very accurately track the clientthroughput using the BDR estimates. We use BDR as a goodestimation of a client throughput to measure the interferencelevel in RADION. The clients use existing (dedicated) uplinkACK channels on robust MCS for transmitting feedback toBS, thereby not imposing additional overhead.

With the help of BDR estimates, the femto BS can nowcompensate for the lack of sensing capability by clientsthrough active probing, whereby the BS schedules its clientson specific resource regions in the frame and estimates theircorresponding BDR to detect resource availability. Such prob-ing can be done by transmitting small amounts of data tospecific clients on specific frame resources for a very shorttime interval (e.g., order of 100 msec) while accounting ACKburst receiving every 5 msec through uplink channel. In thismanner, probing (using BDR) can indeed detect the resourceusage quickly.

4.2 Client CategorizationInitial step: The second component in RADION categorizesclients into two classes; the first needs just link adaptation(class 1) and the second needs resource isolation (class 2).To determine the clients class, we need to estimate thethroughput that a client receives with and without interference.One can expect that if the throughput with interference iscomparable to that without interference, the client is class




1; else it is class 2. This is achieved through active probingand BDR estimation. To facilitate client categorization, theframe structure contains two measurement zones of equal size(only 2 symbols); free and occupied zones as depicted inFig. 2(c). All BSs are required to transmit in the occupiedzone constantly, therefore scheduling a client in the occupiedzone enables BS to calculate the BDR in the presence ofinterference from other BSs. The free zone will be used for ob-taining the interference-free BDR estimation with probabilisticscheduling. However, it is possible that more than one BSwould simultaneously schedule its client in the free zone andit would yield an inaccurate BDR estimation correspondingto the case of sampling collision. In order to avoid this, arandom access mechanism with probability 1

n

is used wheren is the number of interfering BSs (the interfering BS set canbe inferred without cooperation by leveraging signal strengthmeasurements from clients, e.g., each client reports the signalstrength from different cells called the automatic neighborreports in LTE). Each BS performs the following two stepstowards categorizing a client leveraging these two zones:

1) Schedule the clients data in the occupied zone andschedule it in the free zone probabilistically. Keep trackof the resulting BDR in both zones over K frames.

2) Determine the normalized throughput per tile in the twozones T

occ

and Tfree

corresponding to their BDRs. IfT

free

(1 + )Tocc

then the client is categorized asclass 2; otherwise class 1.

Through exhaustive measurements (Table 1), we set = 0.25and K = 25 frames. This parameter setting yields very highaccuracy (>90%) of client categorization and do not dependon the network size or load. Class 1 clients are scheduled inthe reuse zone using all sub-channels while class 2 clients arescheduled in the isolation zone using subset of sub-channels(reuse and isolation zones are depicted in Fig.2(c)).Refinement step: After initial categorization, we refine itwith a sub-classification of class 2 clients that is suited fordistributed operation. In the above step, while clients in class1 are identified accurately, not all clients categorized as class2 may need resource isolation. With resource isolation, a BSallocates only a subset of sub-channels to a client. For clientswith low-moderate interference, link adaptation to cope withinterference may be a better option than sacrificing resourcesthrough isolation. Thus, to further refine the categorizationof class 2 clients, we factor in the loss of resources dueto isolation. This was missing in the step 2) in the initialcategorization, since equal resources were used in the occupiedand free zones. The amount of isolated resources availableto a cell depends on the resource allocation algorithm. Ifthe resources assigned to the isolation zone is a fraction fof that of the reuse zone, the BS refines the status of aclient in class 2 by scheduling the client on resources inthe isolation zone and determining its normalized per tilethroughput in this zone, T

isol

. It retains the client in class 2only if f T

isol

(1+)Tocc

; otherwise the client is revertedto class 1. Here, (0.05, experimentally obtained) is usedto avoid oscillations in categorization. RADION further sub-classifies clients in class 2 as those that benefit significantly

= 0.25 = 0.5 = 1 = 1.5 = 2K = 25 0 0 0 13% 29%K = 50 0 0 0 12% 28%K = 125 0 0 0 12% 27%K = 250 0 0 0 12% 22%K = 500 0 0 0 10% 20%

TABLE 1Categorization error with respect to and K.

1

2

1 2 3 4 5 6 7 8 9 10111213141516171819

Clie

nt ca

tegory

Location

Ground truthInitial step

Refined step

Fig. 3. Refinement step yields accurate categorization.

from resource isolation (class 2h) and those that benefitmarginally from it (class 2l);

Class 2h client, when f TisolT

occ

(1 + ) Class 2l client, when (1 + ) f Tisol

T

occ

< (1 + )

The benefits of such a sub-classification will be discussed insection 4.3.Evaluation: We consider two cells (1 and 2); clients 1 and2 belong to the two cells respectively. We generate multipletopologies with different levels of interference by varying thelocation of client 1 in the presence of interfering cell 2. Theisolation zone of each cell has the same number of symbols asthe occupied zone but operates on an orthogonal half the sub-channels. First, for every location of client 1, its ground truthis determined by scheduling data to the client in the occupiedand isolation zones and determining it as class 2 if 0.5T

isol

T

occ

. Then our three step (two initial steps and refinement step)categorization algorithm is executed. The categorization resultsof the initial and the refinement steps are shown in Fig.3. Itis seen that the initial step wrongly categorizes clients in tenlocations, however, the refinement step corrects most (nine) ofthem. The only erroneous classification (location 2) was due toa change in channel conditions during the process. The clientswho need to be in class 2 are correctly classified even withthe initial step, while refinement only adds more clients fromclass 2 to class 1. The client categorization, both the initial andrefinement steps, yields 94.7% accuracy (comparing groundtruth and refined result in Fig.3).

4.3 Resource DecouplingTwo-zone structure: To schedule clients of both classes inthe same frame, we use two variable size data of reuse andisolation zones. Zones are virtual partitions of frame fortransmitting data burst to different clients. The reuse zoneoperates on all sub-channels and schedules class 1 clients,while the isolation zone schedules class 2 clients on only a




0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

CD

F

Throughput(Mbps)

BaselineZoning

(a) Using two-zone structure

0

0.2

0.4

0.6

0.8

1

4 5 6 7 8 9

CD

F

Throughput(Mbps)

2Z2Z-CR

3Z3Z-Opt

(b) Using three-zone structure

BS1

10

15 10 5

Demanded Reuse Zone Size (symbols)

15 5 10

R R RI

I

IT

15

T

BS2 BS3

(c) Illustration of frame (three zone) usage

Fig. 4. (a) 35% throughput gain achieved with two-zone structure. (b) Three-zone structure yields a 35% throughputimprovement over two-zone structure. (c) Use of reuse, transition and isolation zones in three cells topology.

contiguous subset of sub-channels. To understand the benefitsof the two-zone structure, we conduct experiments with twoBSs. Each BS has two clients (one in each class), and interfereswith the class 2 client of the other BS. The baseline schemeoperates the two BSs on two orthogonal sets of sub-channels,with each BS scheduling both its clients within its ownsubset. This is compared against a two-zone scheme wherea BS schedules its class 1 client on all sub-channels, whilethe isolation zone uses the other half of the frame for itsclass 2 client on half the sub-channels. We generate variousinterference topologies and the CDF of the net throughputis plotted for the two schemes in Fig.4(a). We see that withthe two zones, class 1 clients can be scheduled in tandem toreuse sub-channel resources effectively, yielding over a 35%throughput gain over the conservative isolation scheme.

Drawbacks of two-zone structure: A two-zone structureenables resource reuse, but is insufficient in a multi-cellcontext. Different cells will have different reuse zone sizesbased on the load generated by the class 1 clients. If twointerfering cells were to operate independently, the larger reusezone will overlap with the isolation zone of the other celland hence, interfere with the isolation clients of that cell(Fig.2(a)). Consequently, it is important to use a common reusezone between interfering cells. For the example topology inFig.4(c), the result in Fig.4(b) indicates that the throughputof BS2 (with the smaller reuse zone) is degraded whenits isolation zone starts right after its reuse zone withoutaccounting for the larger reuse zone of BS1 (2Z curve). Thisis in comparison to even a simple scheme (2Z-CR curve) thatstarts the isolation zone of BS2 only after the end of the reusezone of BS1, leaving the region between the reuse zones ofthe two BSs unused in BS2.

Need for common reuse zone: Each cell has to determinethe common reuse zone in a distributed manner. The commonreuse zone size can either be the minimum or maximumof the reuse zone sizes of its neighboring interfering cells.RADION uses the maximum of the reuse zone sizes within theinterference neighborhood as the common reuse zone for tworeasons: (i) since the deployment is un-coordinated and non-cooperative (e.g., residential complexes), there is no incentivefor a cell to decrease its reuse zone size (only other cellssuffer); (ii) given the absence of sensing by the clients, activeprobing by the BS to determine resource availability can beemployed to only determine the reuse zone of a cell with a

larger zone size effectively (elaborated in section 4.4).Three-zone structure: RADION uses a three-zone trans-mission structure as shown in Fig.4(c). While a BSs reusezone remains the same (for class 1 clients), its isolationzone only begins from the end of the common (maximum)reuse zone in its neighborhood. The region between its reusezone and common reuse zone is the transition zone. Blindlyscheduling class 2 clients in the transition zone is harmful dueto interference from neighboring cells with larger reuse zonesizes. Leaving the transition zone empty (2Z-CR curve in Fig.4(b)) will under-utilize resources. RADION intelligently picksclass 2 clients (using the sub-classification in section 4.2) tobe scheduled in the transition zone on the same subset of sub-channels as the isolation zone. Specifically, class 2l clientsare scheduled in the transition zone, while class 2h clientsoperate in the isolation zone. There are two benefits to suchan approach. First, the transition zone allows selected class 2clients to opportunistically reuse resources without incurringsignificant interference. Second, operating the transition zoneon the same subset of sub-channels as the isolation zoneprevents the common reuse zone from propagating to theentire network (across multiple contention domains), therebyeliminating under-utilization due to resource coupling. To seethis, in Fig. 4(c), if the transition zone is allowed to operate onall sub-channels, then the common reuse zone that terminatesat 15 symbols in the contention domain between cells 1 and2, will also propagate to cell 3. This allows cell 3s class2h clients to be scheduled only after 15 symbols, while thecommon reuse zone in its own contention domain (betweencells 2 and 3) terminates at 10 symbols.Evaluation: The benefits of RADIONs three-zone structureare evident in Fig.4(b). While both 3Z and 3Z-Opt employa three-zone structure, 3Z-Opt employs intelligent schedulingof class 2 clients in the transition and isolation zones, 3Zrandomly schedules class 2 clients in the two zones. It is clearthat the three zone structure with sub-classification yields a35% throughput improvement.

4.4 Distributed Allocation FrameworkWe now describe the distributed resource allocation process.The goal of each cell is to determine the size of the reuse (s

r

)and transition (s

t

) zones as well as the specific contiguousset of sub-channels (C) for operation in the isolation zone.Each BS first classifies its clients into classes 1, 2l and 2h.




The preamble used by each BS is chosen from a standardizedset (of orthogonal sequences); using this, the clients measurethe signal strength to various BSs and hence lock on to aBS with strong signal strength. The same process is usedby each client in class 2h to determine its set of stronginterferers (received signal strength over a threshold, S

th

).Using feedback from these clients, the BS determines thesuper-set of strong interferers. The cardinality of this setincluding itself (n) determines the fair share allocation of sub-channels (m = N

n

, where N is total number of sub-channels)for the clients in the isolation zone (class 2h)2. Then, based ondata availability and the set of clients chosen for transmissionby the MAC scheduler (using proportional fair scheduling),the BS determines the desired size of its reuse zone (s

r

). Thisis proportional to the relative traffic load from the clients in thetwo classes. Next, the BS determines the common reuse zone(in time, s

t

) in its interference neighborhood as well as thespecific set of m sub-channels (in frequency, C) for operationin the isolation zone (details to follow). After determiningthe resource allocation parameters, clients in classes 1, 2land 2h are scheduled in the reuse, transition and isolationzones respectively. Determination of the resource allocationparameters in RADION is accomplished with the followingjoint time-frequency probe-adapt mechanism.

4.4.1 Two-phase AdaptationRADION employs a combination of both coarse and fine timescale adaptations (Fig.5). Every BS picks a period P of coarseadaptation (order of several seconds; thousands of frames).P is picked from a set of large prime numbers to reducethe frequency of overlap of adaptation (and probing) periodsacross cells. The goal of coarse adaptation is to track coarsenetwork dynamics such as (de)activation of cells/clients, loadchanges, etc., that happen at the granularity of several seconds.Every P frames, each BS triggers a series of fine timescale adaptations automatically (can also be based on changesdetected in interference conditions as indicated by its clients).Once triggered, the goal of fine adaptation is to quicklyconverge to the right set of allocation parameters (s

t

, C), fora given set of network conditions, while that of the coarseadaptation is to track coarse network dynamics such as trafficload changes, (de)activation of clients, etc. that happen atthe granularity of several seconds. During fine adaptation, theBS performs a probe-and-adapt procedure every q frames tillconvergence, where q is randomly selected from [1, 0.1P ] andoperates at the granularity of hundreds of milliseconds. Therandomness of q minimizes probing collisions.

Employing coarse adaptation in isolation will result in longrecovery times in the event of probing collisions across fem-tocells. This leads to large periods of degraded performance.On the other hand, employing fine adaptation in isolationwill require continuous probing to track network dynamics,thereby resulting in large overhead. RADION strikes a goodbalance between coarse and fine adaptations; fine adaptationis suspended after quick convergence to an efficient resource

2. The solution can be extended to weighted allocations, albeit at the costof information exchange between femtocells.

Invoke coarse adaptationevery P frames

Max reuse zone found?

Select probing res.(t)(sequential or binary)

No

Yes

Same chunk?

NoYes

Wait q framesProbe n chunks

Probe n chunks at tCompare BDR;Select chunkUpdate current BDR

Fig. 5. Flow chart of RADIONs two-phase adaptation.

allocation and invoked again only in the next coarse adap-tation period. Thus, the BS spends a large fraction of its Pframes operating on an efficient allocation with its probingand adaptation mechanism constituting only a small portionof it. Further, even during the probing-adaptation procedure,data scheduling to clients is not interrupted and is seamlesslyincorporated into it.

4.4.2 Probe and AdaptThe goal of fine adaptation is to ensure quick convergencein the determination of resource parameters in both time(common reuse zone) and frequency (sub-channels in isolationzone) domains. This is achieved with a joint time-frequencyadaptation algorithm. Each BS probes a vertical strip ofresources in the frame, of size t N (i.e., encompassingall sub-channels in time t), where t is the granularity ofprobing in the time domain (few symbols). The frequencydomain is further probed in chunks of f contiguous sub-channels (f = m). t and f can be varied to tradeofffine grained allocation and convergence time. Note that ifclients are capable of sensing (e.g., cognitive clients), thenmultiple sub-channels can be sensed simultaneously to allowfine grained allocation without impacting convergence time.Again, RADION uses K=25 frames to probe resource and itonly takes less than 600 msec to probe entire frame resources.The probing mechanism is fast enough to detect the bursty-ness nature of interference.Joint probing in time and frequency: When coarse adap-tation is triggered, the BS probes resource regions after itsown reuse zone to determine the common reuse zone in itsneighborhood (step 2 in Algo.1). The intuition is that sincethe interfering cell with the largest reuse zone will use allsub-channels till its reuse zone, when frequency chunks areprobed within the largest reuse zone, they will exhibit similar(degraded) BDRs, while when probed beyond the largest reusezone, there will be at least one frequency chunk, whose BDRexceeds those of the other chunks by (see inference insection 4.2). This observation is used by every BS to determinethe common reuse zone. Specifically, to probe in a verticalresource region when a vertical resource strip in the frameis probed, the BS transmits data to a client in each of nrandomly chosen frequency chunks of size m. Since P varies




Algorithm 1 RADION: Distributed Resource Allocation Frame-work

1: m = Nn

, ci

2 C, st

sr

, bi

: BDR, ei

: counter, 8ei

=02: Probe c

c

, update bc

, st

+=t, fc

: current frame3: while (f

c

mod P ) 04: for i = 1 : n5: Probe c

i

, update bi

/*probe n frequency chunks*/6: b

u

: u = max

i:ci

2C(bi) /*find the max BDR*/7: if (b

u

> b

c

) k (bc

> )

8: Select cf

: call Algo.2 or 39: c

c

cf

, update bc

(we found st

)10: Pick q 2 [1, 0.1P ], wait q frames /*fine adaptation*/11: for i = 1 : n12: Probe c

i

, update bi

13: Select cf

: call Algo.2 or 314: if c

c

cf

15: 9ci

2 C \ {cf

}, s.t bi

> , ei

++16: 9e

i

2, pick one of the ci

, ei

=0; otherwise ci

=;17: c

c

cc

[ci

, goto step 2 /*frequency converged*/18: else19: c

c

cf

, goto step 10 /*re-do fine adaptation*/20: else21: s

t

+=t, cc

cu

, update bc

, goto step 4

Algorithm 2 Gibbs Sampler: Frequency Resource Selection1: Temperature parameter: T = 0.052: 8c

i

2 C compute the probability:(c

i

) = (e

b

i

1T

)/(

Pn

i=1 eb

i

1T

)

3: Sample a random variable rand with law 4: Select c

f

according to rand

Algorithm 3 Greedy: Frequency Resource Selection1: index: i = argmax

i:ci

2C(bi), cf ci

across cells, and the frequency chunk to be probed is chosen atrandom, probing conflicts across resource regions are avoided(probability of collision is 1

Bi=1PinB

, where B is the numberof BSs and evaluation is in section 5). Each chunk is probedfor twenty-five frames and the BDR on each of these chunksis estimated (step 4,5); the maximum BDR (across chunks) iscompared to the clients current recorded BDR (step 6,7).Convergence in time: We consider two approaches to probingin the time domain: sequential and binary search (Fig.6). Insequential probing (outlined in the pseudo-code), the verticalstrip to be probed is advanced sequentially by t till a gainexceeding or a high value (>=0.83) of BDR (for BSconstituting the maximum reuse zone) is seen compared to thecurrent BDR (step7). Otherwise, the current BDR is updatedbased on the maximum BDR with the recent probing (step21). In binary search, two adjacent vertical strips are probedand the BDR with the left and right strips are compared. IfBDR

right

> BDR

left

then common reuse zone has been

3. Considering the wireless loss of ACK bursts, we set to be lowest butclose to 1, while it yields 0 probing error.

Time(symbols)

2

Sub-channel

Sequential

1.if gain>;

converged

2.otherwise;

move t

t

1

31 Binary

1.BDRr>BDRl

converged

2.if gain>;

left

3.if gain ) or right ( ), and the current BDR isupdated only when the region probed is within the maximumreuse zone. If there are multiple clients in class 2h, they areprobed together in each of the chunks and decisions are madewith respect to each client. Since different clients may receiveinterference from different cells, the common reuse zonevaries with respect to clients. Time domain probing continuestill the common reuse zone for each client in class 2h isdetermined, with the largest common reuse zone determiningthe termination of the transition zone for the cell.Convergence in frequency: Once the common reuse zone isdetected, the BS simultaneously has the BDR information on nfrequency chunks, with multiple frequency chunks potentiallyavailable for operation. We consider two approaches for theselection of a frequency chunk (step 8,9): greedy and Gibbssampler (Algo.2 and 3). While the greedy scheme is deter-ministic and picks the chunk yielding the highest BDR, theGibbs sampler is probabilistic and favors chunks with higherBDR [10]. It has a temperature parameter T , which can bevaried with time to provide an annealed version that convergesto stable states of low potential (low interference and highBDR). While convergence in the time domain (common reusezone) can be achieved with high accuracy, frequency domainconvergence is sensitive to frequency selectivity and channelerrors. Hence, the frequency chunk selected is confirmed byanother iteration of fine adaptation, which probes the fre-quency chunks alone (i.e., no increment of t) after q frames(step 10-13). If the same frequency chunk yields the highestBDR, then there is convergence in the frequency domainand the chosen time and frequency parameters are employedfor operation till the next coarse adaptation (step 14-17).Otherwise, probing in the frequency domain is repeated everyq frames till contention (interference) is alleviated (step 18,19).Thus, by probing vertical strips of frequency chunks, RADIONdetermines both the common reuse zone (s

t

) and the set ofsub-channels (C in isolation zone) simultaneously; this leads toquick convergence. Note that BSs adapt both control channeland frequency resource isolation (i.e., FDIctrl isolation [27])




when converging/operating in frequency domain to completelyavoid the interference from the control channels caused byneighboring cells.

4.4.3 Handling Network DynamicsClient (dis)-associations impact the traffic load of a cell andconsequently, the resource allocations in the isolation zone.From an ideal (centralized) resource allocation standpoint,every cell has to share the frequency resources in the isolationzone of a frame in the contention domains that it belongsto (cliques in the interference/conflict graph), with its idealshare being determined by the size of the largest contentiondomain that it belongs to. When a new cell is introducedor an existing cell leaves (or traffic ceases) the contentiondomain, the existing share of frequency resources decreasesor increases, respectively. However, this change has to bedetected by each cell in a completely distributed manner inRADION; this allows cells to contract and expand their sub-channel allocations in the isolation zone. Such a feature is alsouseful in improving the resource utilization in the network.Note that, since a cell (A) does not have information on itscontention domains (requires global knowledge), it computesits fair share (say x, x ideal share) based on its interferingneighbors. However, if one of its neighboring cells (B) belongsto a larger contention domain (hence has a lower share < x),then some resources (unused by B) will be under-utilized incell As contention domain. The ability to probe and expandresource usage will avoid such under-utilization that is a by-product of distributed operations (smaller granularity of chunksizes (< f ) lowers such under-utilization).

RADION allows a cell to adapt its sub-channel usage asfollows. Although a cell selects one of the frequency chunksfor operation upon convergence, it keeps track of the BDRin other frequency chunks and the potential set of chunksunused by neighboring cells. It continues to monitor suchunused chunks for an additional period of P frames, givingits neighbors enough time to detect and use their fair share. Ifsome of these chunks still continue to be available, then theBS decides to expand its resource usage by adding one of theunused chunks to its allocation (step 16). Adding one chunkat a time, allows other cells in its contention domain to alsoshare the unused resources in a fair manner. This expansionof resource usage will address cases when cells switch off orcease to carry traffic. However, if after expansion, the ceasedtraffic in a cell restarts or a new cell enters the contentiondomain, this will be detected in the form of degraded BDRson the frequency chunks or as a new interferer sensed by itsclients. In either case, the BS will contract to its conservativeshare of sub-channels (in the isolation zone) computed basedon its updated set of interfering neighbors and re-run itsadaptation algorithm. Any resulting under-utilization in itscontention domains will be addressed subsequently throughits resource expansion mechanism.

5 SYSTEM EVALUATIONTestbed: Our testbed consists of three femtocells deployed inan indoor environment reflecting a typical femtocell deploy-ment with walls, doors, and rooms (Fig.7). We use PicoChips

Fig. 7. Deployment and picture of our WiMAX testbed.

femto BSs that run WiMAX (802.16e) [28]. Our clients arelaptops with commercial USB WiMAX cards [29]. Eachfemtocell has multiple clients (class 1, 2l and 2h clients) andall class 2 clients are located within the transmission range ofother femtocells. All the three cells operate on a 8.75 MHzbandwidth with a carrier frequency of 2.59 GHz. We achievesynchronization across femtocells via external GPS modules[30], aligning the start times of frames.

Our testbed is restricted to three femtocells, however, theimpact of interference generated by two interferers is suffi-cient [13] to evaluate the proposed framework. We augmentevaluations with simulations, where we study scalability underdense deployments with large number of BSs. Unlike WiFi,femtocells perform synchronous frame transmissions withoutcarrier sensing. Hence, it makes more sense to generate differ-ent interference topologies by varying the locations of clientsnot the BSs. This provides a finer control on the inter-BSinterference magnitude as opposed to changing the locationsof the BSs, and also covers a wide range of scenarios thatinclude both line-of-sight (LOS) and non-LOS links. We alsoconsider ideal link adaptation where we sequentially run theexperiment over all MCS levels and record the one deliveringthe highest throughput for the given topology.Implementation: RADION is implemented on the PicoChipplatform [31], which provides a base reference implementationof the 802.16e standard. We significantly extend and modifythe MAC scheduler (2000 lines C code) to realize thecomponents in RADION. Our key modifications include: (1)BDR estimator: The BS tracks the BDR status for eachclient connection and updates the BDR estimates as a movingaverage. BDR estimation plays an important role in both clientcategorization and probe-and-adapt mechanisms of RADION.(2) Client categorization: We introduce two measurementzones (i.e., free and occupied zones) for client categorizationand schedule data bursts in each measurement zone. (3)Resource decoupling: Each BS tracks its clients categoriesto schedule them in the appropriate zone. From the set ofclients to be scheduled, the BS determines the traffic demand(data) from clients for each zone and splits the frame toschedule them simultaneously. (4) Probe-and-Adapt module:The two-phase adaptation algorithm is implemented on eachBS; adaptations are triggered at frame boundaries. Further,RADIONs modular nature allows us to incorporate the twovariants of channel selection as outlined in section 4.

5.1 Prototype EvaluationsHaving evaluated the first two components of RADION insection 4, here we focus on evaluating the two-phase adapta-




1-10

11-20

21-30

300 320 340 360 380 400

Su

b-c

ha

nn

els

Time (sec)

BS 1BS 2BS 3

(a) Fine adaptation

1-10

11-20

21-30

400 420 440 460 480 500 520

Su

b-c

ha

nn

els

Time (sec)

BS 1BS 2BS 3

(b) Coarse adaptation

1-10

11-20

21-30

0 100 200 300 400 500 600

Su

b-c

ha

nn

els

Time (sec)

BS 1BS 2BS 3

(c) Two-phase adaptation

Fig. 9. Three convergence patterns. Two-phase adaptation incurs very few false switches (only twice) and providesvery fast recovery from false decisions.

0

4

8

12

16

20

Coarse Fine Two Phase 0

3

6

9

12

Ove

rla

p D

ura

tio

n (

%)

Ove

rla

p D

ura

tio

n

pe

r F

als

e D

ecis

ion

(se

c)Ratio

Duration

(a) Overhead

0

3

6

9

12

15

Coarse Fine Two Phase 0

2

4

6

8

Fa

lse

De

cis

ion

s (

%)

Pro

bin

g C

olli

sio

ns (

%)

False DecisionsProbing Collisions

(b) False decision

Fig. 8. Two-phase adaptation outperforms others.

tion algorithm. To understand the efficiency of algorithm wefirst evaluate the adaptation process in time and frequencydomains in isolation, followed by their joint evaluation. Weconclude the evaluation of RADIONs adaptiveness to networkdynamics. We create multiple clique topologies where clientsin class 2h are within the transmission range of other BSsby varying client locations. The BSs operate on frames with30 sub-channels and 22 time symbols for data bursts. Eachexperiment is run for at least 10 minutes and is repeatedmultiple times to generate confidence results.

5.1.1 Frequency Domain ConvergenceHere, we fix the reuse zone of all the three cells to bethe same, thereby eliminating the need for determining acommon reuse zone. Hence, the focus is only on frequencydomain convergence in the isolation zone for class 2h clients;every cell has to identify and operate on a contiguous setof 10 sub-channels each (3 contending cells). We evaluateRADIONs two-phase (coarse + fine) adaptation against thecoarse and fine time scale adaptations in isolation. The coarseadaptation period is chosen to be a prime number of framesin [1000,6000] for each BS and is fixed for subsequentadaptations. Similarly, the period of fine adaptation is chosenrandomly in [1,600] for every adaptation. Further, selection ofsub-channels after probing follows the greedy approach (Gibbssampling considered later).

To understand the impact of false decisions on system(throughput) performance, we consider the following metrics:(i) fractional overlap duration: the fraction of time that cellsoperate on overlapping resources (leading to collisions); (ii)Overlap duration per false decision: ratio of the net durationof resource overlap to the total number of false decisions

(resource overlap collisions); (iii) Fractional false decisions:ratio of net false decisions to total number of decisions; (iv)Fractional probing collisions: ratio of net probing collisions tototal number of decisions. A probing collision occurs if twoor more cells probe at the same time and choose the sameresources for isolated operations resulting in a false decision.However, a false decision can also occur due to inaccurateBDR estimates and/or asymmetric interference patterns. Weomit the throughput results and rather focus on the conver-gence to orthogonal resources among the BSs. Throughputdegradation only occurs when the BSs operate on the sameresource therefore, the fractional overlap duration reflects thethroughput result indirectly.

The results of multiple runs of frequency convergence arein Fig. 8(a) and 8(b). Coarse adaptation incurs fewer falsedecisions compared to fine adaptation, but the time to recovery(overlap duration) is large per false decision. In contrast,while the overlap duration per false decision is small forfine adaptation, probing collisions and hence, false decisionsincrease with the number of active probes. RADIONs two-phase adaptation strikes a good balance between coarse andfine adaptations, resulting only in 2% of the frames collidingand hence outperforms others in all metrics.

To illustrate, we present one particular example of sub-channel convergence as a function of time in Fig.9. We foundsimilar patterns for all other runs. While the cells maintainorthogonality of sub-channels for most part of the experiment,several differences are observed between the schemes. Withfine adaptation (Fig.9(a)), at 352 seconds, BS2 switches fromsub-channels 21-30 to 1-10, which is already occupied byBS3. This is a false decision due to an inaccurate probingestimate. After BS2 switches to sub-channels 1-10, BS3switches to sub-channels 11-20 when its next adaptation isinvoked (after q frames). Thus, each cell reacts to othersdecisions to avoid interference and maintains orthogonal sub-channel usage for its class 2h clients. While the numberof such switches is small with coarse adaptation (Fig.9(b)),recovery from a false decision is also slow. While recoveryis fast with fine adaptation, multiple switches result in bothincreased probing overhead and false decisions. The two-phaseadaptation (Fig.9(c)) exhibits the best of both adaptations; itincurs fewer switches while also providing fast recovery fromfalse decisions.




0

4

8

12

16

20

Greedy Gibbs 0

2

4

6O

ve

rla

p D

ura

tio

n (

%)

Ove

rla

pp

ing

Du

ratio

n

pe

r F

als

e D

ecis

ion

(se

c)

RatioDuration

(a) Overhead

0

4

8

12

16

20

Greedy Gibbs 0

1

2

3

Fa

lse

De

cis

ion

s (

%)

Pro

bin

g C

olli

sio

ns (

%)

False DecisionsProbing Collisions

(b) False decision

1-10

11-20

21-30

0 50 100 150 200 250 300 350 400

Sub-c

hannels

Time (sec)

[ ]

[ ]

BS 1BS 2BS 3

(c) Frequency convergence

5

9

13

17

21

25

160 180 200 220 240

Tim

e S

ym

bol

Time (sec)

BS 1BS 2BS 3

(d) Time convergence

Fig. 10. Greedy selection outperforms Gibbs sampler (a and b). Convergence patterns of greedy selection (c and d).

5.1.2 Time Domain ConvergenceHere, the three cells are pre-assigned orthogonal sets of sub-channels in the isolation zone. We focus on their convergenceto the common reuse zone. Recall that RADION probes mul-tiple frequency chunks and compares their BDRs to determineconvergence to this common zone. Hence, estimation of BDRplays an important role in the accuracy of common zonedetermination. We study both convergence and the impactof the number of probing frames used, for estimating BDR.Sequential search is used for time domain adaptation. Thethree cells are set different reuse zone demands of 9, 13 and 17symbols, respectively. BS3, having the maximum reuse zone,will quickly converge to the common reuse zone, while theother two cells will require adaptation. For a given number ofprobing frames (per chunk), we run the adaptation experimentover 100 times and determine the fraction of cases where thecommon reuse zone is accurately determined by cells 1 and 2.We repeat this experiment by varying the number of probingframes. We observe that 25 probing frames are sufficient tocorrectly determine the common reuse zone in over 94.5% ofthe cases (more frames only provided marginal improvements).

5.1.3 Joint Time-Frequency ConvergenceWith three BSs (1,2,3) in a clique topology with reuse zonedemands set at 17, 13 and 9 symbols respectively, we runthe joint time-frequency adaptation process of RADION ateach BS. From the results of multiple runs in Fig.10(a) and10(b), we see that the greedy sub-channel selection yieldsquick convergence with a low false error rate. The Gibbssampler incurs a higher fraction of collided frames comparedto the greedy. Since both the schemes employ the two-phaseadaptation, the average collision duration per false decision issimilar with the two schemes (Fig.10(a)). While one mightexpect the probabilistic nature of Gibbs sampling to yieldbetter convergence [10], this is only true if inferences are basedon sensing as opposed to probing; the probabilistic selectionincreases the number of probing collisions and hence falsedecisions (Fig.10(b)). Although the greedy approach is deter-ministic, diversity in sub-channel gain across BSs (and theirclients) implicitly results in cells picking different frequencyresources for their operation. Since both the schemes employthe same time-domain adaptation procedure, their convergenceerror in the time domain remains similar and less than 5%.

The time-frequency convergence patterns for the greedyscheme are shown in Fig.10(c) and 10(d). Here we present

one example of the multiple runs but we confirm very similarpattens for all runs. The frequency adaptations are directlyindicated. False convergences in the time domain are indicatedby brackets to capture the duration for which the cell operateswith a wrong common reuse zone. We see that the numberof probing collisions and the resulting quick switch in sub-channel resources to maintain orthogonality in the frequencydomain. We omit the pattern of Gibbs sampling for the sakeof brevity; however, we highlight that it shows many morefrequent switches as the results (Fig.10(a) and 10(b)) show.The time convergence pattern for the greedy approach isexpanded in Fig.10(d). We see that BS3 computes the commonreuse zone falsely as 21 symbols (instead of 17) at 200 secs;this is corrected in the next coarse adaptation period. Tillthe reuse zone is corrected, BS3s false decision does notcause interference to BS1s and BS2s class 2h clients sinceit continues to use its isolated resources in its transition zone.However, BS3s class 2h clients now incur unfairness as theiroperational resources are reduced by four symbols. Since thenumber of false decisions is very small, we let these correctthemselves in the next coarse adaptation period.

In summary, we find that the joint time-frequency adaptationprocess in RADION yields quick and accurate convergence ateach femtocell to the common reuse zone as well as to provideorthogonal sub-channels in the isolation zone.

5.1.4 Network Dynamics (Single contention domain)We now evaluate RADIONs ability to adapt to networkchanges. First, we consider a single contention domain whereall BSs cause interference to each other. In the clique topologyof three cells, RADIONs adaptation algorithm allows each BSto detect its fair share and converge to orthogonal sets of 10sub-channels in the isolation zone. The frequency convergencepattern is shown in Fig.11, where frequency resources areprobed at the granularity of five sub-channel chunks. RA-DIONs resource expansion mechanism allows each BS toexpand its frequency resources in the isolation zone if it probesempty resources for more than a coarse adaptation period.When BS3 is switched off at about 210 seconds (indicated bythe arrow), its sub-channels 11-20 become free. BS1 probesthe availability of chunks 11-15 and 16-20 but decides toexpand its resources only to 1-15, to allow fair access to othercells in the contention domain. However, since the chunk 16-20 remains available in the next adaptation period as well, BS1continues to expand its resources to 1-20. BS2 fails to grab the




1-5

6-10

11-15

16-20

21-25

26-30

0 100 200 300 400 500

Su

b-c

ha

nn

els

Time (sec)

BS 1

BS 2

BS 3

Fig. 11. Frequency convergence (single contention).

chunk 16-20 due to an inaccurate BDR estimate. Thus, whileRADION paves the way for a distributed fair sharing of unusedresources, distributed operations without information sharingprevents it from controlling the utilization-fairness tradeoff inthe best way possible.

5.1.5 Network Dynamics (Multiple contention domain)

Here we evaluate the adaptiveness in a multiple contentiondomain by creating chain topology BS3-BS1-BS2; the idealfair shares of all BSs are 15 sub-channels each. However, sinceBS1 does not have global information, it computes its fairshare to be 10 sub-channels (based on two interfering neigh-bors) without realizing that its neighbors belong to differentcontention domains. BS2 and BS3 compute their fair sharesas 15 sub-channels each. RADIONs resource expansion canhelp BS1 salvage under-utilized resources. The convergenceis shown in Fig.12. Initially, while BS2 and BS3 convergeto operate on sub-channels from 16-30, BS1 converges tosub-channels 1-10. RADIONs resource expansion/contractionfeature is enabled after 200 seconds. At this point, BS1 probesthe available chunk from 11-15 and expands its allocationto 1-15; this remains stable till 370 secs. At 370 secs, BS2expands its allocation to 11-30 due to an inaccurate probing,which is immediately rectified in the next adaptation. Thisagain happens at around 510 secs. However, this time, beforeBS2 rectifies its decision, BS1 responds to interference bycontracting its allocation back to its initial fair share of 1-10.This in turn prompts BS3 to probe and expand its allocationto 11-30 (similar to BS2).

The above experiment demonstrates that when a new BSjoins the contention domain, it is detected by other cells in thedomain; these BSs update (contract) their fair share and runthe adaptation process to determine their isolated resources.Stated otherwise, it succinctly captures both the expansionand contraction features incorporated into RADION to tracknetwork changes. It also indicates the transition between a fairallocation and one with high utilization. However, to finelycontrol such transitions, information exchange across multiplecells is required, which may not be feasible in residential envi-ronments. RADIONs best-effort utilization-fairness tradeoffsare particularly suited for such environments.

1-5

6-10

11-15

16-20

21-25

26-30

0 100 200 300 400 500 600

Su

b-c

ha

nn

els

Time (sec)

BS 1

BS 2

BS 3

Fig. 12. Frequency convergence (multiple contention).

5.2 Evaluation through SimulationsThe purpose of simulations is to help evaluate the effectivenessof RADION in a large-scale network (severe interferenceand complex topologies). Given that probing chunks in thefrequency domain (sub-channel selection) indirectly controlsconvergence in the time domain, here we only focus on conflictresolution (i.e., probing) in the frequency domain. We increasethe number of BSs to stress test RADIONs convergence inthe frequency domain.Simulation set-up: We consider two versions of coarse adap-tation: (i) a BS picks a fixed prime number of frames Pfrom [8000, 16000], and (ii) P is varied (randomly) acrossadaptation periods. A larger range is chosen for P to allowmultiple cells to pick the prime periods without inducingexcessive probing collisions.

Fine adaptation period (q) is drawn from the range [n s,600], where n is the number of BSs and s is the number ofprobes sent on each chunk (set to 10 frames in the simulation).The variable range is based on the observation that smallerranges are unfair to the case where large number of BSssimultaneously probe and adapt. In addition, when we evaluatefine adaptation in isolation, we simulate by changing themaximum period (max. q) to 600, 1600, 2600 and 3600 (min.period of n s is still common). We use the same evaluationmetrics as in the prototype evaluation and simulate by varyingthe number of BSs that simultaneously sample chunks anddecide on the chunk of operation. Each measurement is theaverage of 25 randomly generated parameters (P and q).Two-phase adaptation: It is seen from Fig. 13(a) and 13(b)that RADIONs two-phase adaptation results in a lower frac-tional collisions than coarse (both fixed and varied) and fineadaptation in isolation. These results clearly corroborate ourfindings from the prototype evaluation, thereby demonstratingRADIONs scalability in dense deployments. In addition, theGreedy selection outperforms Gibbs sampling. Similar to theprototype evaluation, the reason behind this is the fact thatsince the samples are actively taken (by transmitting bursts)from each chunk, they may not be representative of the state(occupied vs. free) of each chunk due to sampling collisions.This coupled with the probabilistic nature of the frequencyselection incurs a higher fraction of false decisions (plotted inFig. 13(c) and 13(d)) and hence a larger number of collisions.Probing duration: While fewer false decisions are incurredwith coarse adaptations for Greedy selection (Fig. 13(c)),




0

0.05

0.1

0.15

0.2

0.25

5BS 7BS 10BS 12BS

Colli

sion T

ime F

ract

ion

Fixed CoarseTwo-Phase

Varied Coarse

(a) Greedy selection

0

0.05

0.1

0.15

0.2

0.25

5BS 7BS 10BS 12BS

Colli

sion T

ime F

ract

ion


Varied Coarse

(b) Gibbs selection

0

0.003

0.006

0.009

0.012

0.015

5BS 7BS 10BS 12BS

Fals

e D

eci

sion F

ract

ion


Varied Coarse

(c) Greedy selection

0

0.006

0.012

0.018

0.024

0.03

5BS 7BS 10BS 12BS

Fals

e D

eci

sion F

ract

ion


Varied Coarse

(d) Gibbs selection

Fig. 13. Greedy selection outperforms Gibbs selection in collision and false decision. RADIONs two-phase adaptationprovides a significant reduction in the collision time.

0

0.2

0.4

0.6

0.8

1

A B C D A B C D A B C D A B C D

Co

llisi

on

Tim

e F

ract

ion

Max q: A=600, B=1600, C=2600, D=3600

5 BS 7 BS 10 BS 12 BS

GreedyGibbs

(a) Collision time fraction

0

0.05

0.1

0.15

0.2

0.25

0.3


Fa

lse

De

cisi

on

Fra

ctio

n

Max q: A=600, B=1600, C=2600, D=3600

5 BS 7 BS 10 BS 12 BSGreedy

Gibbs

(b) False decision fraction

0

200

400

600

800

1000

1200

1400


Co

llie

d F

ram

es

pe

r F

als

e D

eci

sio

n

Max q: A=600, B=1600, C=2600, D=3600

5 BS 7 BS 10 BS 12 BS

(c) Number of collied frames per false decision

Fig. 14. Results for the fine adaptation process. Fine adaptation is simulated with different period of range (max q).

the collision duration per false decision is longer than two-phase adaptation. This is mainly attributed to the fact that theutilizing a fine adaptation period reduces the observation timebefore sampling and making a decision. When a BS observesthe state of its own chunk (i.e., not sampling), other BSs alsoprobe the same chunk and result in collisions. If the obser-vation duration is not long enough, these collisions result inan inaccurate BDR average for the current chunk. This resultsin a few back-to-back false decisions until the BS obtains anaccurate measurement from each chunk. However, we showthat even with this characteristic, two-phase adaptation stillprovides significant benefits of reduced fractional collisions.The importance of the length of observation duration can alsobe seen from Fig. 14, where a larger range helps in reducingfractional collisions for fine adaptation in isolation. Note thateven with larger ranges, fine adaptation alone performs worsethan RADIONs adaptation.

The contrary is true with fine adaptation (smaller collisionduration per false decision but a higher number of falsedecisions) as seen in Fig.14. RADION combines the best ofcoarse and fine adaptation and provides a significant reductionin the time spent in collisions. As with prototype results,the greedy approach to sub-channel selection outperformsGibbs sampling; increased probing collisions and hence, falsedecisions are seen with Gibbs sampling. Fig.14 shows theimpact of qs range on fine adaptation with respect to thenumber of BSs. Increasing the range of q provides more timefor estimating BDR on the current frequency chunk. Hence,false decisions due to inaccurate BDR estimates and thus,collision durations are reduced. However, adapting qs rangeis a double-edged sword. While it results in less number offalse decisions, it incurs a higher penalty of a false decisionsince the probing collisions cannot completely be eliminated.In practice, the design of such a system that eliminates probing

0

3000

6000

9000

12000

5BS 7BS 10BS 12BS

Colli

ded F

ram

es

per

Fla

se D

eci

sion


Varied Coarse

(a) Greedy Selection

0

3000

6000

9000

12000

5BS 7BS 10BS 12BS

Colli

ded F

ram

es

per

Fals

e D

eci

sion


Varied Coarse

(b) Gibbs selection

Fig. 15. Two-phase adaptation very quickly corrects falsedecision, and hence minimizes the collisions of operatingon the same frame resources.

collisions is hard (without coordination) due to the asynchronyof decisions among BSs. We also show that fine adaptationalone does not perform well since it incorporates a shortobservation duration which results in many wrong decisions.Hence, adapting qs range alone is not sufficient; the two-phaseprocess as in RADION is needed.False decision: In Fig. 15, we plot the number of collidedframes per false decision. We observe that the penalty of afalse decision is significantly smaller in two-phase adaptationdue to fine adaptation periods that are more reactive thancoarse adaptation. On the other hand, for coarse adaptations(both fixed and varied), the penalty of a false decision istypically on the order of a P value (there is negligible variationbetween Greedy and Gibbs). This means that once a falsedecision is made, it is corrected only until the next coarseadaptation period of another BS which is chosen among largeprimes to address sampling collisions.Convergence: We evaluate the convergence of our algorithmwith large-scale network (12 BSs). The time duration whenmore than two BSs operating on the same frequency resourcesis about only 30 seconds during 100 minutes simulation. Inother words, 12 BSs stay on orthogonal resources (converged)




0

0.2

0.4

0.6

0.8

1

0 3000 6000 9000 12000 15000

Col

lisio

n Pe

rcen

tage

Time (sec)

Fig. 16. Total collision duration is less than 0.7%.

99.5% of time. We also measure how quickly BS converges tothe orthogonal resources after collisions and the average timetakes to correct wrong decision is 1.62 seconds for two-phaseadaptation while coarse adaption takes 50.68 seconds. Theseresults show quick convergence and correction of algorithm.Stress test: Here, we perform a stress testing of RADIONsfrequency selection in the 10 BS scenario and present anexample convergence trace. Initially, all BSs start on the samefrequency chunk and try converging on orthogonal chunks. If,in a frame, at least two BSs use the same chunk, we markthis frame as a collision frame. We then calculate the numberof such frames in a second and get the ratio over 200 frames(in WiMAX, 200 frames are transmitted in a second). Thesimulation is run for 3,000,000 frames (15,000 seconds). Fig.16 plots the collision percentage; spikes indicate that some BSmade a wrong decision and resulted in collisions. A collisionpercentage of 0 means that all BSs operate on orthogonalchunks at that particular instant. It is seen that with RADION,BSs stay on orthogonal chunks with very short interruptionsdue to false decisions (resolved quickly with fine adaptation).We observed that the total collision duration due to falsedecisions was 100 seconds - less than 0.7% of the totalactive time of the BSs.Comparison to FERMI: We compare the performance ofRADION against other resource management schemes, e.g.,FERMI and UseAll. Again, we only compare the frequencyresource allocation since FERMI and RADION adapt dif-ferent reuse zone convergence strategies. Our purpose is tounderstand the relative sub-optimality introduced due to thelack of cooperation across BSs in RADION. In FERMI, acentral controller generates an interference map between BSsand assigns frequency resource to BSs using clique basedallocation [13]. On the other hand, each BS in RADION deter-mines the frequency allocation based solely on its neighboring(interfering) cells. For the baseline strategy, we introduce theUseAll scheme where each BS uses all available resources in agreedy manner. Fig. 17 shows the CDF of throughput of eachBS (simulated with 40 BSs). The median throughputs are 1.4 /4.1 / 5.1 Mbps for UseAll, RADION and FERMI, respectively.The performance of RADION is superior to UseAll and closeto that of FERMI. Since FERMI constructs the global view ofinterference relations across the BSs, it optimally allocates theresources (maximizing frequency reuse), and hence providesthe best performance. However, RADION still provides 36%

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

CD

F

Throughput (Mbps)

FERMIRADION

UseAll

Fig. 17. Performance comparison.

higher throughput (on aggregate) compared to the baselinestrategy (without the inter-cell coordination) without incurringthe central computation overhead of FERMI, thereby balanc-ing both performance and overhead.

6 CONCLUSIONWe design and implement RADION, arguably the first selforganizing resource management framework for OFDMA fem-tocell networks. RADION consists of four key building blocks,client throughput estimation, client categorization, resourcedecoupling and two-phase adaptation and allocation. RA-DION allows appropriately chosen clients to opportunisticallyreuse the spectrum while isolating resources for the otherclients in a distributed way. We implement RADION using aWiMAX testbed to show its quick convergence to an efficientresource allocation in real settings. We also demonstrate thescalability and efficacy of RADION in larger scale settingswith simulations. We only consider downlink performance,however, a similar approach can be applied to the uplink.Improving fairness and utilization via increased informationexchange and the impact of power control are features thatwe plan to investigate in the future.

ACKNOWLEDGMENTA preliminary version of this paper appeared at ACM Mobi-Hoc 2012 [32].

REFERENCES[1] Cisco Visual Networking Index: Global Mobile Data Traffic Forecast

Update, 2011-2016, Feb 2012.[2] R. Van Nee and R. Prasad, OFDM for Wireless Multimedia Communi-

cations, Artech House, 2000.[3] J. G. Andrews, H. Claussen, M. Dohler, S. Rangan and M. C. Reed,

Femtocells: Past, Present, and Future, IEEE Journal on Selected Areasin Communications, vol. 30, no. 3, pp. 497-508, 2012.

[4] H. Ekstrom, A. Furuskar, J. Karlsson, M. Meyer, S. Parkvall, J. Torsner,and M. Wahlqvist, Technical Solutions for the 3G Long Term Evolu-tion, IEEE Communications Magazine, vol. 44, no. 3, pp. 38-45, 2006.

[5] S. Yeh, S. Talwar, S. Lee, and H. Kim, WiMAX Femtocells: APerspective on Network Architecture, Capacity, and Coverage, IEEECommunications Magazine, vol. 46, no. 10, pp. 58-65, 2008.

[6] R. Kim, J. S. Kwak, and K. Etemad, WiMAX Femtocell: Requirements,Challenges, and Solutions, IEEE Communications Magazine, vol. 47, no.9, pp. 84-91, Sep. 2009.

[7] V. Chandrasekhar, J. G. Andrews, and A. Gatherer, Femtocell Networks:A Survey, IEEE Comm. Magazine, vol. 46, no. 9, pp. 59-67, Sep. 2008.

1536-1233 (

ieee transactions on mobile computing 1 self … ieee papers/ns2/ns2... · self-organizing resource...

Documents