Attribute-based bandwidth reservation scheme for mobile multimedia wireless networks

2005InternationalConferenceonWirelessNetworks, Communications and MobileComputing

Attribute-Based

Bandwidth Reservation Scheme for

Mobile

Multimedia

Wireless

Networks

Jau-Yang

Changf t and

Hsing-Lung

Chent

§ Department of

Electronic

Engineering

Chinmin

Institute

ofTechnology

Toufen,

Taiwan, R.

O. C.

t

Department

ofElectronic

Engineering

National

Taiwan

University

of

Science

and

Technology

Taipei, Taiwan, R.O.C.

jychang@ms.

chinmin. edu.

tw,

hlchen@et.

ntust.

edu.

tw

Abstract

In ordertoprovide seamless muiltimediaservices inthe next-generation wireless networks, efficient bandwidth reservation scheme must bedeveloped. A novel attribute-based bandwidth reservation scheme isproposed in this

paper as a solution tosupport quiality-of-service

guaran-tees ( oS) in the mobile multimedia wireless networks. Based on the existing network conditions, theproposed

scheme makes an adaptive decisionjbr bandwidth

reser-vation andcall admission in each base station (BS). We

use the dynamically adaptive approaches to reduce the

connection-blocking probability (CBP) and connection-droppingprobability (CDP), and to increase the

band-width utilization (BU) formobile multimediawireless

net-works. Simulation results show thatouirproposedscheme

outperjbrm the previouslv proposed scheme in terms of

CBP, CDP, and BUinmobile communicationsystems.

1.

Introduction

Providingmultimedia services with aQoS guarantee in

mobile wireless networks presents morechallenges dueto

user's mobility and limited bandwidth resource. Since

bandwidthis themost criticalresourceinmobile

multime-dia wireless networks, it is important to employ mecha-nisms forefficiently using theavailable bandwidth. In

mo-bile cellularcommunication networks carrying multimedia traffic,it becomesnecessarytoprovide QoSguaranteesfor multimediatraffic connections. Inordertoutilize the radio

spectrum efficiently, microcellular and picocellular

net-works are deployed. These networks have the inherent

problem of rapid handoffs duetosmallercoverage areasof

BSs. This problem leads to higher CDP and results in bandwidth resource availability varying repeatedly.

Fre-quentchanges inthenetworktraffic make theprovisionof guaranteed QoS more difficult. Reserving bandwidth for

handoff mobile hosts (MHs) can reduce CDP in wireless

networks. However, itmaylead tohigh CBP fornew con-nections and low BU. Hence, research inthe area ofQoS provisioning in the next-generation wireless networks fo-cuses onthe

integration

ofresourcereservation and admis-sioncontrolpolicies

[1]-[6].

Recently, several bandwidth reservation schemes have

been proposed to support QoS provisioning in wireless

environments. In[2], the shadow clusterconcept is usedto predict

future

resource

allocation

andto

perform

bandwidth reservation inwireless networks. Inthis scheme, the home BS ofthe MH informs the BSs of its shadow cluster area about the MH's requirement, position, and movement pa-rameters. Itrequires the exchange ofalarge amount of de-tailed knowledge ofMHs between the BSs. This scheme increasestoo manyoverheads between the BSs,especially in microcellular and picocellular wireless systems. More-over, it isdifficult to accurately estimate future movement pattems in real wireless environments. In [3],the adaptive bandwidth reservation scheme providesQoS guaranteesby reservingbandwidth in allneighboring cells.Inthis scheme, ifanyreservation in neighboring cells fails, the new con-nection and handoff concon-nection are rejected. It increases CBP and CDP.

Moreover,

bandwidth is reserved redun-dantly because the MH moves to only one ofneighboring cells. It decreases the BU. Anadaptive resource allocation for multimedia QoS management in wireless networks is proposed in [1] to improve CBP, CDP, and BU by using thedynamic guard-channel scheme. However, this scheme

does not appropriately adjust the reserved bandwidth for

each BS in bursts of traffic conditions. It reduces the BU and increases theCDP.

Thispaperintroducesanovel attribute-based bandwidth reservation scheme that supports QoS guarantees in the next generation mobile multimedia wireless networks. In theproposedscheme, each BSmakes anadaptive decision for bandwidth reservation by employing attribute-measurement mechanism and traffic-based reservation strategy. Theamountofreservedbandwidthforeach BS is dynamically adjusted, according to the on-line traffic

in-formation of each BS. The traffic-based reservation strat-egymakes the systemmoreresponsivetothecurrenttraffic conditions, providing efficient bandwidth reservation for thehandoffconnections. The control of bandwidth reserva-tion is adaptive under overloading traffic conditions, thus caneffectively deal with sudden traffic surges. Multimedia traffic is acombination of both real-time andnonreal-time traffic. Real-time traffic includes voice and video while data and graphics makeupnonereal-time traffic.Wedesign efficient calladmissioncontrolalgorithms for real-time and nonreal-time traffic, according to the different multimedia services. On-line bandwidth reservation scheme does not requireprediction offuture traffic intensity anduser mobil-ity. This approach has significantly lower complexityand system overhead than mobilityoriented methods. Our pro-posed scheme is an adaptive and on-line mechanism. It only requires minimal computational overhead ineachBS, and nocommunication overhead betweenneighboring BSs torequestand release reservations.

The remainder of this paper is organized as follows. Section 2 describes our assumed model ofwireless net-works. In Section 3, we illustrate the proposed scheme in detail. In Section 4, we present our simulation model and analyze the comparative evaluationresults oftheproposed schemethrough simulations. Finally, Section5offerssome

concludingremarks.

2.

System

Model

We consider a multimedia mobile wireless network withacellular network infrastructure. Eachcell containsa BS that needstoallocate andreserve bandwidthfor MHs, and communicates withotherBSs. BSsare connected to a switching networkthrough wired lines. Two typesof con-nectionsshare the bandwidth ofthe BS for each cell: new connections and handoff connections. A BS in a cellular network may receive new connections ofMHs within its cell andhandoffconnectionsofMHs fromtheneighboring cells. Whenaconnection arrives at acell where the band-width is not available, this situation is called connection blocking for new connections or connection dropping for handoff connections. For amobile user,droppingan ongo-ing connection is generallymoreunacceptable than block-ing a new connection request. Therefore, reducing the CDP is usually a main objective in the mobile wireless communication systems. The connectionblockingrateand connection droppingrate aretwo most significant connec-tion-level QoSparameters inthe mobilewirelessnetworks

[1], [5],

[6].

Intheproposedscheme, itis assumed that whenanMH requestsa newconnection in thecurrentcellormoves into the neighboring cell, the following parameters are pro-vided: 1) The traffic class (I or [I). 2) The required band-width for the connection. 3) Theminimum required band-width for the connection. The minimum required band-widthconnection is the smallest amount of bandwidth that the connection can be assured of a certainacceptable QoS, e.g.,the smallestencodingrateofits codec [3].

Hand-out connections MHO0 MHO, MH, t Hand-in connections

Figure 1. Hand-in andhand-out connections within the time interval from 1 to inthe cell .

3.

Proposed Bandwidth

Reservation Scheme

Bandwidth reservationis one ofthecritical elements in ensuring the QoS in mobile wireless networks and thus, requiresmechanisms to efficiently usethe available band-width. The proposed scheme combines attribute-measurement mechanism, traffic-based reservation strat-egy, and call admission control algorithms to guarantee QoS requirements in mobile wireless environments. The attribute-measurement mechanism is employedto measure the traffic information. Thetraffic-based reservation strat-egy is employed to calculate the amount of reserved bandwidth. The call admission control algorithms are em-ployed tocontrol whether the connectionscanbe served or not. These mechanisms are executed ateach BS in a dis-tributed manner.

3.1

Attribute-measurement

mechanism

In order to implement the attribute-measurement mechanism, we monitor several traffic parameters within each time interval in the cell. Intheproposed scheme, we classifyall datatraffic intoreal-time (Class I) and nonreal-time (Class II) according to the different multimedia services. Wedefine the hand-in and hand-out connections for Class I (or Class II) data traffic within each time interval in the cell. The connection for an incoming handoff MH in the cell is called hand-in connection and the connection for an outgoing handoffMH in the cell is called hand-out connection. Figure 1 shows the hand-in and hand-out connections within the time interval from to-

itO

toto in cellj,where

to

is the current time and

It,

is the length oftime interval at time to, and

MH,

denotes the rth hand-in connection and MH,, denotes the sth hand-out connection. Let

BIk1

and

Bbfk,

respectively, be thesum ofrequiredbandwidthsof hand-in connections and the sum of released bandwidths of hand-out connections for the class k data traffic at time tin cell j, where t is measured time from t, - Ik j to

t,

The 1k

denotesthe lengthoftime interval for class k data traffic at time t,, in cellj. Let

ak

denote the bandwidth rate of hand-in and hand-outconnections for class k data trafficat timet, in

cellj.

The

ak;

canbeexpressedby

t=,,,-1 kt,=t,,-l, j t=t,-I_1k a,1 = t,, t=t,, _1k III'i max.

I

BItfj

, t= .k . (1) BOk.j ,.1J

The parameter ak represents bandwidth variation of

MHs moving into and departing from the cellj for the class k data traffic within thetime interval [t,-i j, t4] If

k

attmen

ha h smo

a,1,

j isapositive value, itmeans thatthe sumofrequired

bandwidths ofhand-inMHs is larger than thesum of re-leased bandwidths of hand-out MHs for the class k data traffic within the timeinterval

[t,

-

Itfj,

t4] in

cellj.

kk

Let

HIt,j

and

HOtfj,

respectively, be the number of hand-in connections and the number ofhand-out connec-tions for the class kdata trafficattimetin

cellj.

Letk

denote the numeral rate ofhand-in and hand-out connec-tions for class k data traffic at time t, in cellj. The

kSf1

canbeexpressedby A

'n",i

E HIft,j- HOtj t=t,,-1k j =t,,-lk j ,a-lij_tnI Hj _I,,.

max. L

HIf,,

>,

HOtfj

t=t _l

.

t=t,

_lk

The parameter

,f;,j

represents variation of MHs mov-ing into anddeparting from the

cellj

within thetime inter-val [t,-

Itf,j,

t]. If

,8fj

is anegative value, itmeansthat the number ofhand-out MHs is largerthan the numberof hand-inMHswithinthe timeinterval

[t,

- If

j,

tn] in

cellj.

Let DI,1 denote the number ofdropping connections for the class k data trafficattimetin

cellj.

Let

7tf;

j be the

CDPofclassk datatrafficattime t,in

cellj,

whichcanbe expressedby k

)"YI,,,

E

DIt1j

tnl.1 t=, Hik ,t, t=t,- Ik

Inorder tomake thesystemresponsivetothe burstsof

k

traffic

conditions,

the

length

of time interval

[t,

-

It{

"j

4,]

ischosen basedon the variation of CDP for each time in-terval. Let k,/ denote variation of

Yt;

,_j

and Yt;j for class k data traffic at time t, in cellj, which can be ex-pressedby k

5k _

Yt,,,

j

5t,,,i

j k t,,, ,,j

(4)

The

length

of time interval

[t,

- Ik j,

t,]

in

cellj

is

dy-namically adjusted

according to ,L; . The value of

It

canbecalculated

by

wk jkk g by

whereuiisanunittimeand

isk

gienb

[ t;i l gint="k (5) if

.t;

>1 (6) ifdk 0

The

a,,,,

flt"j

and

yf,

j highlyaffect theoperationof

a reservation system in mobile communication networks.

k k ~~~~k

We define atk j,

8tL;

, and y,,'j as

reservation-factors

for class k data traffic at time t, in cellj. When ak > 0

(or

<0)and

h;f,,

>0(or>0), itindicates under the

heavy

(or light)

traffic load for class k data traffic at time tn in

cellj. On the otherhand, When af,,j >0 and

L;J,j

<0,or

at,

f <0 and

it,fj

> 0, it represents less traffic fluctua-tions for class k data traffic at time tn in

cellj.

According

k k

to ytl j , sft j can be calculated which

represents

the

variation of CDP for class k data traffic attime tnin cellj.

When

,5t

j >1 (or<1), itmeans that the CDP for class k data trafficisanincreased(or decreased)trendattime tn in

cell j. Hence, for efficiently adjusted the

reservation-factors and bandwidth reservation, the value of

,k

should be increased (or decreased). The initial value of

It&;j

and u may be chosen basedon the real operation of

k

system

(e.g.,

It'

j = 60 sec and u= 10

sec).

The

reserva-tion-factors bring an immediate response on the traffic conditions. They not only make the system responsive to

the bursts of trafficconditions, but also get thebandwidth reservationmoreefficiently.

3.2 Trafric-based reservation strategy

The traffic-based reservation strategy is employed to calculate the amount of reserved bandwidth based on the reservation-factors and networkconditions. Thebasic idea of this strategy is that a reservation system with lower complexitydoesnotrequireadvance knowledgeor

predic-tion offuturetraffic intensityand usermobility. This strat-egy makes lower system overhead and causescontrol deci-sions ofbandwidth reservation in realtime. In this frame-work, the BSs record the requested bandwidths ofMHs within each time interval for each traffic class. Let R'_I,,+1

denotethe amountofnewreservedbandwidthfor the next

time and Rk the amount of reserved bandwidth at the

currenttimet,ofor the class k data traffic in the BS ofcellj.

Let B,,k is the amount of hand-in requested bandwidths 1..I...

for class k data traffic within the time interval win theBS

of

cellj,

which can beexpressed by

BW;j

E

BItj

(7)

k,

t=tf}Ie-f..i

Then, the Rk can becalculatedby

RJtl+I

2

jRill

+ 2

BWkj

(8)

where £ is the explicit reservation-factor

calculatedby which can be

to''

+

fl,1.

(9)

The explicit reservation-factor controls the relative weights for therecent andpast traffic histories. If the mo-bility ofMHs is uniformly distributed, the new reserved bandwidth for thenext time in theBS of celljcanrelyon the current reserved bandwidth

(

k ). However, when the

traffic distribution is non-uniform, due to the temporal traffic fluctuations, the new reserved bandwidth for the next time in the BS of cellj should depend on recent re-quested bandwidths of hand-in connections (

Bk

j). We

decide the value of e. byconsidering thecurrentnetwork traffic conditions. Hence the value of reservation-factor

k canaffect the networkperformance significantly.

For efficient BU, the control decisions of bandwidth reservation must be dynamically adjustable. In our pro-posed scheme, we take into account the existing network conditions. Theamountof reservedbandwidth for eachBS is dynamically adjusted, according to the on-line traffic information of each BS. It reserves bandwidth only when necessary. So the BSs make bandwidth reservation effi-ciently. To provide QoS guarantees, if the reserved band-width for the Class I is not sufficient, our scheme allows the use of the bandwidth from the reserved bandwidth for Class IL. The traffic-based reservation strategy makes the system more responsive to the current traffic conditions, providing efficient bandwidth reservation for the handoff connections.Therefore, theCDP canbe reduced.

3.3 Call Admission

Control

Algorithms

When MH xrequests a new connection in the cell

bj,

the BS of cell

bj

tries to admit this request by using the available bandwidth. The new connection ofMH x is re-jected in the cell

bj

ifits required bandwidth is large than the unusedbandwidth. Otherwise, the BS of cell

bj

exam-ines the reservation factors, according to the attribute-measurement mechanism. If there are high reservation-factors (e.g.,

cabj

- +£bj - >0.6) inthe BSof cell bj, then the new connection will be blocked. Otherwise, the BS of cell

bj

accepts theconnection and allocates the cor-responding bandwidthfor the newconnection ofMHx.

When MHx moves fromcell bjto the cellbk, the BS

ofcell

bj

releases allocatedbandwidth of the MH x. There

aretwoclasses ofthe call admission control algorithms for handoff connection in the cell bk, according to the traffic class of MHx.Ifthe handoff connection of MHxis a real-time data traffic, the BS of cell bk checks whether the available bandwidth plus total reserved bandwidth RRbk'S_bk I + R_b~kcls _{)i ufceto}is sufficient or not._{o.I h}Ifthe mini-_ii

mumrequired bandwidth for handoff connection ofMH x is not sufficient, the BS of cell bk drops the handoff con-nectionofMHx. Otherwise, the BSof cell bk accepts the connectionand allocates the corresponding bandwidth for the handoffconnectionof MH x.

If the handoff connection of MH x is a nonreal-time data traffic, it is dropped only when there is no residual bandwidth for the MH x in theBSof cell bk. Ifthereis not any available bandwidth and reserved bandwidth

(Rbkas H )for class I1 traffic in theBS of cellbk,the BS

of cell bk drops the handoffconnection ofMH x. Other-wise, the BS of cell bk accepts the connection and allo-catesthecorresponding bandwidth for the handoff connec-tion ofMHx.

4.

Performance

Analysis

Inthis section, we presentperformance analysis for the proposed scheme. We describe our simulation model and illustrate the simulation results, comparing our scheme with thetraditionally non-reserved bandwidth mechanism. Theperformance ofourproposedscheme is also compared anexisting adaptiveresourceallocation scheme [I]. Inour simulation modelwehavethefollowing assumptions.

1)The simulation environment is composed of 100 cells and available bandwidth of each cell is 30 Mb/s, each cellkeeping contact withits sixneighboring cells. The distance betweentwoBSsis 1 km.

2) The arrival process fornewconnection requests is Pois-sondistribution withrateX, which is uniform inallcells. 3) Three cases of mobility are considered: MHs move

hardly, in which case the speed of MHs are uniformly distributed between0 and 5 kmlh; MHs move slowly, in which case the speed of MHs are uniformly distrib-uted between6 and 40km/h;andMHs move rapidly,in whichcase the speed ofMHsare uniformly distributed between41 and90km/h.

4) Six different traffic types are assumed based on band-widthrequirement and traffic class showninTable 1 [3]. 5) The durations of connections are assumed to follow a

geometric distribution with different means for different multimedia traffic types [3].

WecomparetheCBPs, CDPs,andBUs of three mecha-nisms viasimulations. Figure2 showsthe CBP. TheCBPs increase in allmechanisms withincrease inthe arrivalrate of new connection request. The unsuitable reservation scheme results in the most ofconnection blocking. It re-serves too many bandwidths for the handoff connection. Our proposed scheme can predict that a little adjustment for bandwidth reservation is enough, resulting in lower CBP. AccordingtotheFigure 2, whenthe connection

arri-I

Table 1. Multimedia traffic used in simulation mode

Traffic Bandwidth Average Connection Average

Application Class Requirement Bandwidth Duration Connection Example

Requirement Duration

I Real-time (CBR) 60- 600 seconds 180 seconds

phone

2 Real-time 256_(CBR)Kb/s 60-1800seconds 300 seconds Video-phone & Video-_conference

1-6Mb/s (average) Interact.Multimedia

3 Real-time 2.5-9 Mb/s (peak) 3Mb/s 300-18000seconds 600seconds &VideoonDemand

(VBR)

4 Nonreal-time_{Nonrea-time(UBR)}5-20Kb/s 10Kb/s 10-120seconds 30seconds Email, Paging&Fax

5 Nonreal-time 64-512Kb/s 256Kb/s 30-36000 seconds 180 seconds R t L n& D t

1-10Mb/s File Transfer& Retrieval

6 Nonreal-time (UBR) 5Mb/s 30- 1200seconds 120 seconds Service

0.4 -- 0.35 _ C 0.3 0l.2 ._ 0.25 o0 o 0.1 o. Q .05 0.05 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 O. 0.9

Connection arrivalrate(requests/sec)

Figure2. Connection-blocking probability (CBP).

0.1 .-0 O.I 0. 3 0.1 0.0 o 0.0 . = O C-0.(: 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Connectionarrivalrate(reqitests/sec)

Figure 3. Connection-dropping probability (CDP). valrateislarger than 0.1 requests/sec,itis evident that the CBPs ofnon-reserved bandwidth mechanismandour

pro-posedschemearebetterthan that of the adaptiveresource

allocation scheme. Figure 3 shows that the CDP ofour

proposed scheme is better than that of the adaptive

re-source allocation scheme and non-reserved bandwidth

1 0.9 0.S - 0.7 .21'_C:) 0.6 1E 0.4 -,: M 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.9

Connection arrival rate(requests/sec)

Figure4. Bandwidthutilization (BU).

mechanism. Furthermore, as the connection arrival rate

increases,it isobvious that the proposed schemeresults in

lowerCDP. Forexample, when the connectionarrival rate is 0.5requests/sec, the proposed scheme iseasy to achieve a reduction in percentage of dropped connections from 15% to 6%, which is compared with the non-reserved bandwidth mechanism. Furthermore, in this case, the pro-posed scheme achieves the CDP that is approximately2% lower than the adaptive resource allocation scheme. The reason for this behavior is that the proposed scheme dy-namically adjusts the bandwidth reservation for each BS byusing traffic-based reservation strategy,providingmore precise bandwidth reservation andefficient call admission control for handoffconnections. Therefore, the CDP can beredtuced. Reduction of CDP andCBP arecontradictory. Hence, we must make a tradeoff between the BU and bandwidth reservation such that the CDP forhandoff con-nections can be reduce, while the CBP for new connec-tionscanbemaintainedatanacceptable level. Via simula-tions, we show that our proposed scheme reaches these significant objectives.

Figure 4 shows the BU. The non-reserved bandwidth mechanism results in higher BU than adaptive resource allocation scheme and our proposed scheme. This is

be-,,

...

. . -- -- --Non-reserved - - - Adaptive Proposed 180 - --Non-reserved I.. Adaptive_ _ Proposed 4 1 12 )4 / )2 f 0

cause the non-reserved bandwidth mechanism does not reserve any bandwidth for handoffconnections, and thus nobandwidth is wasted. The adaptive resource allocation scheme is suitable for the uniformly traffic conditions. However, when the traffic conditions are bursts, this scheme does not appropriately adjust the reserved band-width foreach BS. Hence the bandwidth is wasted. In con-trast, our proposed scheme adjusts bandwidth reservation byusingtheattribute-measurementmechanism and traffic-based reservation strategy, leading to handling bursts of traffic conditions efficiently. The amount of reserved bandwidth for eachBS isdynamically adjusted, according to the on-line traffic information of each BS. It reserves bandwidth only when necessary, resulting in higher BU. According to the Figure 4, when the connection arrival rate is lower than 0.1 requests/sec, the BUs of three mechanisms are similar. However, when the connection arrival rate increases, the BU of our proposed scheme is close to that of the non-reserved bandwidth mechanism and superior to that of the adaptive resource allocation scheme.

5.

Conclusion

The majoradvantage of this paper is that our proposed scheme is successfully appliedto deal withbandwidth res-ervation problems and provides better QoS guarantees in mobile multimedia wireless networks. In the proposed scheme, the amount of reserved bandwidth for each BS is dynamically adjusted, according to the on-line traffic in-fonnation of eachBS.The simulation results show that the bandwidth reservation scheme using dynamically adaptive techniques improves CBP, CDP, and BU significantly. The

proposed scheme has significantly lower complexity and system overhead than mobility oriented methods. It not only outperforms the pervious scheme, but is also suitable for thenextgenerationmobile communicationsystems.

References

[1] L. Huang, S. Kumar, and C.-C. Jay Kuo, "Adaptive Resource Allocation for Multimedia QoS Management in wireless Networks,"IEEE Trans. Veh. Technol.,vol. 53, no.

2, pp. 547-558, March 2004.

[2] D. A. Levine, I. F. Akyildiz, and M. Naghshineh, "A Resource Estimation and Call Admission Algorithm for Wireless Multimedia Networks Using the Shadow Cluster Concept," IEEE/ACMTrans.Networking, vol. 5, no. 1, pp.

1-12, Feb. 1997.

[3] C. Oliverira, J. B. Kim, and T. Suda, "An Adaptive

BandwidthReservation Scheme forHigh-Speed Multimedia Wireless Networks," IEEE Select. Areas Commun., vol.

16,no.6, pp.858-874, Aug. 1998.

[4] J. Ye, J. Hou, and S. Papavassiliou, "A Comprehensive Resource Management Framework for Next Generation Wireless Networks,"IEEE Trans. Mobile Computing, vol. 1, no.4,pp.249-264, Oct.-Dec.2002.

[5] L.Badia,M.Zorzi, andA.Gazzini, "AModelfor Threshold Comparison Call Admission Control in Third Generation Cellular Systems," in Proc. IEEE Int. Conf: Communications(ICC'03),2003, pp. 1664-1668.

[6] J. Ye, J. Hou, and S. Papavassiliou, "Integration of Advanced Reservation and Bandwidth Reconfiguration Based Admission Control in Wireless Networks with Multimedia Services," in Proc. IEEE Int. Conf on Distributed Computing Systems Workshops (ICDCSW'03), 2003. pp. 844-849.