474 IEEE COMMUNICATIONS LETTERS, VOL. 12, NO. 6, JUNE 2008

A Divide-and-Conquer Technique forThroughput Enhancement of RFID Anti-collision Protocol

Jeong Geun Kim, Member, IEEE

Abstract— A novel anti-collision technique is proposed tomaximize identification performance in slotted Aloha basedRadio Frequency Identification (RFID) systems. We observe thatmuch higher throughput can be achieved by identifying tagsin a divide-and-conquer method, in which the set of tags ispartitioned into multiple subsets of roughly equal size and theneach subset is identified in sequence. Numerical results showthat the throughput performance of our proposal outperformsexisting methods by a significant margin.

Index Terms— RFID, anti-collision, partitioning.

I. INTRODUCTION

COLLISION in Radio Frequency Identification (RFID)systems refers to simultaneous responses of multiple

radio frequency tags to the reader’s inquiry. Collision incursa mixture of scattered waves, thereby prohibiting correctidentification by the reader. To cope with this problem a certainform of coordination protocol named anti-collision protocol isrequired between the reader and the tags [1][2].

Designing an efficient anti-collision protocol for theresource-constrained RFID system has become an active re-search topic [1]–[6]. It is well understood that throughput offramed slotted Aloha is maximized when the frame size (i.e.,the number of slots in a frame) is equal to the number oftags [1]. Since no information about the tag population isavailable a priori, the number of tags must continually be es-timated while initial rounds of slot allocation proceed blindly.In general, the time taken to arrive at a reasonably accurateestimate is a significant fraction of overall latency. One wayto solve this is to exercise the blind estimate for a smallerset of tags from which the size of the entire tag populationcan be easily inferred. Based on this idea, our algorithm firstblindly partitions the tags and probes the first subset of tags toestimate the tag population. Once the estimation is obtained,the remaining subsets of tags are repartitioned into an optimalnumber of subsets that yields the maximum throughput. Therationale behind this method is that higher throughput canbe obtained by performing identification for multiple smallersubsets of equal size instead of doing so for all tags at once.We refer to the throughput gain obtained by partitioning overnon-partitioning as the partitioning gain. After the optimumrepartitioning is finished, the identification process is thenperformed sequentially for each subgroup. In particular, theestimate of the number of tags in a subgroup becomes moreaccurate as more subgroups have been identified. By taking the

Manuscript received February 22, 2008. The associate editor coordinatingthe review of this letter and approving it for publication was F. Granelli. Thiswork was supported by Kyung Hee University.

J. G. Kim is with the Department of Electrical Engineering, Kyung HeeUniversity, Yongin, Korea (e-mail: jg kim@khu.ac.kr).

Digital Object Identifier 10.1109/LCOMM.2008.080277.

Slot 1

Tag ID

Frame 1 Frame 2 Frame 3 Frame n...

Slot n...Slot 4Slot 3Slot 2

Query QueryRep QueryRep

Tag ID

Tag ID

AckNack

QueryRep QueryRep

Select

Fig. 1. System model for performance evaluation.

average of the tag populations in subgroups that have alreadybeen processed, we can obtain a more accurate estimate of thesize of the next subgroup.

The benefits of our partitioning algorithm are twofold: fasterestimation on the tag population and higher throughput byexploiting the partitioning gain. Both benefits contribute to asignificant improvement of throughput. Extensive simulationsare performed with a discrete-event simulation library todemonstrate superiority of our algorithm over previous non-partitioning approaches.

II. SYSTEM MODEL

A system model for framed slotted Aloha is established hereto compare the performance of various anti-collision schemes.The system model is mostly based on the EPC standard [2],though some details are simplified for the sake of analysis. Asshown in Fig. 1, our system model begins the identificationprocess by sending the Select message to the tags. The roleof the Select message is to partition the tags into multiplesubgroups. The Query message following the Select messageinforms the tags of the frame size. Unlike the EPC standard,in which the only valid frame sizes are powers of two (i.e.,2Q, Q = 1, 2, · · · ), no such restriction exists in our systemmodel. In addition, the unique tag ID is directly transmittedto a chosen slot instead of using a short reservation packet,i.e., an RN16 message in the EPC standard. This simplificationis adopted for the sake of analysis.

III. PARTITIONING ALGORITHM

Our partitioning algorithm consists of three stages. In thefirst stage, an initial estimate of the tag population is made.In the second stage, the tags are repartitioned into an optimalnumber of subgroups based on this estimate. In the final stage,the remaining subgroups of tags are identified in sequence.

In the beginning of the first stage, all tags are blindlypartitioned into s subgroups, The initial value of s is set to 4.For the first subgroup, the identification process starts with

1089-7798/08$25.00 c© 2008 IEEE

KIM: A DIVIDE-AND-CONQUER TECHNIQUE FOR THROUGHPUT ENHANCEMENT OF RFID ANTI-COLLISION PROTOCOL 475

an initial frame size (n) of 16. Instead of completing theidentification for the first subgroup, we take an alternativeapproach in which the tags are continuously repartitioneduntil the first subgroup of tags is reliably being read withthe initial frame size of 16. Repartitioning occurs wheneverthe kth consecutive collision or empty slots are observed. Ofcourse, kth consecutive collisions lead to a smaller subgroupby doubling the number of subgroups (i.e., s = 2 ·s), whereaskth consecutive empty slots require a larger subgroup byhalving it (i.e., s = s/2). Once a frame is completed withoutrepartitioning, the next frame size is set to 2.39 · nc slots,where nc is the number of collision slots in a frame.

Once the identification of the first subgroup is completed,we have a rough estimate of the number of tags. This informa-tion is then utilized to repartition the remaining tags into anoptimal number of subgroups. The model used to determinethe optimum value was presented in [5].

After the optimal repartitioning, we have a rough estimateof the number of tags in each subgroup. When we begin theidentification of the second subgroup, the initial frame sizeis set according to the rough estimate, thereby yielding athroughput close to its possible maximum. In particular, theseestimates become more accurate as the remaining subgroupsare identified. Before identifying the ith subgroup, we alreadyknow the tag population of previous subgroups. By taking theaverage of those, we can produce a more accurate estimate n̂i

of number of tags in the ith subgroup

n̂i =i−1∑j=1

mj/(i − 1) (1)

where mj is the number of tags in jth subgroup.Our partitioning algorithm is summarized in the following

pseudo-code:

Partitioning-Algorithm(m) {n ← 16 // initial frame size

s ← 4 // initial group size

kth ← 5 // threshold for regrouping

// 1st Stage - Initial Estimation on Tag Population

do {[kc, ke] ← Read-Slot(kc, ke)

// record the outcome of slot. kc and ke represent the

// number of consecutive collision and empty slots

if (kc ≥ kth)thens ← s/2; kc ← 0; Regroup(s)

else if (ke ≥ kth)thens ← 2 · s; ke ← 0; Regroup(s)

else if (Is-End-Of-Frame())thenn ← 2.39 · nc; Start-New-Frame(n)

end} until (nc = 0)

// 2nd Stage - Optimum Repartitioning

nest ← Estimate-Number-Of-Tags(m1, s);s∗ ← Optimum-Partition-Size(nest);

// 3rd Stage - Sequential Identification

for each subgroup i from s∗ subgroupsn ← ∑i−1

j=1 mj/(i − 1);do {

[ns, nc] ← Read-Frame(n);if (Is-End-Of-Frame())

thenn ← 2.39 · nc; Start-New-Frame(n)

end} until (nc = 0)

}

IV. PERFORMANCE ANALYSIS

In this section, we present probabilistic analytical modelsto determine the expected identification time. The followingnotations are used throughout the rest of this paper:

• Ns(n,m): the expected identification time to read out mtags by partitioning the tags into s subgroups with aninitial frame size of n.

• P [ns, nc | n,m]: the probability that, given n slots andm tags, the number of successful and collision slots isns and nc, respectively.

Consider first the non-partitioning case with m tags andan initial frame size of n. The expected identification timewithout partitioning was derived in [3]:

N1(n,m) = n +min(n,m)∑

ns=0

z∑nc=0

P [ns, nc | n,m]

· N1((β · nc)∗,m − ns)

(2)

where z =⌊min

(n − ns,

m−ns

2

)⌋, (x)∗ � round(x), and β

is set to 2.39 as will be derived later. A recursive equation toobtain the probability P [ns, nc | n,m] is also available in [3].

Now, consider the partitioning case in which all the tagsrandomly choose one of s subgroups with equal probability.In this case, the expected identification time with an initialframe size of n0 is given by

Ns(n0,m) =∑m1

· · ·∑ms

m!m1!m2! · · ·ms!

(1s

)m

· (N1(n0,m1) + · · · + N1(n̂s,ms))

(3)

where mi is a random variable indicating the number of tagsin ith subgroup and m = m1 + m2 + · · · + ms. Note that n̂s

is defined in (1).

V. NUMERICAL RESULTS AND SIMULATIONS

To evaluate the performance of the partitioning algorithm,we have performed extensive simulations with a discrete-event simulator. We consider a single reader and severaltags under ideal (i.e., error-free) channel conditions. Thetwo primary performance metrics in our experiments arethe average throughput and average identification time. Theaverage throughput is defined as a ratio of the number of tags

476 IEEE COMMUNICATIONS LETTERS, VOL. 12, NO. 6, JUNE 2008

1 2 3 4 5 6 7 8250

260

270

280

290

300

310

Number of Groups

Iden

tific

atio

n Ti

me

(slo

ts)

Blind(n0=16)

Perfect(n0=m/s)Estimation Gain

Partitioning Gain

Fig. 2. Two types of gains from partitioning (m = 100).

to the number of slots used to read the tags. The averageidentification time is the number of slots used to identify allthe tags.

We have mentioned the benefits of partitioning are twofold:fast estimation of the tag population and throughput gain.To verify those benefits, we measure the identification timeusing the equation (3) as the subgroup counts (s) rangesfrom 1 to 8. To determine the magnitudes of those gains,we test the cases of blind estimation (n0 = 16) and perfectestimation (n0 = 100/s), i.e., the case in which the numberof tags is known in advance. Figure 2 shows Ns(16, 100)and Ns(100/s, 100) versus the number of subgroups, whereNs(16, 100) and Ns(100/s, 100) corresponds to the blind andperfect estimation cases, respectively. The results in this figureclearly show that partitioning achieves two distinct types ofbenefits. In the case of 6 subgroups (s = 6), for example, thepartitioning gain refers to the improvement in throughput ofthe partitioning algorithm over non-partitioning ones, on thecondition that tag population is known in advance.

To prove superiority of our algorithm, we compare it withother algorithms including 2.39nc Rule [3] and EDFSA [6]via simulations. In the 2.39nc Rule, the number of collisionsis counted at the end of frame, then the next frame size isset to 2.39nc. In the EDFSA [6], tag partitioning was used toallow only a fraction of tag population to access the channelwhen the maximum frame size is finite. Figure 3 showsestimation latency and throughput of anti-collision algorithmsunder comparison as a function of the number of tags, whichranges from 100 to 1000. The estimation latency in Fig. 3(a) is defined as the time in slots taken to estimate thenumber of remaining tags within 3% or 5% error. As shownin the figure, our estimation procedure, executed at the firststage of our algorithm, outperform the estimation techniqueusing the equation (12) in [1]. In particular, the performancegap becomes wider as higher accuracy is required. Figure 3(b) shows that the partitioning algorithm has a throughput6 − 7% and 4% higher over the 2.39nc Rule and EDFSA,respectively. Considering the theoretical throughput limit 1/e(36.8%) of slotted Aloha that has been a common reference,

100 200 300 400 500 600 700 800 900 10000.3

0.32

0.34

0.36

0.38

0.4

0.42

Number of T agsT

hrou

hgpu

t

P artitioningE DF S A2.39nc R ule

100 200 300 400 500 600 700 800 900 10000

100

200

300

400

500

600

Number of T ags

Est

imat

ion

Late

ncy

(slo

ts) V ogt (3 %)

V ogt (5 %)partitioning (3 %)partitioning (5 %)

(a) Estimation latency of anti-collision algorithms

(b) Throughput of anti-collision algorithms

Fig. 3. Estimation latency and throughput of anti-collision algorithms.

such improvement is significant. These improvements arecontributed by both fast estimation and partitioning gain whichare well illustrated in Fig. 3 (a) and Fig. 2, respectively.

VI. CONCLUSIONS

In this work, we proposed a novel anti-collision algorithmthat utilizes the partitioning technique. We have shown thatpartitioning the tags into multiple subgroups produces twosignificant advantages: fast estimation of the tag populationand throughput gain. A significant gain in throughput overthe existing methods is observed via extensive simulation.

REFERENCES

[1] H. Vogt, “Efficient object identification with passive RFID tags,” in Proc.International Conference on Pervasive Computing, pp. 1854–1858, 2002

[2] EPCglobal, EPC Radio-Frequency Identity Protocols Class-1 Generation-2 UHF RFID Protocol for Communications at 860 MHz - 960 MHHz,ver 1.0.9, 2004

[3] F. S. Schoute, “Dynamic frame length Aloha,” IEEE Trans. Commun.,vol. 31, no. 4, pp. 565–568, Apr. 1983

[4] C. Floerkemeier and M. Wille, “Comparison of transmission schemes forframed Aloha based RFID protocols,” in Proc. International Symposiumon Applications on Internet, pp. 92–97, 2006

[5] W. Shin and J. G. Kim, “Partitioning of tags for near-optimum RFIDanti-collision performance,” in Proc. WCNC 2007, pp. 1673–1678, 2007

[6] S. Lee, S. Joo, and C. Lee, “An enhanced dynamic framed slottedAloha algorithm for RFID tag identification,” in Proc. MobiQuitous 2005,pp. 166–172, 2005