Versatile Access Control for Massive IoT: Throughput ...wicl.cnu.ac.kr/wordpress/DB/International_Journal/2019_04_Versatile... · ments including access throughput, access delay,

1536-1233 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://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: DOI 10.1109/TMC.2019.2914381, IEEETransactions on Mobile Computing

1

Versatile Access Control for Massive IoT:Throughput, Latency, and Energy Efficiency

Han Seung Jang, Member, IEEE , Hu Jin, Senior Member, IEEE ,Bang Chul Jung, Senior Member, IEEE , and Tony Q. S. Quek, Fellow, IEEE

Abstract—In this paper, we propose a novel access control (AC) mechanism for cellular internet of things (IoT) networks with massivedevices, which effectively satisfies various performance metrics such as access throughput, access delay, and energy efficiency. Basicidea of the proposed AC mechanism is to adjust access class barring (ACB) factor according to performance targets. For a givenperformance target, we derive the optimal ACB factors by considering not only the conventional preamble collision detection techniquebut also the early preamble collision detection technique, respectively. In addition, the proposed AC mechanism considers overall radioresources to optimize the ACB factor, which includes the number of preambles, random access response (RAR) messages, andphysical-layer uplink shared channels (PUSCHs), while most conventional ACB schemes consider only the number of preambles. Inparticular, the proposed AC mechanism is illustrated with two representative performance metrics: latency and energy efficiency.Through extensive computer simulations, it is shown that the proposed versatile AC mechanism outperforms the conventional ACBschemes in terms of various performance metrics under diverse resource constraints.

Index Terms—Internet of things, low-latency, energy-efficiency, access class barring (ACB), ACB factor, backlog estimation.

F

1 INTRODUCTION

1.1 Background

T YPICAL communication networks, which are opti-mized for human-to-human (H2H) communications,

are evolving to support machine-to-machine (M2M) com-munications or internet of things (IoT) services [1]. In par-ticular, 5G cellular wireless networks have been developedand standardized under careful consideration on IoT ser-vices in the third generation partnership project (3GPP)and international telecommunication union (ITU) [2], [3].For supporting massive connections, the IoT networks mayrequire various technologies including edge computing [4],artificial intelligence (AI) [5], etc. Optimizing M2M com-munication protocols is another interesting topic to satisfyvarious quality of service (QoS) requirements in the IoTnetworks [6], [7]. For example, in smart city, autonomousvehicles require very low latency communications withnearby vehicles, while temperature sensors require an ultra-high energy efficiency (EE) to prolong their battery life sincethe battery lifetime is required to be around 10 years. In par-ticular, Sivanathan et al. [8] proposed a robust framework for

• H. S. Jang is with the School of Electrical, Electronic Communication, andComputer Engineering, Chonnam National University, Yeousu, 59626,Republic of Korea.E-mail: [email protected]

• H. Jin is with the Division of Electrical Engineering, Hanyang University,Ansan 15588, Republic of Korea. E-mail: [email protected]

• B. C. Jung is with the Department of Electronics Engineering, ChungnamNational University, Daejeon 34134, Republic of Korea.E-mail: [email protected]

• T. Q. S. Quek is with the Information Systems Technology and DesignPillar, Singapore University of Technology and Design (SUTD),Singapore 487372. E-mail: [email protected]

• Manuscript received XX. XXX. 2018; revised XX. XXX. 2019, acceptedXX. XXX. 2019 (Corresponding authors: Hu Jin and Bang Chul Jung).

IoT device classification using network traffic characteristicsbased on machine learning schemes.

Existing commercial cellular networks may encountercritical bottlenecks if they are adopted for massive IoTnetworks without significant modifications. First of all,the amount of radio resources in the cellular network isvery limited to accommodate a rapidly growing numberof IoT devices, and thus careful resource management isrequired [9], [10], [11]. Andrade et al. [9], [10] addressed thatthe control channels including physical downlink channel(PDCCH) may be limited to support overload control, andthey proposed a prioritized control resource schedulingalgorithm considering both of H2H and M2M communi-cations. Xia et al. [11] provided a comprehensive technicalsurvey on the resource management techniques for M2Mcommunications in commercial LTE/LTE-A cellular net-works.

1.2 Related works

Technical challenges may come from massive nodes that areactivated simultaneously due to bursty traffic. In general,the bursty traffic pattern causes severe congestion or over-load problem in a radio access network (RAN). To overcomethe congestion problem, several overload control schemeshave been proposed [12], [13], [14], [15], [16], [17], [18], [19].Among them, the access class barring (ACB) scheme hasbeen considered as a promising technology since it effec-tively controls bursty traffic in cellular M2M networks [14].In particular, the dynamic access class barring (D-ACB)schemes have been actively studied in order to support avarying number of access nodes in IoT networks [17]. TheD-ACB schemes estimate the number of backlogged nodesto attempt access at each time slot, and then it generatesan optimal ACB factor, also known as access probability, to



2

allow nodes to send a preamble via physical random accesschannel (PRACH), by considering the available resources.Jin et al. [18] proposed a recursive pseudo-Bayesian ACBscheme in order to lower total service time of a burstytraffic, and a practical backlogged estimation scheme basedon the number of undetected preambles at the first step ofthe random access procedure. On the other hand, Tavanaet al. [15] proposed a backlog estimation method based onKalman filtering for the adaptive ACB scheme, where it isshown that total service time becomes close to that withthe optimal ACB scheme that exactly knows the numberof backlogged nodes. Duan et al. proposed a dynamicpreamble allocation scheme and an iterative algorithm toadaptively update the ACB factor for M2M devices. Vuralet al. [19] proposed an adaptive preamble slicing algorithm,called multi-preamble random access, to dynamically deter-mine the preamble set size according to different load condi-tions and service classes. However, most of existing studiesfocused on the dynamic access control mechanisms to meeta single performance target such as maximization of thenumber of successful access or equivalently minimizationof access delay and total service time. In addition, they onlyconsidered the preamble resources at the first step of RAprocedure even though other radio resources are involvedin the overall RA procedure.

EE is one of the most important performance metricsespecially for battery-powered IoT devices, and thus manytechniques have been proposed in literature [20], [21], [22],[23], [24], [25], [26], [27], [28], [29]. Ho et al. [20] proposed ajoint massive access control and resource allocation schemeto improve the EE, which adopts machine node grouping,coordinator selection, and optimal power allocation for co-ordinators. Zhang et al. [26] proposed an integrated energyefficient system for 5G IoT networks, which holisticallycombines the wireless and wired parts together in orderto optimize the EE of the system. Azari and Miao [28]proposed a resource allocation framework based on themax-min lifetime-fairness, which exploits the informationon the battery time of devices. Yang et al. [27] proposedan energy efficient resource allocation scheme with energyharvesting technology in time division multiple access andnon-orthogonal multiple access strategies. However, theeffect of the access control technique on the EE have beenrarely investigated in massive IoT networks to the best ourknowledge.

1.3 Contributions

In this paper, we propose a versatile access control mecha-nism that effectively satisfies various performance require-ments including access throughput, access delay, and theEE for massive IoT networks. To obtain the ACB factor forsatisfying latency or EE requirement, the proposed mecha-nism considers all the radio resources involved in the overallRA procedure: preambles on PRACH, RAR messages onphysical downlink shared channel (PDSCH) or PDCCH,and uplink data resources on physical uplink shared chan-nel (PUSCH). In addition to the conventional preamblecollision detection (C-PCD) technique, we also exploit theearly preamble collision detection (E-PCD) technique at thefirst RA procedure [30], [31]. Furthermore, we propose a

comprehensive access control mechanism which can simul-taneously support two groups of devices requesting low-latency and ultra-low power consumption services, respec-tively. Simulation results show that the proposed versatileAC mechanism efficiently satisfies various performance re-quirements under various resource constraints and it out-performs the existing AC schemes. The main contributionsof this paper can be summarized as follows:

• We propose a versatile AC mechanism to efficientlycontrol random access traffic from a massive num-ber of IoT nodes while achieving various perfor-mance requirements such as low-latency and energy-efficiency.

• We derive the ACB factors for low-latency accessand energy-efficient access with the conventionalpreamble collision detection technique and the earlypreamble collision detection technique, respectively.

• We also propose a comprehensive AC mechanism inorder to simultaneously support both of low-latencyaccess group and energy-efficient access group.

The rest of this paper is organized as follows. In SectionII, we describe a system model considered in this paper. InSection III, we derive preliminary probabilities of collision-free and collided preambles which are utilized to optimizethe ACB factor of the proposed mechanism. In Section IV,we propose two types of AC mechanisms: low-latency ACand energy-efficient AC mechanisms. Then, we explain theestimation algorithm for the number of backlogged nodesin Section V. The proposed AC mechanisms are validatedvia extensive computer simulations in Section VI. Finally,conclusions are drawn in Section VII.

2 SYSTEM MODEL

Fig. 1 shows the system model of a single cell massive IoTnetwork which is considered in this paper. We assume thatthe eNodeB controls uplink access from two representativegroups1, each of which requires a certain performance met-ric such as low-latency (LL) and energy efficiency (EE). Ac-tivation of nodes is assumed to be bursty due to an externalevent for both groups. In addition, the eNodeB is assumedto adopt a probability-based dynamic access class barring(ACB) technique as used in the commercial cellular network,i.e., 3GPP LTE system [32], which inherently controls anumber of simultaneous accesses by using a single numbercalled ACB factor p ∈ (0, 1). Let pLL and pEE denote theACB factors for LL access and EE access, respectively. In theproposed versatile AC mechanism, we consider followingparameters:

• M : number of available preambles• Q: number of available RAR messages• U : number of available PUSCH resources• N : total number of activated nodes in a group• νi: estimated number of backlogged nodes for the ith

slot

1. In practice, there may be a number of different groups. For exam-ple, a group of devices may require a mixed performance requirementamong maximum throughput, low-latency, and low-power consump-tion.



3

pEE

pLL

Low-latency Ultra-low power

Time

Num

ber of

acti

vate

d n

odes

Time

Num

ber of

acti

vate

d n

odes

Bursty

traffic

Bursty

traffic

Fig. 1. System model of a massive IoT network with versatile accesscontrol mechanism.

• pi: ACB factor for the ith slot

The above parameters will be explained in the next subsec-tion in detail. Note that Q depends on PDSCH or PDCCHresources, and U depends on PUSCH resources. Since asingle RAR message can deliver a single uplink resourcegrant for PUSCH [33], the number of allocable PUSCHresources is limited to

K = min{Q,U}. (1)

Thus, we only consider the number of available pream-bles M and the number of allocable PUSCH resources Kthroughout the paper.

2.1 Overall Random Access Procedure

The overall RA procedure consists of five steps, whichinclude the access check (Step 0) and the typical four-stepcontention-based RA procedure (Step 1-4) [34].

(Step 0) Access check: The eNodeB calculates an ACBfactor for a specific purpose based on the estimated num-ber of backlogged nodes νi and the number of availableresources M and K, and then it notifies the ACB factor ofthe ith PRACH pi ∈ [0, 1] at the beginning of the ith PRACHslot. Then, each activated node generates a random numberq ∈ [0, 1] and compares q with pi. If q ≤ pi, then the nodeattempts an RA on the ith PRACH slot. Otherwise, it defersan RA attempt to the (i+ 1)th PRACH slot.

(Step 1) Preamble transmission and detection: Eachnode which passed the access check at the Step 0 randomlyselects a single preamble among M available preambles,and then sends the selected preamble on the ith PRACHslot. More than one node may select the same preamble dueto the random preamble selection property.

(Step 2) Random access response (RAR): After detectingthe preambles from nodes, the eNodeB sends the RAR mes-sages, each of which conveys identity of detected preambles,timing advance information, and an initial uplink PUSCHresource grant for uplink data transmission.

(Step 3) Uplink data transmission: Using the PUSCHresource indicated via the RAR message at the Step 2, thecorresponding node transmits uplink data which conveysa data packet, a radio resource control (RRC) connectionrequest, a tracking area update, or a scheduling request.

(Step 4) ACK message transmission: If the eNodeBsuccessfully decodes the data transmitted by a single node,

Preamble

Transmission

(on PRACH)

RA Response

(on PDCCH)

Data

Transmission

(on PUSCH)

Contention

Resolution

(on PDCCH)

Device

eNodeB

Collisiondetection

1 2 3 4 Collisionnotified

· Detected PA

· Undetected PA

Resourceallocation to detected PAs

No ACK for collided preamble

(a) Conventional PCD technique

Preamble

Transmission

(on PRACH)

RA Response

(on PDCCH)

Device

eNodeB

1 2 Collisionnotified

· Collision-free PA

· Collided PA

· Undetected PA

Resource allocation only to collision-free PAs

(b) Early PCD technique

Fig. 2. Comparison of preamble collision detection (PCD) techniques:(a) conventional PCD technique and (b) early PCD technique.

it sends back the ACK message including the node identity(ID) which is obtained from the decoded data, otherwise theeNodeB sends nothing back.

2.2 Preamble Collision Detection TechniquesAs noted before, we consider two preamble collision detec-tion (PCD) techniques: conventional PCD (C-PCD) and earlyPCD (E-PCD) [35]. Fig. 2 compares the two techniques. TheC-PCD technique is performed based on unsuccessful datadecoding at Step 3 of the RA procedure. In other works, ifmultiple nodes send the same preamble at Step 1 and receivethe identical RAR message from the eNodeB at Step 2, thenall relevant nodes transmit their own data via the sameuplink resource and the eNodeB may not decode the datasuccessfully. From this unsuccessful decoding, the eNodeBdetects the preamble collision. If the preamble collisions aredetected, the eNodeB does not send the ACK message tothe corresponding devices at Step 4. Nodes which do notreceive the ACK message try to perform the RA procedureat the next PRACH slot.

On the other hand, the PCD can be performed at Step1 of the RA procedure with the E-PCD techniques whichexploit tagged preambles [30], [31], ID transmission [36], ormachine learning algorithms [37]. If the preamble collisionsare detected with E-PCD techniques at Step 1, the eNodeBdoes not send the RAR message to the corresponding nodesat Step 2. Then, nodes which do not receive the RARmessage defer to attempt the RA to the next PRACH slotwithout transmitting uplink data. It is worth noting that theE-PCD techniques can separate the detected preambles intocollision-free preambles and collied preambles and allocatesthe PUSCH resources only for the collision-free preamblesthrough the RAR messages at Step 2.

2.3 Traffic ModelWe assume that each of N nodes in a group is activated atthe time x ∈ (0, Tact), which follows Beta distribution with



4

parameters α = 3 and β = 4 [38]:

fX(x) =xα−1(Tact − x)β−1

(Tact)α+β−1B(α, β),

where B(α, β) =∫ 10 x

α−1(1 − x)β−1dx denotes the betafunction. It is assumed that there exist Iact slots within timeTact and the expected number of newly activated nodes inthe ith RA slot is given by λi = N

∫ titi−1

fX(x)dx for i =1, 2, . . . , Iact.

3 PRELIMINARY PROBABILITIES ON COLLISION-FREE PREAMBLE AND COLLIDED PREAMBLE

In this section, we first define the following discrete randomvariables (RVs):

• D: number of detected preambles,• S: number of collision-free preambles,• C: number of collided preambles,• R: number of undetected preambles,

where D = S+C and R = M −D. Then, we derive severalpreliminary probabilities related to the above RVs, whichare used for designing the versatile AC mechanism.

At Step 1 of the RA procedure, each activated nodewhich passes the access check sends an arbitrary preambleamong M available preambles via PRACH, and then theeNodeB detects preambles. In general, detected preamblesare classified into collision-free preambles and collidedpreambles. The collision-free preamble indicates the de-tected preamble which is sent by a single node, while thecollided preamble indicates the detected preamble which issent by more than two nodes. However, depending on thePCD techniques (e.g., E-PCD or C-PCD), the eNodeB canor cannot identify the detected preamble is collision-free orcollided at Step 1 of the RA procedure.

We first consider the following joint probability thatwhen m nodes simultaneously attempt the RAs, any (r+ s)preambles among M preambles are chosen, and then eachof s preambles among chosen (r + s) preambles is selected(collision-free) only by a sinlge node, and each of remainingr preambles is not selected by any nodes:

ΨMr+s(s|m)

=

(M

r + s

)(r + s

s

)(m

s

)s!{M − (r + s)}m−s

Mm, (2)

where( Mr+s

)(r+ss

)represents the total number of cases that

chooses (r + s) preambles among M , and then chooses scollision-free preambles among the chosen (r+s) preambles[15]. In addition, the term

(ms

)s!(d − s)m−s represents the

total number of cases that s nodes among m RA-attemptingnodes transmits an exclusive preamble and the remaining(m− s) nodes chooses one of preambles among {M − (r +s)} preambles. Then, the denominator Mm represents thetotal number of cases that m RA-attempting nodes select apreamble among M preambles.

Let Ei denote the event that the ith preamble is selectedby at most one node, and then the probability of the eventEi when m nodes attempt RAs simultaneously, i.e, m nodespass the access check at Step 0, is given by the sum of the

probability that ith preamble is not selected by any nodesand the probability that ith preamble is selected only by asingle node, i.e.,

Pr{Ei|m} =

(1− 1

M

)m+

(m

1

)1

M

(1− 1

M

)m−1, (3)

where 1M denotes the probability that a node selects the ith

preamble among M possibilities.Based on the inclusion-exclusion principle [39], Pr{E1 ∪

· · · ∪ EM |m} is written as

Pr{∪Mi=1Ei|m

}=

M∑i=1

(−1)i+1i∑

j=0

ΨMi (j|m)

=M∑i=1

i∑j=0

(−1)i+1

(M

i

)(i

j

)(m

j

)j!

(M − i)m−j

Mm, (4)

which denotes the probability that at least one preambleamong M preambles is selected by at most one node withm RA-attempting nodes. Besides, the complementary prob-ability to Pr

{∪Mi=1Ei|m

}is given by

Φ(M |m) = 1− Pr{∪Mi=1Ei|m

}= 1−

M∑i=1

i∑j=0

(−1)i+1

(M

i

)(i

j

)(m

j

)j!

(M − i)m−j

Mm

=M∑i=0

i∑j=0

(−1)i(M

i

)(i

j

)(m

j

)j!

(M − i)m−j

Mm, (5)

which represents the probability that each of all M pream-bles are selected by at least two nodes (collided) with mRA-attempting nodes. As a result, we obtain the joint prob-ability that s preambles are collision-free among d detectedpreambles when m nodes simultaneously attempt the RA:

Pr{D = d, S = s|m} = ΨMr+s(s|m)Φ(M − r − s|m− s)

= ΨMM−d+s(s|m)Φ(d− s|m− s)

=d−s∑i=0

i∑j=0

(M

d

)(d

s

)(m

s

)s!

(d− s)m−s

Mm×

(−1)i(d− si

)(i

j

)(m− sj

)j!

(d− s− i)m−s−j

(d− s)m−s, (6)

where we utilize r = M − d and( MM−d+s

)(M−d+ss

)=(M

d

)(ds

).

We assume that a Priori distribution of the number ofbacklogged nodes n follows Poisson distribution with meanof ν, i.e.,

P(n|ν) =νn

n!e−ν . (7)

It is shown that this assumption significantly reduces thecomputation complexity of designing algorithms [40]. Then,for given ACB factor p and ν, the joint probability that spreambles are collision-free among d detected preambleswhen m out of n backlogged nodes simultaneously attemptthe RA is given by

Pr{d, s, n,m|p, ν} = Pr{d, s|m}︸︷︷︸(6)

Pr{m|p, n}︸︷︷︸Bnm(p)

Pr{n|ν}︸︷︷︸P(n|ν)

, (8)



5

where Bnm(p) =(nm

)pm(1 − p)(n−m) denotes binomial dis-

tribution with ACB factor of p. By marginalizing (8) withregard to n and m, the joint probability that s preambles arecollision-free among d detected preambles for given p and νis given by

Pr{D = d, S = s|p, ν} =∞∑n=0

n∑m=0

Pr{d, s, n,m|p, ν}

=

(M

d

)(d

s

)e−pν

(pνM

)s (−1− pν

M+ e

pνM

)d−s. (9)

Then, by marginalizing Pr{D = d, S = s|p, ν} with regardto S and D, respectively, Pr{D = d|p, ν} and Pr{S =s|p, ν} are also given by

Pr{D = d|p, ν} =d∑s=0

Pr{D = d, S = s|p, ν}

=

(M

d

)e−pν

d∑s=0

(d

s

)(pνM

)s (−1− pν

M+ e

pνM

)d−s=

(M

d

)e−pν

(−1 + e

pνM

)d, (10)

and

Pr{S = s|p, ν} =M∑d=0

Pr{D = d, S = s|p, ν}

= e−pν(pνM

)s M∑d=0

(M

d

)(d

s

)(−1− pν

M+ e

pνM

)d−s=

(M

s

)e−pν

(pνM

)s M∑d=0

(M − sd− s

)(−1− pν

M+ e

pνM

)d−s=

(M

s

)e−pν

(pνM

)s (−pνM

+ epνM

)M−s, (11)

respectively. Since C = D − S, we have Pr{C = c, S =s|p, ν} by substituting s + c for d in Eq. (9). Thus, theprobability that the number of collided preambles is equalto c is given by

Pr{C = c|p, ν} =M−c∑s=0

Pr{C = c, S = s|p, ν}

=

(M

c

)e−pν

(−1− pν

M+ e

pνM

)c (1 +

pν

M

)M−c. (12)

As a result, for given p and ν, the expected values of RVs D,

S and C are given by

E[D|p, ν] =M∑d=0

d

(M

d

)e−pν

(−1 + e

pνM

)d= Me−pν

(−1 + e

pνM

) M∑d=0

(M − 1

d− 1

)(−1 + e

pνM

)d−1= Me−pν

(−1 + e

pνM

)epνM (M−1)

= M(

1− e−pνM

), (13)

E[S|p, ν] =M∑s=0

s

(M

s

)(pνM

)se−pν

(−pνM

+ epνM

)M−s= pνe−pν

M∑s=0

(M − 1

s− 1

)(pνM

)s−1 (−pνM

+ epνM

)M−s= pνe−pνe

pνM (M−1)

= pνe−pνM , (14)

andE[C|p, ν] = E[D|p, ν]− E[S|p, ν]

= M(

1− e−pνM

)− pνe−

pνM

= M(−1− pν

M+ e

pνM

)e−

pνM , (15)

respectively.

4 VERSATILE ACCESS CONTROL MECHANISMS

In this section, we propose two representative AC mecha-nisms: low-latency AC mechanism and energy-efficient ACmechanism. For each mechanism, we derive the ACB factorsby considering both the C-PCD technique and the E-PCDtechnique. Here, we assume that the estimated number ofbacklogged nodes is equal to ν. The estimation algorithmfor the number of backlogged nodes will be explained inthe next section.

4.1 Low-Latency Access Control

The low-latency requirement can be satisfied by maximizingthe access throughput in each slot. Thus, we derive theACB factor for maximizing the access throughput which isdefined as the number of nodes that succeed in the RA.

4.1.1 With C-PCD technique

Basically, the RA of a certain node is successful whenit sends an exclusive preamble and receives an exclusivePUSCH resource grant. Let A denote the access throughput,i.e, the number of nodes that succeed in the RA. Recall thatK denotes the number of allocable PUSCH resources at Step2 of the RA procedure and D denotes the number of de-tected preambles. By using the well-known total probabilitytheorem, the probability that A = a can be obtained as

Pr{A = a} = Pr{A = a,D ≤ K}+ Pr{A = a,D > K}= Pr{S = a,D ≤ K}+ Pr{S ≥ a,D > K}Bsa

(KD

), (16)

where Bsa(KD

)=(sa

) (KD

)a (1− K

D

)s−aand K

D < 1. Theterm K

D < 1 can be regarded as the resource scheduling



6

probability. Then, the conditional expected value of A forgiven p and ν is given by

E[A|p, ν] =K∑a=0

K∑d=0

aPr{D = d, S = a|p, ν}+

K∑a=0

M∑d=K+1

d∑s=a

aPr{D = d, S = s|p, ν}Bsa(KD

). (17)

Then, we obtain the optimal ACB factor p to maximize theaverage access throughput as follows:

pLLconv = arg max0≤p≤1

E[A|p, ν]. (18)

In particular, when K = M , Eq. (17) can be rewritten by

E[A|p, ν] =M∑a=0

M∑d=0

aPr{D = d, S = a|p, ν}

= E[S|p, ν]. (19)

Thus, in this case, the optimal pLLconv is found by takingderivative of E[S|p, ν] = pνe−

pνM with respect to p, i.e,

pLLconv =M

ν. (20)

When K < M , on the other hand, it is difficult to findthe closed-form expression for the optimal ACB factor andthus we approximate E[A|p, ν] as

E[A|p, ν] ≈ E[S|p, ν] ·min

{1,

K

E[D|p, ν]

}, (21)

where we have two components: E[S|p, ν] represents theaverage number of nodes that succeed in preamble trans-missions, and min

{1, K

E[D|p,ν]

}represents the successful

resource scheduling probability. In case that E[D|pLLconv =Mν , ν] ≤ K, the optimal ACB factor can be approximated

as Eq. (20) since min{1, KE[D|pLLconv,ν]

} = 1 and E[A|p, ν] ≈E[S|p, ν]. On the other hand, when E[D|pLLconv = M

ν , ν] > K,E[A|p, ν] increases until the ACB factor p approaches topLLconv such that E[D|pLLconv, ν] = K and E[A|p, ν] becomesdecreased if p > pLLconv since E[S|p, ν] < E[D|p, ν] andE[D|p, ν] = M

(1− e−

pνM

)is an increasing function of p. It

implies that the eNodeB should control the ACB factor p sothat it allocates PUSCH resources to all detected preamblesin order to maximize access throughput. In other words, theresource scheduling probability needs to approach to 1 formaximizing the access throughput.

As a result, we obtain the following ACB factor bysolving E[D|p, ν] = M

(1− e−

pνM

)= K:

pLLconv = − ln

(1− K

M

)M

ν, K < M. (22)

4.1.2 With E-PCD techniqueWith E-PCD techniques, the eNodeB allocates resourcesonly for the collision-free preambles, and thus the averageaccess throughput is given by

E[A|p, ν] =M∑a=0

min(a,K) Pr{S = a|p, ν}, (23)

where S denotes the number of collision-free preambles.Then, we obtain the optimal ACB factor p to maximize theaverage access throughput as follows:

pLLearly = arg max0≤p≤1

E[A|p, ν]. (24)

Similar to the case of C-PCD technique, when K = M ,

E[A|p, ν] =M∑a=0

aPr{S = a|p, ν} = E[S|p, ν], (25)

and thus the optimal pLLearly is also found by

pLLearly =M

ν. (26)

Note that, when K = M , the optimal ACB factors with theC-PCD and the E-PCD are identical, i.e., pLLconv = pLLearly = M

ν .When K < M , on the other hand, it is difficult to find

the closed-form expression for the optimal ACB factor, andthus, we approximate E[A|p, ν] as

E[A|p, ν] ≈ E[S|p, ν] ·min

{1,

K

E[S|p, ν]

}, (27)

where we have two components: E[S|p, ν] represents theaverage number of nodes that succeed in preamble trans-missions, and min

{1, K

E[S|p,ν]

}represents the successful

resource scheduling probability. Note that Eq. (27) is dif-ferent from Eq. (21), which comes from the fact that theE-PCD technique allocates resources only for the collision-free preambles. In case that E[S|pLLearly = M

ν , ν] ≤ K, theoptimal ACB factor can be approximated as Eq. (26) sincemin{1, K

E[S|pLLearly,ν]} = 1 and E[A|p, ν] ≈ E[S|p, ν]. However,

when E[S|pLLearly = Mν , ν] > K, E[A|p, ν] increases until the

ACB factor p approaches to pLLearly such that E[S|pLLearly, ν] = K

and E[A|p, ν] ≈ K if pLLearly < p ≤ pLLearly. It implies that theeNodeB cannot allocate resources even for some collision-free preambles when p > pLLearly, which is called resource allo-cation failure in this paper. The average number of collision-free preambles that experience the resource allocation failureis given by

F (p, ν) =

{0, if p ≤ pLLearly,E[S|p, ν]−K, if pLLearly < p < pLLearly.

We obtain the ACB factor p such that E[S|p, ν] = K tomaximize the access throughput and minimize the numberof resource allocation failures (the number of unnecessaryaccesses). As a result, we obtain the following ACB factorby solving E[S|p, ν] = pνe−

pνM = K:

pLLearly = −W0

(−KM

)M

ν, K < M, (28)

where W0(x) represents the principle (upper) branch Lam-bert W function withW0(x) > −1 [41].

However, if pLLearly is used as the ACB factor, the av-erage access throughput becomes slightly lower than Kbecause the transmissions occur in a probabilistic man-ner. To validate this, let p denote the ACB factor andBNn (p) =

(Nn

)(p)n(1 − p)(N−n) denotes the probability that

n nodes attempt RAs among N backlogged nodes with the



7

ACB factor of p. Then, the average number of successfulRA-attempts is given by

S(p,K) =N∑n=0

min[n(1− 1

M

)n−1,K]BNn (p), (29)

where S(pLLearly,K) is smaller than S(pLLearly,K).On the other hand, let us define unsuccessful RA-attempts

as the attempts either with collided preambles or withcollision-free preambles experiencing the resource allocationfailure. The average number of unsuccessful RA-attempts isgiven by

U(p,K) =N∑n=0

nBNn (p)− S(p,K). (30)

Note that the unsuccessful RA-attempt causes energy wasteof nodes. Hence, the ACB factor needs to be carefully con-trolled between pLLearly and pLLearly by considering both S(p,K)and U(p,K) in practice. We will further discuss a detailedsetting for p in the section of numerical results.

4.2 Energy Efficient Access ControlEnergy efficiency (EE) is defined as the ratio of energyconsumed for the successful uplink transmissions to energyconsumed for all uplink transmissions during the RA pro-cedure at a node.

4.2.1 With C-PCD techniqueThe EE with C-PCD technique is given by

E[ηEEconv(p)] =(1 + χ)E[S|p, ν] min

(1,E

[KD |p, ν

])pν + χpνmin

(1,E

[KD |p, ν

])=

(1 + χ)E[S|p, ν] min(1,E

[KD |p, ν

])pν{

1 + χmin(1,E

[KD |p, ν

])} , (31)

where χ denotes the ratio of energy consumed for thepreamble transmission at Step 1 over energy consumedfor the data transmission at Step 3 at a node. If wecontrol the ACB factor such that all nodes with detectedpreambles receive resource grants, i.e., E[D|p, ν] ≤ K,min

(1,E

[KD |p, ν

])= 1, then E[ηEEconv(p)] is rewritten as

E[ηEEconv(p)] =(1 + χ)E[S|p, ν]

(1 + χ) pν=pνe−

pνM

pν= e−

pνM . (32)

The EE with C-PCD technique is a decreasing function of p,and thus there exists a trade-off between the average accessthroughput and the energy efficiency. It is worth noting that(32) does not depend on the energy ratio of preamble overdata transmissions χ. In particular, if pLLconv and pLLconv are usedfor the ACB factor in (32), then we have

E[ηEEconv(p

LLconv)] = e−1 (33)

and

E[ηEEconv(p

LLconv)] = 1− K

M , (34)

respectively.For achieving a certain EE of θ for a group of nodes, the

ACB factor is controlled as:

pEE:θconv = − ln(θ)M

ν, (35)

which satisfies

E[ηEEconv(p)] = e−pνM = θ.

For energy-efficient AC, the target value of θ needs to behigher than both E[ηEE

conv(pLLconv)] and E[ηEE

conv(pLLconv)], which is

set to

θ >

{e−1, if E[D|pLLconv = M

ν , ν] ≤ K,1− K

M , if E[D|pLLconv = Mν , ν] > K.

4.2.2 With E-PCD technique

The EE with E-PCD technique is given by

E[ηEEearly(p)] =(1 + χ)E[S|p, ν] min

(1,E

[KS |p, ν

])pν + χE[S|p, ν] min

(1,E

[KS |p, ν

]) . (36)

If we control the ACB factor such that all nodeswith collision-free preambles receive resource grants, i.e.,E[S|p, ν] ≤ K, min

(1,E

[KS |p, ν

])= 1, then E[ηEEearly(p)] is

rewritten as

E[ηEEearly(p)] =(1 + χ)E[S|p, ν]

pν + χE[S|p, ν]=

(1 + χ)pνe−pνM

pν + χpνe−pνM

=(1 + χ)e−

pνM

1 + χe−pνM

. (37)

The EE with E-PCD technique is a decreasing function ofp, and thus there exists a trade-off between the averageaccess throughput and the energy efficiency as well. Incontrast to E[ηEEconv(p)] in (32), the EE with E-PCD techniqueE[ηEEearly(p)] depends on the energy ratio of preamble overdata transmissions χ. In particular, if pLLearly and pLLearly are usedfor the ACB factor in (37), then we have

E[ηEE

early(pLLearly)] =

(1 + χ)e−1

1 + χe−1(38)

and

E[ηEE

early(pLLearly)] =

(1 + χ)eW0(−KM )

1 + χeW0(−KM )

, (39)

respectively.For achieving a certain EE of θ for a group of nodes, the

ACB factor is controlled as:

pEE:θearly = − ln

(θ

1 + (1− θ)χ

)M

ν, (40)

which satisfies

E[ηEEearly(p)] =(1 + χ)e−

pνM

1 + χe−pνM

= θ.

For energy-efficient AC, the target value of θ needs to behigher than both E[ηEE

early(pLLearly)] and E[ηEE

early(pLLearly)], which is

set to

θ >

(1+χ)e−1

1+χe−1 , if E[S|pLLearly = Mν , ν] ≤ K,

(1+χ)eW0(−KM )

1+χeW0(−KM )

, if E[S|pLLearly = Mν , ν] > K.



8

4.3 Comprehensive Access Control

In subsections 4.1 and 4.2, we proposed low-latency AC andenergy-efficient AC mechanisms, respectively. However, inpractical situation, the eNodeB is required to support ac-cesses both from nodes with low-latency QoS and fromnodes with energy-efficient QoS at the same time. First, weassume that eNodeB can identify the access ratio betweenlow-latency access and energy-efficient access. Let γ and(1−γ) denote the access ratios of the low-latency access andthe energy-efficient access, respectively. Then, when thereexists Nt backlogged nodes at time slot t in total, γNt and(1 − γ)Nt backlogged nodes attempt random access withlow-latency QoS and energy-efficient QoS, respectively, onthe average. In the proposed comprehensive access controlmechanism, the eNodeB broadcasts two respective ACBfactors pLL and pEE:θ, and each node utilizes one of ACB fac-tors depending on its QoS (low-latency or energy-efficiency)requirement. Consequently, we have the following equation:

γNtpLL + (1− γ)Ntp

EE:θ = Ntpcomp., (41)

where pcomp. = γpLL+(1−γ)pEE:θ denotes a comprehensiveACB factor, which is utilized to estimate Nt as ν. Based onthe estimated backlogged size ν, ACB factors pLL and pEE:θ

are calculated based on the proposed low-latency AC in (20),(22), (26), and (28) and energy-efficient AC mechanisms in(35) and (40), respectively.

5 ESTIMATION ON THE NUMBER OF BACKLOGGEDNODES

We extend the pseudo-Bayesian backlog estimation scheme[18] in order to estimate the number of backlogged nodesfor the proposed versatile AC mechanism. The pseudo-Bayesian backlog estimation scheme operates based on thenumber of available preambles M and the number of un-detected (idle) preambles r, which can be exactly identifiedduring the preamble detection procedure at Step 1 of theRA procedure for both C-PCD and E-PCD techniques. Algo-rithm 1 summarizes the estimation and update algorithm ofν. The eNodeB first estimates ν0 as the number of preamblesM . Then, the estimated value is updated as ν = ν+∆ν. Theestimation offset ∆ν > 0 is given by [18]

∆ν(p) = E[n|r, p, ν]− ν = νp

(e−

pνM − r

M

1− e− pνM

). (42)

For various ACB factors, (42) is rewritten as:

∆ν(pLLconv) = ∆ν(pLLearly) =Me−1 − r1− e−1

, (43)

∆ν(pLLconv) = − ln

(1− K

M

)(M

K

)(M −K − r), (44)

∆ν(pLLearly) = −W0

(−KM

)M

(eW0(−K

M ) − rM

1− eW0(−KM )

), (45)

∆ν(pEE:θconv ) = − ln(θ)M

(θ − r

M

1− θ

), (46)

∆ν(pEE:θearly ) = − ln(ξ)M

( θ1+(1−θ)χ −

rM

1− θ1+(1−θ)χ

). (47)

Algorithm 1 Estimation and update algorithm of ν1: Initialize ν0 = M , k0 = 0, and p0.2: Count the number of undetected preambles ri−1

3: Calculate ∆ν according to (42) with ri−1, pi−1, and νi−1

4: Update νi−1 = νi−1 + ∆ν5: if ∆ν > 0 then6: ki = ki−1 + 1 and νi−1 = νi−1 + ki · ∆ν . Boosting7: else8: ki = 0 and νi−1 = νi−1

9: end if10: νi = max(M,νi−1 − ci−1) . New estimation of ν for slot i

TABLE 1Simulation parameters and values

Parameters Values

The number of preambles, M 64

The number of allocable PUSCH resources, K 10 ∼ 64

Energy consumption ratio, χ 3

PRACH interval, Tinterval 1 msActivation duration, Tact 500 ms

Total number of RA-attempting nodes, N 10000

Access ratio, γ 0.5

Target energy efficiency, θ 0.6/0.8/0.9

For the comprehensive AC mechanism, the estimation offsetis calculated by ∆ν(pcomp.), where pcomp. = γpLL + (1 −γ)pEE:θ. As in [18], we introduce a boosting factor ki inorder to reflect characteristics of the bursty traffic. It impliesthat when we observe that the traffic continuously increasesbased on ∆ν > 0, we boost the estimation ν. At the last line,the new estimation of νi is calculated by νi = νi−1 − ci−1,where ci−1 denotes the number of successful accesses on the(i− 1)th PRACH slot.

6 NUMERICAL RESULTS

Table 1 summarizes parameters used for computer simula-tions. We assume that random access opportunity (PRACH)appears every 1 ms, i.e., Tinterval = 1 ms, and the numberof preambles is equal to M = 642. The number of allocableresources K is assumed to vary from 10 to 64 and eachgroup of nodes consists of 10000 nodes which try to connectto the network through the RA procedure. The proposedlow-latency AC mechanism aims at the lowest access delayor equivalently the maximum access throughput. For theproposed energy-efficient AC mechanism, the target EEvalue is assumed to be θ. In particular, for the comprehen-sive AC mechanism, we set the access ratio γ to 0.5. Weevaluate the performances of the proposed AC mechanismsin terms of access throughput, average access delay, andenergy efficiency. The access delay is defined as the timedifference between the access completion time and the ac-tivation time of a node. For comparison, the performanceof the conventional low-latency AC mechanisms [12] – [18]with the ACB factor of p = M

ν is also illustrated. Among[12] – [18], the backlog estimation scheme of [18] is utilizedto obtain ν representatively.

2. The proposed versatile AC mechanism can dynamically adjustACB factors according to the number of preambles, but we focus onthe case of the fixed number of preambles in this paper.



9

0 100 200 300 400 500 600 700 800 900 10000

1000

2000

3000

4000

5000

6000

7000

Low-latency AC

Energy-efficient AC

Fig. 3. Estimation of backlogged nodes for N = 10000 machine nodesin case of energy-efficient AC and low-latency AC with K = 64.

6.1 With C-PCD technique

First of all, we evaluate the performance of backlog estima-tion. Fig. 3 shows the pseudo-Bayesian estimation result forthe number of backlogged nodes in case of energy-efficientAC and low-latency AC, respectively, with K = 64. Weobserved that there exist a gap between the estimation andactual values, especially during the bursty period. However,the estimation algorithm keeps track of the actual valuesquite well during the traffic descent period. In particular,the backlog estimation under the low-latency AC is quiteaccurate.

Fig. 4 shows (a) access throughput, (b) average accessdelay, and (c) energy efficiency of the proposed low-latencyAC mechanism with the C-PCD technique. The proposedlow-latency AC mechanism achieves higher access through-put and energy efficiency, and lower average access delaythan those of the conventional low-latency AC mechanismwhen K ≤ 42 since it optimizes the ACB factor by consid-ering not only M but also K, while the conventional ACmechanism takes into account only M . According to theresult, both the proposed and the conventional low-latencyAC mechanism results in the same performances whenthe number of allocable PUSCH resources is sufficientlylarge. The performance gap between two mechanisms issignificant especially when K is small. For example, theaverage access delay of the proposed low-latency AC mech-anism is lower than 200ms, while the average access delayof the conventional low-latency AC mechanism is higherthan 300ms when K = 18. Furthermore, the EE of theproposed low-latency mechanism (0.80) is approximately 4times higher than the EE of the conventional low-latency ACmechanism (0.21) when K = 10. According to simulationstatistics even though they are not shown in the figure, whenK = 10, the proposed low-latency AC mechanism spends10983 PUSCH resources and 983 PUSCH resources arewasted due to preamble collisions, while the conventionallow-latency AC mechanism spends 16738 PUSCH resourcesand 6738 PUSCH resources are wasted due to preamblecollisions for accommodating 10000 nodes.

Fig. 5 shows (a) access throughput, (b) average access de-lay, and (c) energy efficiency of the proposed energy-efficientAC mechanism with the C-PCD technique for various targetEEs (θ = 0.9, 0.8, and 0.6). As shown in Fig. 5 (c), theproposed energy-efficient AC mechanism satisfies the targetEE regardless of K at the cost of increase of average accessdelay as shown in Fig. 5 (b). As expected, there exists atradeoff relationship between the EE and throughput/delayperformances. Note that the conventional AC mechanismonly yields the maximum EE of 0.38 when K ≥ 42. Forthe case that θ = 0.6, the proposed energy-efficient ACmechanism achieves higher EE than the target 0.6 whenK ≤ 26 since the proposed low-latency AC mechanismalready achieves higher EE than 0.6 as shown in Fig. 4 (c).When θ = 0.6 and K ≤ 38, the proposed energy-efficiencyAC mechanism outperforms the conventional AC mecha-nism in terms of all performance metrics such as accessthroughput, average access delay, and energy efficiency. Inaddition, when θ = 0.8 and K ≤ 20, the same performancetendency is observed.

Fig. 6 shows cumulative distribution functions (CDFs)of (a) the number of access reattempts and (b) the accessdelay of the proposed AC mechanisms when K = 20. Theproposed energy-efficient AC mechanism with the targetEE of 0.9 and 0.8 controls almost 89% and 78% of nodesto succeed in the RA in a single attempt. The averagenumber of reattempts is equal to 0.12 and 0.27 for givenEE targets of 0.9 and 0.8, respectively. Note that, withthe conventional low-latency AC mechanism, the averagenumber of reattempts is equal to 4.27 and only 21% of nodessucceeds in the RA in a single attempt. As shown in Fig. 6(b), nodes experience longer access delay with the proposedenergy-efficient AC mechanisms when θ = 0.9, comparedwith both the proposed low-latency AC mechanism andthe conventional low-latency AC mechanism. However, ifthe target EE is reduced from 0.9 to 0.8 with the proposedenergy-efficient AC mechanism, then the access delay issignificantly reduced from 1168 ms to 528 ms at the 90thpercentile.

Figs. 7 (a) and (b) show the average access delay and theenergy efficiency of the proposed comprehensive AC mech-anism with the C-PCD technique. Here, we set the accessratio γ to 0.5, which implies that 5000 nodes belong to thelow-latency access group and the other 5000 nodes belongto the energy-efficient access group. In particular, we setthe target energy efficiency θ to 0.9 for the energy-efficientgroup. Since the low-latency group and the energy-efficientgroup share all preamble and PUSCH resources together,the low-latency group under the proposed comprehensiveAC shows lower average access delay than that of the low-latency access only (all 10000 nodes attempt low-latencyaccess) as shown in Fig. 4 (b), and the energy-efficient groupunder the proposed comprehensive AC also shows loweraverage access delay than that of the energy-efficient accessonly (all 10000 nodes attempt energy-efficient access) asshown in Fig. 5 (b). In Fig. 7 (b), the energy-efficient groupunder the proposed comprehensive AC achieves the EE ofaround 0.8, which is lower than 0.9 of the energy-efficientaccess only as shown in Fig. 5 (c), whereas the low-latencygroup under the proposed comprehensive AC benefits con-siderably from the energy-efficient group in terms of EE, and



10

10 14 18 22 26 30 34 38 42 46 50 54 58 62

0

2

4

6

8

10

12

14

16

18

20

22

24

(a) Access throughput

10 14 18 22 26 30 34 38 42 46 50 54 58 62

0

100

200

300

400

500

600

700

800

(b) Average access delay

10 14 18 22 26 30 34 38 42 46 50 54 58 62

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

(c) Energy efficiency

Fig. 4. Performances of the proposed low-latency AC mechanism with the C-PCD technique for varying K.

10 14 18 22 26 30 34 38 42 46 50 54 58

0

2

4

6

8

10

12

14

16

18

20

22

24


10 14 18 22 26 30 34 38 42 46 50 54 58 62

0

100

200

300

400

500

600

700

800


10 14 18 22 26 30 34 38 42 46 50 54 58 62

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1


Fig. 5. Performances of the proposed energy-efficient AC mechanism with the C-PCD technique for varying K.

0 5 10 15 20 25

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

(a) Number of access reattempts

0 200 400 600 800 1000 1200 1400 1600

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

(b) Access delay (msec)

Fig. 6. Cumulative distribution function (CDF) of (a) the number of access reattempts and (b) the access delay in energy-efficient AC and low-latencyAC when K = 20.



11

thus, achieves higher EE than that of the low-latency accessonly as shown in Fig. 4 (c). According to the performanceresult, we observe that access delay is significantly sacrificedin order to achieve a very high energy efficiency duringbursty traffic period.

6.2 With E-PCD techniqueIt is worth noting that there exist no previous ACB tech-nique that exploits the E-PCD techniques. Thus, with theE-PCD technique, we only evaluate the performance of theproposed AC mechanism by changing its own parametersor compare the proposed low-latency AC mechanism to theproposed energy-efficient AC mechanism.

Fig. 8 shows the number of successful and unsuccessfulattempts of the proposed low-latency AC mechanism withthe E-PCD technique, which are denoted by S(p,K) in(29) and U(p,K) in (30), respectively, when K = 16 andN = 300. If p ≤ pLLearly, then S(p,K) > U(p,K). However,S(p,K) < U(p,K) as p approaches to pLLearly. There existsa tradeoff relationship between the access throughput andthe EE for varying the ACB factor. Hence, we define a novelACB factor as

p = εpLLearly + (1− ε)pLLearly (48)

for 0 ≤ ε ≤ 1.Fig. 9 shows (a) access throughput, (b) average access

delay, and (c) energy efficiency of the proposed low-latencyAC mechanism with the E-PCD technique for varying εof the ACB factor in (48). When ε = 1 and ε = 0, theACB factors becomes pLLearly in (26) and pLLearly in (28), respec-tively. As ε increases, the EE of the proposed mechanismdecreases while the access throughput increases and theaverage access delay decreases, but they are saturated oversome ε. All performances become improved as K increases.Furthermore, for a given K, we optimize ε so that the EEand the access throughput of the proposed low-latency ACmechanism are maximized while the average access delay isminimized, which is denoted by ε. For example, the accessthroughput is maximized and the average access delay isminimized for ε ∈ [0.6, 1] and the maximum EE is achievedwhen ε = 0.6 for ε ∈ [0.6, 1] when K = 16. Thus, inthis case, ε = 0.6. Then, for a given K, the ACB factor inthe sense that the maximum EE and the maximum accessthroughput are achieved while the average access delay isminimized is obtained by

pLLearly = εpLLearly + (1− ε)pLLearly. (49)

Fig. 10 shows (a) access throughput, (b) the averageaccess delay, and (c) the energy efficiency of the proposedlow-latency AC mechanism with the E-PCD technique forvarious ACB factors including pLLearly in (26) when ε = 0,pLLearly in (28) when ε = 0, and pLLearly with the best ε in(49). Even though the proposed low-latency AC mechanismwith pLLearly (ε = 0) results in a slightly longer averageaccess delay when K ≤ 24 than other cases, it significantlyimproves the EE. When K > 24, the proposed low-latencyAC mechanism with the E-PCD technique utilizes only pLLearlysince E[S|pLLearly = M

ν , ν] ≤ K.Fig. 11 shows (a) the access throughput, (b) the average

access delay, and (c) the energy efficiency of the proposed

energy-efficient AC mechanism with the E-PCD technique.As shown in the figure, the proposed energy-efficient ACmechanism effectively satisfies the target EE requirementseven though there exist a little dissatisfaction for some Kdue to approximation error in obtaining the ACB factor.

Fig. 12 shows the CDF of (a) the number of accessreattempts and (b) the access delay of the proposed energy-efficient AC mechanism with θ = 0.9, 0.8 and the pro-posed low-latency AC mechanism with the best ε. With theproposed energy-efficient AC mechanism, 63% and 50% ofnodes succeed the RA in a single attempt and the averagenumber of access reattempts is equal to 0.51 and 0.95 whenθ = 0.9 and 0.8, respectively. On the other hand, withthe proposed low-latency AC mechanism, 35% of nodessucceed the RA in a single attempt and the average numberof access reattempts is equal to 1.73. However, as shown inFig. 12 (b), at the 90th percentile, the access delays of theproposed energy-efficient AC mechanism with the E-PCDtechnique is equal to 325 ms and 270 ms for θ = 0.9 and0.8, respectively, while that of the proposed low-latency ACmechanism with the E-PCD technique is equal to 250 ms.

7 CONCLUSION

In this paper, we proposed a versatile access control mech-anism for massive IoT networks, which effectively satisfiesvarious QoS requirements: low-latency for mission-criticalservices and high energy efficiency (EE) for battery-poweredsensor applications. We adopted the early collision detectiontechnique as well as the conventional preamble collision de-tection technique at the first step of the random access (RA)procedure. For optimizing the access class barring factorof the proposed AC mechanism, we considered all radioresources which are involved in the overall RA procedure.The proposed low-latency AC mechanism achieves lowestaccess delay by aggressively allowing for IoT devices toattempt the RA at the cost of the decreased EE, while theproposed energy-efficient AC mechanism achieves a targetEE by restricting attempts from the IoT devices at the costof increased access delay. Compared to the conventionalAC mechanism, the proposed low-latency AC mechanismfurther reduces the access delay, while improving the EEunder diverse resource constraints. Furthermore, we pro-posed a comprehensive access control mechanism whichcan simultaneously support two groups of devices requiringlow-latency and ultra-low power consumption, respectively.

ACKNOWLEDGEMENT

This work was supported in part by the SUTD-ZJU ResearchCollaboration under Grant SUTD-ZJU/RES/01/2016, inpart by the Basic Science Research Program through theNRF funded by the Ministry of Science and ICT (NRF-2016R1A2B4014834), and in part by the National ResearchFoundation of Korea (NRF) grant funded by the Koreagovernment (MSIT) (NRF-2018R1C1B6008126).

REFERENCES

[1] C. X. Mavromoustakis, G. Mastorakis, and J. M. Batalla, Internet ofThings (IoT) in 5G mobile technologies. vol. 8. Springer, 2016.



12

10 14 18 22 26 30 34 38 42 46 50 54 58 62

0

100

200

300

400

500

600

(a) Average access delay (msec)

10 14 18 22 26 30 34 38 42 46 50 54 58 62

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

(b) Energy efficiency

Fig. 7. Performances of the proposed comprehensive AC mechanism with the C-PCD technique for varying K. The access ratio γ is set to 0.5.

0 0.1 0.2 0.3 0.4 0.5 0.6

0

20

40

60

80

100

120

140

160

Fig. 8. Average numbers of successful and unsuccessful attempts ofthe proposed low-latency AC mechanism with the E-PCD technique forvarying ACB factor p.

[2] Study on Scenarios and Requirements for Next Generation AccessTechnologies, 3GPP TR 38.913 V14.3.0, June 2017.

[3] Minimum requirements related to technical performance for IMT-2020radio interface(s), ITU-R M.2410-0, Nov. 2017.

[4] M. Li, R. Yu, P. Si, and Y. Zhang, “Energy-efficient machine-to-machine (M2M) communications in virtualized cellular networkswith mobile edge computing (MEC),” IEEE Trans. Mobile Comput.,2018 (to appear).

[5] H. Li, K. Ota, and M. Dong, “Learning IoT in edge: Deep learningfor the internet of things with edge computing,” IEEE Network,vol. 32, no. 1, pp. 96–101, Jan.-Feb. 2018.

[6] S.-C. Lin and K.-C. Chen, “Cognitive and opportunistic relay forQoS guarantees in machine-to-machine communications,” IEEETrans. Mobile Comput., vol. 15, no. 3, pp. 599–609, Mar. 2016.

[7] J. Huang, C. Xing, S. Y. Shin, F. Hou, and C. Hsu, “OptimizingM2M communications and quality of services in the IoT forsustainable smart cities,” IEEE Trans. on Sustain. Comput., vol. 3,no. 1, pp. 4–15, Jan.-Mar. 2018.

[8] A. Sivanathan, H. H. Gharakheili, F. Loi, A. Radford, C. Wije-nayake, A. Vishwanath, and V. Sivaraman, “Classifying IoT de-vices in smart environments using network traffic characteristics,”IEEE Trans. Mobile Comput., 2018 (to appear).

[9] T. P. de Andrade, C. A. Astudillo, and N. L. da Fonseca, “Alloca-tion of control resources for machine-to-machine and human-to-

human communications over LTE/LTE-A networks,” IEEE Inter-net Things J., vol. 3, no. 3, pp. 366–377, June 2016.

[10] T. P. de Andrade, C. A. Astudillo, L. R. Sekijima, and N. L.da Fonseca, “The random access procedure in long term evolutionnetworks for the internet of things,” IEEE Commun. Mag., vol. 55,no. 3, pp. 124–131, Mar. 2017.

[11] N. Xia, H.-H. Chen, and C.-S. Yang, “Radio resource managementin machine-to-machine communications-a survey,” IEEE Commun.Surveys Tuts., vol. 20, no. 1, pp. 791–828, First Quarter 2018.

[12] S. Duan, V. Shah-Mansouri, and V. W. S. Wong, “Dynamic accessclass barring for M2M communications in LTE networks,” in Proc.IEEE GLOBECOM, Dec. 2013, pp. 4747 – 4752.

[13] Z. Wang and V. W. Wong, “Joint access class barring and timingadvance model for machine-type communications.” in Proc. IEEEICC, June 2014, pp. 2357–2362.

[14] A. Laya, L. Alonso, and J. Alonso-Zarate, “Is the random accesschannel of LTE and LTE-A suitable for M2M communications? Asurvey of alternatives,” IEEE Commun. Surveys Tuts., vol. 16, no. 1,pp. 4–16, First Quarter 2014.

[15] M. Tavana, V. Shah-Mansouri, and V. W. Wong, “Congestioncontrol for bursty M2M traffic in LTE networks,” in Proc. IEEEICC, June 2015, pp. 5815–5820.

[16] Z. Wang and V. W. Wong, “Optimal access class barring for station-ary machine type communication devices with timing advanceinformation,” IEEE Trans. Wireless Commun., vol. 14, no. 10, pp.5374–5387, Oct. 2015.

[17] S. Duan, V. Shah-Mansouri, Z. Wang, and V. W. Wong, “D-ACB:Adaptive congestion control algorithm for bursty M2M traffic inLTE networks,” IEEE Trans. Veh. Technol., vol. 65, no. 12, pp. 9847–9861, Dec. 2016.

[18] H. Jin, W. Toor, B. C. Jung, and J. B. Seo, “Recursive Pseudo-Bayesian access class barring for M2M communications in LTEsystems,” IEEE Trans. Veh. Technol., vol. 66, no. 9, pp. 8595–8599,Sept. 2017.

[19] S. Vural, N. Wang, P. Bucknell, G. Foster, R. Tafazolli, and J. Muller,“Dynamic preamble subset allocation for RAN slicing in 5G net-works,” IEEE Access, vol. 6, pp. 13 015–13 032, 2018.

[20] C. Y. Ho and C.-Y. Huang, “Energy-saving massive access controland resource allocation schemes for M2M communications inOFDMA cellular networks,” IEEE Wireless Commun. Lett., vol. 1,no. 3, pp. 209–212, June 2012.

[21] H. S. Dhillon, H. C. Huang, H. Viswanathan, and R. A. Valenzuela,“Power-efficient system design for cellular-based machine-to-machine communications,” IEEE Trans. Wireless Commun., vol. 12,no. 11, pp. 5740–5753, Nov. 2013.

[22] A. Azari and G. Miao, “Energy efficient mac for cellular-basedM2M communications,” in Proc IEEE Global Conference on Signaland Information Processing (GlobalSIP), Dec. 2014, pp. 128–132.

[23] G. Miao, A. Azari, and T. Hwang, “E2-MAC: Energy efficientmedium access for massive M2M communications,” IEEE Trans.Commun., vol. 64, no. 11, pp. 4720–4735, Nov. 2016.



13

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

2

4

6

8

10

12

14

16

18

20


0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

50

100

150

200

250

300

350


0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1


Fig. 9. Performances of the proposed low-latency AC mechanism with the E-PCD technique for varying ε when K = 10, 16, and 20.

10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40

0

2

4

6

8

10

12

14

16

18

20

22

24


10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40

0

50

100

150

200

250

300

350

400


10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1


Fig. 10. Performances of the proposed low-latency AC mechanism with the E-PCD technique for varying K.

10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40

0

2

4

6

8

10

12

14

16

18

20

22

24


10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40

0

50

100

150

200

250

300

350

400


10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1


Fig. 11. Performances of the proposed energy-efficient AC mechanism with the E-PCD technique for varying K.

[24] A. Aijaz, M. Tshangini, M. R. Nakhai, X. Chu, and A.-H. Aghvami,“Energy-efficient uplink resource allocation in LTE networks withM2M/H2H co-existence under statistical QoS guarantees,” IEEETrans. Commun., vol. 62, no. 7, pp. 2353–2365, July 2014.

[25] M. A. Mehaseb, Y. Gadallah, A. Elhamy, and H. Elhennawy,“Classification of LTE uplink scheduling techniques: An M2Mperspective,” IEEE Commun. Surveys Tuts., vol. 18, no. 2, pp. 1310–1335, Second Quarter 2016.

[26] D. Zhang, Z. Zhou, S. Mumtaz, J. Rodriguez, and T. Sato, “One in-tegrated energy efficiency proposal for 5G IoT communications,”IEEE Internet Things J., vol. 3, no. 6, pp. 1346–1354, Dec. 2016.

[27] Z. Yang, W. Xu, Y. Pan, C. Pan, and M. Chen, “Energy efficientresource allocation in machine-to-machine communications withmultiple access and energy harvesting for IoT,” IEEE InternetThings J., pp. 229–245, Feb. 2018.

[28] A. Azari and G. Miao, “Network lifetime maximization for

cellular-based M2M networks,” IEEE Access, vol. 5, pp. 18 927–18 940, 2017.

[29] N. M. Balasubramanya, L. Lampe, G. Vos, and S. Bennett, “Ontiming reacquisition and enhanced primary synchronization sig-nal (ePSS) design for energy efficient 3GPP LTE MTC,” IEEE Trans.Mobile Comput., vol. 16, no. 8, pp. 2292–2305, Aug. 2017.

[30] H. S. Jang, S. M. Kim, H.-S Park, and D. K. Sung, “An earlypreamble collision detection scheme for cellular M2M randomaccess,” IEEE Trans. Veh. Technol., vol. 66, no. 7, pp. 5974–5984,July 2017.

[31] ——, “A preamble collision resolution scheme via tagged pream-bles for cellular IoT/M2M communications,” IEEE Trans. Veh.Technol., vol. 67, no. 2, pp. 1825–1829, Feb. 2018.

[32] Study on RAN improvements for machine-type communications, 3GPPTR 37.868 V11.0.0, Oct. 2011.

[33] Medium access control (MAC) protocol specification (Release 14), 3GPP



14

0 2 4 6 8 10 12 14 16

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

(a) Number of access reattempts

0 50 100 150 200 250 300 350 400 450 500

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

(b) Access delay (msec)

Fig. 12. Cumulative distribution function (CDF) of (a) the number of access reattempts and (b) the access delay of the proposed AC mechanismswith the E-PCD technique when K = 20.

TS 36.321 V14.6.0, Mar. 2018.[34] S. Sesia, I. Toufik, and M. Baker, LTE - The UMTS Long Term

Evolution: From Theory to Practice. John Wiley & Sons Ltd., 2009.[35] H. S. Jang, B. C. Jung, and D. K. Sung, “Dynamic access control

with resource limitation for group paging-based cellular IoT sys-tems,” IEEE Internet Things J., vol. 5, no. 6, pp. 5065–5075, Dec.2018.

[36] N. Zhang, G. Kang, J. Wang, Y. Guo, and F. Labeau, “Resourceallocation in a new random access for M2M communications,”IEEE Commun. Lett., vol. 19, no. 5, pp. 843–846, May 2015.

[37] D. Magrin, C. Pielli, C. Stefanovic, and M. Zorzi, “EnablingLTE RACH collision multiplicity detection via machine learning,”arXiv preprint arXiv:1805.11482, 2018.

[38] [70bis#11]-LTE: MTC LTE simulations, 3GPP 3GPP TSG-RAN WG2meeting# 71 R2-104663, Aug. 2010.

[39] W. Feller, An Introduction to Probability Theory and Its Application.John Wiley & Sons Ltd., 1968.

[40] R. Rivest, “Network control by bayesian broadcast,” IEEE Trans.Inf. Theory, vol. IT-33, no. 3, pp. 323–328, May 1987.

[41] R. M. Corless, G. H. Gonnet, D. E. Hare, D. J. Jeffrey, and D. E.Knuth, “On the LambertW function,” Advances in Computationalmathematics, vol. 5, no. 1, pp. 329–359, 1996.

Han Seung Jang (S’14–M’18) received the B.Sdegree in Electronics and Computer Engineer-ing from Chonnam National University, Gwangju,Korea, in 2012, and the M.S. and Ph.D. degreesin Electrical Engineering from Korea AdvancedInstitute for Science and Technology (KAIST),Daejoen, Korea in 2014 and 2017, respectively.

From May 2018 to February 2019, he was aPost-Doctoral Fellow with the Information Sys-tems Technology and Design (ISTD) Pillar, Sin-gapore University of Technology and Design

(SUTD), Singapore. He was also a Post-Doctoral Fellow from Septem-ber 2017 to April 2018 with Chungnam National University, Daejoen,Korea. He is currently an Assistant Professor with the School of Elec-trical, Electronic Communication, and Computer Engineering, ChonnamNational University, Yeosu, Korea. His research interests include cellu-lar Internet-of-things (IoT)/machine-to-machine (M2M) communications,machine learning, smart grid, and energy ICT.

Hu Jin (S’07–M’12–SM’18) received the B.E.degree in electronic engineering and informationscience from the University of Science and Tech-nology of China, Hefei, China, in 2004, and theM.S. and Ph.D. degrees in electrical engineeringfrom the Korea Advanced Institute of Scienceand Technology (KAIST), Daejeon, South Korea,in 2006 and 2011, respectively. From 2011 to2013, he was a Post-Doctoral Fellow with TheUniversity of British Columbia, Vancouver, BC,Canada. From 2013 to 2014, he was a Research

Professor with Gyeongsang National University, Tongyeong, South Ko-rea. Since 2014, he has been with the Division of Electrical Engineering,Hanyang University, Ansan, South Korea, where he is an AssociateProfessor. His research interests include medium-access control andradio resource management for random access networks and schedul-ing systems considering advanced signal processing and queueingperformance.



15

Bang Chul Jung (S’02–M’08–SM’14) receivedthe B.S. degree in Electronics Engineering fromAjou University, Suwon, Korea, in 2002 and theM.S. and Ph.D. degrees in Electrical & Com-puter Engineering from KAIST, Daejeon, Korea,in 2004 and 2008, respectively. He was a seniorresearcher/research professor with KAIST Insti-tute for Information Technology Convergence,Daejeon, Korea, from January 2009 to February2010. From March 2010 to August 2015, he wasa Faculty of Gyeongsang National University. He

is currently an Associate Professor of the Department of ElectronicsEngineering, Chungnam National University, Daejeon, Korea. He hasserved as Associate Editor of IEICE Transactions on Fundamentals ofElectronics, Communications, and Computer Sciences since 2018. Hisresearch interests include wireless communication systems, internet-of-things (IoT) communications, statistical signal processing, informationtheory, interference management, random access, radio resource man-agement, cooperative relaying techniques, in-network computation, andmobile computing.

Dr. Jung was the recipient of the Fifth IEEE Communication SocietyAsia-Pacific Outstanding Young Researcher Award in 2011 and he hasbeen selected as a winner of Haedong Young Scholar Award in 2015,which is sponsored by the Haedong foundation and given by Korea Insti-tute of Communication and Information Science (KICS). He was also therecipient of the Bronze Prize of Intel Student Paper Contest in 2005, theFirst Prize of KAIST’s Invention Idea Contest in 2008, the Bronze Prizeof Samsung Humantech Paper Contest in 2009, the Best Paper Award inSpring Conference of Korea Institute of Information and CommunicationEngineering (KIICE) in 2015, the Best Paper Award in the Institute ofElectronics and Information Engineering (IEIE) Symposium in 2017, andthe Best Paper Award in Winter Conference of KICS in 2018.

Tony Q.S. Quek (S’98-M’08-SM’12-F’18) re-ceived the B.E. and M.E. degrees in electri-cal and electronics engineering from the TokyoInstitute of Technology, Tokyo, Japan, in 1998and 2000, respectively, and the Ph.D. degreein electrical engineering and computer sciencefrom the Massachusetts Institute of Technology,Cambridge, MA, USA, in 2008. Currently, he is atenured Associate Professor with the SingaporeUniversity of Technology and Design (SUTD).He also serves as the Acting Head of ISTD Pillar

and the Deputy Director of the SUTD-ZJU IDEA. His current researchtopics include wireless communications and networking, internet-of-things, network intelligence, wireless security, and big data processing.

Dr. Quek has been actively involved in organizing and chairing ses-sions, and has served as a member of the Technical Program Commit-tee as well as symposium chairs in a number of international confer-ences. He is currently an elected member of IEEE Signal ProcessingSociety SPCOM Technical Committee. He was an Executive EditorialCommittee Member for the IEEE TRANSACTIONS ON WIRELESS COM-MUNICATIONS, an Editor for the IEEE TRANSACTIONS ON COMMUNICA-TIONS, and an Editor for the IEEE WIRELESS COMMUNICATIONS LET-TERS. He is a co-author of the book “Small Cell Networks: Deployment,PHY Techniques, and Resource Allocation” published by CambridgeUniversity Press in 2013 and the book “Cloud Radio Access Networks:Principles, Technologies, and Applications” by Cambridge UniversityPress in 2017.

Dr. Quek was honored with the 2008 Philip Yeo Prize for OutstandingAchievement in Research, the IEEE Globecom 2010 Best Paper Award,the 2012 IEEE William R. Bennett Prize, the 2015 SUTD OutstandingEducation Awards – Excellence in Research, the 2016 IEEE Signal Pro-cessing Society Young Author Best Paper Award, the 2017 CTTC EarlyAchievement Award, the 2017 IEEE ComSoc AP Outstanding PaperAward, and the 2016-2018 Clarivate Analytics Highly Cited Researcher.He is a Distinguished Lecturer of the IEEE Communications Society anda Fellow of IEEE.

Documents

Versatile Access Control for Massive IoT: Throughput ...wicl.cnu.ac.kr/wordpress/DB/International_Journal/2019_04_Versatile... · ments including access throughput, access delay,