Load Estimation in a Two-Priority mMTC Random

Access Channel
Ahmed O. Elmeligy Ioannis Psaromiligkos Au Minh
McGill University McGill University Hydro-Quebéc Research Institute (IREQ)
[email protected] [email protected] [email protected]

Abstract—The use of cellular networks for massive machine- joint probability of the number of successful and collided
arXiv:2405.14667v2 [eess.SP] 8 Oct 2024

type communications (mMTC) is an appealing solution due to UE preamble transmissions within a RACH slot is derived
the wide availability of cellular infrastructure. in [6]. Maximum likelihood (ML) estimation and Bayesian
Estimating the number of devices (network load) is vital
for efficient allocation of the available resources, especially for techniques are used to estimate the network load. Two algo-
managing the random access channel (RACH) of the network. rithms are proposed based on whether the number of successes
This paper considers a two-priority RACH and proposes two and collisions are known at the BS or only the number of
network load estimators: a maximum likelihood (ML) estimator successes is known. However, the probability of unselected
and a reduced complexity (RCML) variant. The estimators are preambles is not considered, which is additional information
based on a novel model of the random access behavior of the
devices coupled with a flexible analytical framework to calculate known to the BS and can be used to improve the estimation
the involved probabilities. Monte Carlo simulations demonstrate accuracy.
the accuracy of the proposed estimators for different network A combinatorial model presented in [4] investigates the
configurations. Results depict increased estimation accuracy us- transient behavior of RACHs with bursty arrivals. The model
ing non-uniform preamble selection probabilities compared to the obtains the average number of successful and collided UEs
common uniform probabilities at no extra computational cost.
Index Terms—Load estimation, massive machine-type commu- in a single RACH slot. The authors then estimate the number
nications (mMTC), random access channel (RACH). of contending UEs in each slot. Simulations show that the
estimation is accurate in the case of a high UE count but
I. I NTRODUCTION suffers for few UE.
The use of already deployed cellular infrastructure is The number of active UEs in
an appealing solution to providing wireless connectivity irregular repetition slotted ALOHA (IRSA) access protocol
for massive machine-type communication (mMTC) applica- is estimated in [7]; successive interference cancellation (SIC)
tions [1]. At the same time, combining multi-priority applica- is adopted at the receiver to recover the transmitted packets.
tions in a single network is a promising cost-efficient approach The authors propose a maximum a posteriori (MAP)
as it avoids deploying multiple networks. However, servicing detector for the number of unrecovered UEs in a specific
the massive number of devices in mMTC poses a significant SIC iteration. A sub-optimal detector is also proposed to
challenge to the cellular network, particularly, by congesting reduce the computational complexity via approximations.
and overloading the random access channel (RACH) [2], [3]. Numerical results illustrate that the suboptimal detector’s
The RACH, which is usually available periodically, is the mean absolute error (MAE) increases with the overloading.
channel the devices use to request access to the network [4]. In [8], an estimate of the network load is used in two
Each time the RACH is available is known as a RACH dynamic access class barring (ACB) algorithms that operate
slot, and at each slot, a device (henceforth referred to as a without the BS knowing the number of devices. The first
user equipment (UE)) randomly chooses a preamble from a algorithm optimizes the ACB factor for a fixed number of
finite set of preambles and transmits it to the base station (BS); preambles, where the ACB factor determines the backoff time
if a preamble is selected by only one device, the BS can of collided devices. The second algorithm optimizes the ACB
successfully decode it. On the other hand, if two or more factor and the number of available preambles.
devices choose the same preamble a collision happens [5]. Finally, a backoff scheme is developed by [9] that uses a
The network load is defined as the number of UEs the net- backoff indicator to determine a random waiting period for the
work serves, and the ratio of UEs to the number of preambles collided UEs. The backoff indicator is dynamically adjusted
is known as the overloading factor; a high overloading factor in [10] based on the availability of resources and the number
indicates that the network is congested, thus increasing the of backlogged UEs.
number of collisions and reducing the network throughput. In all these works, only a single UE priority class is
Knowing the load is vital for the network operator to decide considered; hence superimposing numerous applications with
how to allocate the available finite resources to the UE pool. varying quality of service (QoS) requirements in the same
Several methods have been proposed to estimate the net- network to save resources would not be possible. Additionally,
work load in the RACH. A closed-form expression for the only a single RACH slot is used, and the estimation techniques
are not extended to multiple RACH slots, which could signif- Resource
Block (RB)
icantly improve the estimator’s accuracy.
Several additional works consider single and multi-UE pri-
π1 π2 π3 πT
ority classes, focusing on improving the RACH performance      
by first introducing a performance metric, and then optimizing M  h  x  h   φ 
 h  x  l 
M −1      h 
the RACH parameters, such as the number of preambles, the  
M − 2  l  h  φ   x 
     
 ..  ..  ..   .. 
    
RACH periodicity, or the ACB factor, to maximize the met- ..  
.  .  .  .   . 
ric [11]–[15]. However, no network load estimation techniques 1

l

φ

x
 
h


are proposed, and the number of contending UEs is assumed

to be known. 1 2 3 T
Time
Slot
In this paper, we propose a load estimation method that
addresses the above shortcomings. Specifically, the main con- Fig. 1. Example of a pattern sequence over T time slots.
tributions of this paper are as follows:
chooses an RB. The probability that a H-UE chooses RB i
• We present a novel approach to model a two-priority
is phi ; likewise, the probability that an L-UE chooses RB i
RACH, which allows us to define access patterns that
is pli . Let ph = [ph1 , . . . , phM ] be the vector collecting the RB
describe the random access behavior of UEs as observed
selection probabilities for an H-UE, while pl = [pl1 , . . . , plM ]
by the BS.
be
PMthe corresponding PMvector for an L-UE. Clearly, we have
• We develop an analytical framework to calculate the h l
i=1 i p = 1 and i=1 ip = 1.
probability of observing a given access pattern.
For each of the M RBs, one of the four events listed below
• Our approach allows for non-uniform preamble selection
may occur:
probabilities. To our knowledge, the literature lacks a
model that considers non-uniform preamble selection, 1) Event h: A single H-UE selects the RB.
which can provide greater flexibility when allocating 2) Event l: A single L-UE selects the RB.
resources to different UE priority classes. 3) Event φ: No UE selects the RB, i.e., the RB is unoccu-
• We propose an ML estimator of the network load from pied or empty.
observed access patterns over multiple RACH slots. 4) Event x: More than one UE selects the RB, i.e., a
• We formulate a reduced complexity maximum likelihood (RCML)collision occurs.
estimator that discards a portion of the available Each of these events can be detected by the BS [16]. However,
information at the BS. in the case of a collision, the BS does not know how many UEs
• Our results show an increase in estimation accuracy are involved [17]–[19] nor their priority. Thus, at each time
using non-uniform preamble selection probabilities at no slot t, the BS observes an access pattern, πt ∈ {h, l, φ, x}M ,
additional computational cost relative to the traditional that is the sequence of events that occur across the M RBs;
uniform approach. Fig. 1 shows an example of such patterns observed over T
Monte Carlo simulation results validate the analytical frame- time slots.
work and showcase the proposed estimators’ accuracy for The problem we consider in this paper is to estimate the net-
various system setups. work load defined as the number of H-UEs, nh , and L-UEs, nl ,
The rest of this paper is organized as follows: Section II by observing a sequence of T access patterns, π1 , π2 , . . . , πT .
presents the system model that formulated the observations of We assume that the preamble selection probability vectors
the RACH at the BS as patterns. The ML and RCML esti- ph and pl are known at the BS, and that π1 , π2 , . . . , πT
mators are developed in Section III along with the algorithm are independent and identically distributed (i.i.d.). This is a
that computes the pattern probabilities. Section IV presents common assumption [7], meant to simplify the system model.
the simulation setup and numerical results that showcase the Extension of the model to the non-i.i.d. case is possible
estimators’ accuracies. Finally, Section V concludes the paper. through the use of conditional probabilities. Such extension,
however, relies heavily on the implementation of the RACH
II. S YSTEM M ODEL AND P ROBLEM S TATEMENT
protocol, e.g., number of retries, back-off time, and is beyond
We consider a system model with nh the scope of this paper.
high-priority user equipments (H-UEs) and nl
low-priority user equipments (L-UEs). The H-UEs and III. P ROPOSED M ETHOD
L-UEs differ in QoS requirements, whereas the H-UEs have
stricter requirements, such as lower latency and higher data This section is divided into thee parts. First, we propose
rates. Note that nh and nl are fixed and unknown to the BS. a ML estimation technique to estimate the network load
Transmitting preambles over the RACH can be viewed as given the observed patterns π1 , π2 , . . . , πT . As the proposed
having a finite number of resource blocks (RBs) randomly algorithm requires the evaluation of the probabilities of the
accessed by the pool of UEs. Time is discretized into slots, observed patterns, in Section III-B we present an algorithm
indexed by t, and each slot is divided into M RBs. We to compute these probabilities as a function of the number of
assume that at slot t, each of the n = nh + nl UEs randomly H-UEs and L-UEs. Finally, we develop an RCML estimator
    
Next, we derive the probabilities in the
 
 l   L1 1 2, 6, 8 =H
   
right hand side (RHS) of (3).

h H2 2
   
  







 The first factor in (3), P H , is the probability that each RB



x 





X3 3

   in H is occupied by one H-UE. Let H = {h1 , h2 , . . . , hH }.


 φ





 φ4 4


1 =L Then we have:
     

 x 


 X5 5

4, 7 =Φ

P H = P (Hh1 )P (Hh2 |Hh1 ) × · · · ×
   
h H6 6
   
  
   


 φ





 φ7

7
 P (HhH |Hh1 , . . . , HhH−1 ) (4)
     

h
 
H8

8 3, 5 =X The probabilities in (4) can be calculated using the follow-
ing proposition.
Fig. 2. Example pattern of length M = 8. In slots 2, 6, and 8 we
have collision-free H-UE transmissions, denoted as events H2 , H6 and H8 ; Proposition 1. The first factor in the RHS of (4) is given by:
hence, H = {2, 6, 8}. Similarly, the event in slot 1 a collision-free L-UE h l
−1
transmission, the events in slots 4 and 7 no transmissions, and the events in P (Hh1 ) = nh phh1 (1 − phh1 )n (1 − plh1 )n (5)
slots 3 and 5 signify collisions.
The rest of the factors are given by:
that discards a portion of the available information at the BS
to reduce the computational complexity. P (Hhk |Hh1 , . . . , Hhk−1 ) =
h l
A. Maximum Likelihood Estimation Nkh p̂hhk (1 − p̂hhk )Nk −1 (1 − p̂lhk )n (6)
h l
Let us denote by P ( · ; n , n ) the probability of observing where Nkh , p̂hhk , and p̂lhk are calculated as follows:
a given pattern or sequence of patterns, which, of course,
depends on nh and nl . Then, the ML estimator is given by: Nkh = nh − (k − 1) (7)
phhk
n̂h , n̂l = arg max P (π1 , π2 , . . . , πT ; nh , nl ) p̂hhk = (8)
phhk + . . . + phH + phi
P
nh ,nl i∈L,Φ,X
T plhk
p̂lhk =
Y
= arg max P (πt ; nh , nl ) (1) (9)
plhk + . . . + plH + pli
P
nh ,nl t=1 i∈L,Φ,X


where the last equality holds due to the i.i.d. assumption on The proof of Proposition 1 is given in Appendix A. P H
π1 , π2 , . . . , πT . can be calculated recursively using Algorithm 1 also given in
Appendix B.
B. Derivation of Pattern Probabilities Following the same procedure, the second factor in (3),


We start with an alternative representation of an access P L H , is the probability that each RB in L is occupied
pattern, π. Note that the subscript t and the dependence on nh by one L-UE given that the RBs in H are unavailable. Let
and nl are omitted throughout this section to avoid overloading L = {l1 , l2 , . . . , lL }. Then we have:
the notation. Define Hi as the event that one H-UE occupies



RB i. Similarly, define Li , Φi , and Xi as the events that P L H = P L l1 H P L l2 L l1 , H × · · · ×


one L-UE occupies RB i, no UE occupies RB i (RB i is P LlL Ll1 , . . . , LlL−1 , H (10)
empty), and two or more UEs occupy RB i (collision occurs),
The probabilities in (10) are calculated using Proposition 2
respectively. Let set H contain the indices of the RBs that
.
contain only one H-UE. Likewise, sets L, Φ, and X contain the
indices of the RBs that contain only one L-UE, are empty, and Proposition 2. The first factor of the RHS of (10) is given
have collisions, respectively. Hence, a pattern can be written by:
l h
P Ll1 H = nl p̂ll1 (1 − p̂ll1 )n −1 (1 − p̂hl1 )N


as the intersection of the events Hi for i ∈ H, Li for i ∈ L, (11)
Φi for i ∈ Φ, and Xi for i ∈ X :
The rest of the factors are given by:
l h
π = H, L, Φ, X (2) P Llk Ll1 , . . . , Llk−1 , H = Nkl p̂llk (1 − p̂llk )Nk −1 (1 − p̂hlk )N


T T T (12)
T H = i∈H Hi , L = i∈L Li , Φ = i∈Φ Φi , and
where
where N h , Nkl , p̂hlk , and p̂llk are as follows:
X = i∈X Xi In Fig. 2, we show an example for a pattern
of length M = 8 along with the corresponding sequence of N h = nh − H, Nkl = nl − (k − 1) (13)
events Hi , Li , Φi , and Xi , as well as the index sets H, L, Φ,
phlk
and X . p̂hlk = (14)
phlk + . . . + phL + i∈Φ,X phi
P
From (2), we can decompose P (π) as follows:
pllk
p̂llk = (15)
pll1 + . . . + plL + i∈Φ,X pli




P
P (π) = P H P L H P Φ H, L P X H, L, Φ (3)
 The proof of Proposition 2 follows the same procedure as where Njh , Njl , p̂hxj and p̂lxj are as follows:
Proposition 1 and is omitted due to lack

of space. Algorithm 2
in Appendix B shows how P L H is calculated. Njh = nh − H − ij−1 − (j − 1), with i0 = 0 (23)

Njl l
= n − L − (kj−1 − ij−1 ) − (j − 1), with k0 = i0 = 0
Moving on to the third factor in (3), P Φ H, L is the (24)
probability that no UEs occupy Φ given that the RBs in H
phxj plxj
and L are unavailable. Let Φ = {φ1 , φ2 , . . . , φΦ }. Then we p̂hxj = , p̂lxj = (25)
have: phxj + . . . + phX plxj + . . . + plX



The proof of Proposition 4 is omitted due to lack of space.
P Φ H, L = P Φφ1 H, L P Φφ2 Φφ1 , H, L × · · · ×


Algorithm 3 in Appendix B shows how P X H, L, Φ is
P ΦφΦ Φφ1 , . . . , ΦφΦ−1 , H, L (16) calculated.
C. Reduced Complexity Maximum Likelihood Estimation
The probabilities in (16) are calculated using Proposition 3
the proof of which is omitted due to lack of space. The ML estimator in (1) is computationally
complex to
implement. This is mainly due to P X H, L, Φ in (3) having
a large number of combinations for certain values of X , N1h ,
Proposition 3. The first factor of the RHS of (16) is given
and N1l . Thus,
we propose an RCML estimator that discards
by:
h l P X H, L, Φ and relies on the remaining factors in (1) to
P Φφ1 H, L = (1 − p̂hφ1 )N (1 − p̂lφ1 )N


(17)
estimate the network load. In other words, the probability of
The rest of the factors are given by: an access pattern π is approximated as follows:
h l





P Φφk Φφ1 , . . . , Φφk−1 , H, L = (1 − p̂hφk )N (1 − p̂lφk )N P (π) ≈ P̂ (π) = P H, L, Φ = P H P L H P Φ H, L

(18) (26)
where N h , N l , p̂hlk , and p̂llk are as follows: The RCML estimator is then obtained by (1) with P̂ (π) in
place of P (π).
N h = nh − H, Nkl = nl − L (19)
IV. S IMULATION R ESULTS
phφk
p̂hφk = (20) In this section, we evaluate the performance of the proposed
phφk + . . . + phΦ + phi
P
i∈X estimators in terms of the MAE for varying numbers of
plφk UEs and the number of patterns observed. We consider three
p̂lφk = (21)
plφk + . . . + plΦ + pli simulation setups. In the first setup, all UEs are assumed
P
i∈X
to be in the same priority class; hence, ph = pl . This
The algorithm that calculates the third factor in (3) is similar scenario replicates the RACH in LTE and the contention-
to Algorithm 2, hence it has been omitted for brevity. based RACH procedure in 5G, with the RBs representing

preambles in the random access procedure [20], [21]. In the
Finally, the last factor in (3), P X H; L Φ , is the proba- second setup, the UEs are split into two priority classes, with
bility that two or more UEs occupy each RB in X . Let X =
the H-UEs having more RBs available than the L-UEs. This
{x1 , x2 , . . . , xX }. Proposition 4 depicts how P X H, L, Φ setup corresponds to virtually allocating the RBs to the UEs
is calculated. depending on their priority class [22], [23]. Finally, the third
setup utilizes non-uniform RB access probabilities introducing
a generalized approach to the RACH protocol, thereby opening


Proposition 4. P X H, L, Φ is given by:

up opportunities for integrating different priority classes more
P X H, L, Φ = effectively. The results presented are averages of 50 Monte
N1h +N1l h
NX l
+NX N1h NXh Carlo simulations. In all simulations, the number of RBs is
X
···
X X
···
X set to M = 6, T is set to 1, 3, or 10, and the number of
k1 =2 kX =2 i1 =0 iX =0
L-UEs ranges from 0 to 7.
X  h A. Single Priority Class
Y N h
(p̂hxj )ij (1 − p̂hxj )N −ij ×
i j In the first simulation, nh is fixed to 2, while the elements in
j=1
p and pl are set to 1/M , indicating that all RBs are equally
h
Nl
 
l
(p̂lxj )kj −ij (1 − p̂lxj )N −kj +ij (22a) likely to be chosen by any UE. The MAE of the ML and
kj − ij RCML estimators against nl are shown in Fig. 3. As we can
see, on average, the ML estimator (Fig. 3a) outperforms the
s.t. km − im ≥ 0, m∈X (22b) RCML estimator (Fig. 3b) due to the RCML not using all the
X
X X
X available pattern information. As expected, the MAE decreases
im = N h , km = N l − N h (22c) as T increases due to the extra information available to the
m=1 m=1 BS. Furthermore, for a fixed T , an increase in the overloading
 5 5 6 6
T =1 T =1 T = 1, nh = 1 T = 1, nh = 1
T =3 T =3 T = 3, nh = 1 T = 3, nh = 1
4 T = 10 4 T = 10 T = 10, nh = 1 T = 10, nh = 1
4 T = 1, nh = 2 4 T = 1, nh = 2
3 3
MAE

MAE

MAE

MAE
T = 3, nh = 2 T = 3, nh = 2
T = 10, nh = 2 T = 10, nh = 2
2 2
2 2
1 1

0 0 0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Number of L-UEs (nl ) Number of L-UEs (nl ) Number of L-UEs (nl ) Number of L-UEs (nl )
(a) (b) (a) (b)
Fig. 3. MAE vs. nl for ML (3a) and RCML (3b) estimators for different T . Fig. 5. MAE vs. nl for ML (5a) and RCML (5b) estimators for different T
and nh .
·10−3
8
0
Estimated number of H-UEs (n̂h )

6 6
T = 1, nh = 1 T = 1, nh = 1
1 T = 3, nh = 1 T = 3, nh = 1

2 6 T = 10, nh = 1 T = 10, nh = 1
4 T = 1, nh = 2 4 T = 1, nh = 2

MAE

MAE
T = 3, nh = 2 T = 3, nh = 2
3 T = 10, nh = 2 T = 10, nh = 2

4 4 2 2
5
6 0 0
2 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
7 Number of L-UEs (nl ) Number of L-UEs (nl )

8 (a) (b)
0
0 1 2 3 4 5 6 7 8 Fig. 6. MAE vs. nl
for ML (6a) and RCML (6b) estimators for different T
Estimated number of L-UEs (n̂l ) and nh using non-uniform ph and pl .

C. Two Priority Classes with Non-Uniform RB Allocation

Fig. 4. Likelihood heatmap for nh = 2, nl = 4, and T = 1.
In the third simulation, we set ph =
factor leads to a higher number of collisions resulting in a [1/12, 1/12, 2/12, 2/12, 3/12, 3/12] and p l
=
higher MAE. A high collision number in the patterns results [4/12, 3/12, 2/12, 1/12, 1/12, 1/12]. Fig. 6 showcases
in less useful information for the estimators, leading to more the MAE for the estimators, which follow a similar trend to
errors. To further drive this point, consider a pattern consisting the results of Section IV-B. We note that for a relatively
of only collisions; we cannot distinguish between the H-UEs low nl , the RCML estimator performs similarly compared
and L-UEs. On the other hand, if a pattern consists of only to the ML estimator but starts to diverge as nl increases.
non-collisions, we can exactly determine the number of H-UEs Hence the last factor in (3) becomes more significant as the
and L-UEs. network load increases. Additionally, the RCML estimator
Finally, to illustrate the estimators’ operation, Fig. 4 shows is significantly faster than the ML estimator; on average,
a heatmap of the likelihood function for nh = 2, nl = 4, across all three simulations, the RCML estimator was 46
and T = 1. The heatmap shows that the likelihood function is times faster than the ML estimator. Furthermore, there is a
maximized for the above parameters when n̂h = 2 and n̂l = 4, tradeoff between the estimators’ accuracy and the time taken
which are the true number of H-UEs and L-UEs. It is noted to estimate the load; the larger T is, the higher the accuracy.
that the T = 1 scenario is similar to other load estimation Finally, comparing the results in Fig. 5 and Fig. 6 we see
work that utilize one RACH slot in their formulations [6], [7]. that utilizing non-uniform RB selection probabilities leads to
increased load estimation accuracy. It is important to note that
the computational complexity in both cases is the same.
B. Two Priority Classes V. C ONCLUSION
In the second simulation, nh is set to 1 or 2. Further- In this paper, we proposed two network load estimators
more, the elements in ph are fixed to 1/M , while pl = for a two-priority mMTC RACH that operate by observing
[1/3, 1/3, 1/3, 0, 0, 0], indicating that the H-UEs have twice multiple RACH slots. First, we modelled the RACH slots by
as many RBs available as the L-UEs. Fig. 5 shows the MAE RB access patterns observed by the BS. Then, we formulated
against nl for different values of nh and T for both estimators. the likelihood of the number of high and low priority UEs
For a fixed nh , the MAE decreases as T increases. On the given the observed access patterns. Based on the formulated
other hand, for a fixed T , the MAE increases as nh increases likelihood function, we developed a ML estimator and a
since the overloading factor increases. RCML estimator, where the latter approximated the likelihood
function by discarding a portion of the pattern information. Thus RB h1 can no longer be chosen by any other UE, i.e.,
Finally, we conducted numerical simulations to evaluate the phh1 = plh1 = 0; we have:
performance of the proposed estimators. The results showed  h
N2 h h
that utilizing non-uniform RB selection probabilities may P (Hh2 |Hh1 ) = p̂h2 (1 − p̂hh2 )N2 −1 ×
lead to increased load estimation accuracy. Importantly, these 1
 l
accuracy gains come at no additional computational cost. n l h l
(p̂lh2 )0 (1 − p̂lh2 )n = N2h p̂hh2 (1 − p̂hh2 )N2 −1 (1 − p̂lh2 )n
Additionally, for a low overloading factor, we found that the 0
RCML estimator offered comparable performance to its ML (31)
counterpart at significant reduction of computational cost. where from (7), N2h = nh − 1 is the remaining H-UEs. p̂hh2
Our work can be further extended along several dimensions. and p̂lh2 are the probabilities that an H-UE and an L-UE can
For example, by considering the arrival rate of UEs in each occupy RB h2 given that RB h1 is unavailable, which are
priority class, or by considering more than two priority classes. given by:
Moreover, we aim to expand the system model to more ph pl
realistically represent bursty mMTC traffic by adopting the p̂hh2 = PMh2 , p̂lh2 = PMh2 (32)
h l
traffic model developed by [24]. i=1 pi i=1 pi
i6=h1 i6=h1
In general, the probability that an H-UE occupies RB hk given
A PPENDIX A
that RBs h1 , . . . , hk−1 are unavailable is:
P ROOF OF P ROPOSITION 1  h
Nk h h
The upcoming derivations of the probabilities in (3) require P (Hhk |Hh1 , . . . , Hhk−1 ) = p̂hk (1 − p̂hhk )Nk −1 ×
the use of the following lemma. 1
 l
n l
Lemma 1. Consider the mutually exclusive and exhaus- (p̂lhk )0 (1 − p̂lhk )n (33)
0
tive events A1 , . . . , An with the corresponding probabilities
π1 , . . . , πn . Given that the event Ak does not occur, the where Nkh is obtained from (7). While p̂hhk and p̂lhk are the
probability of each remaining event Ai , i 6= k, is given by: probabilities that an H-UE and an L-UE can occupy RB
πi hk given that RBs h1 , . . . , hk−1 are unavailable, which are
π̂i = Pn , ∀i ∈ {1, . . . , n} \ {k} (27) calculated using (8) and (9), respectively. Simplifying (33),
j=1 πj
j6=k results in (6), which completes the proof.
Proof. Let A = {A1 , A2 , . . . , An }. Given that event Ak does
not occur, the conditional probability of each remaining event A PPENDIX B
Ai ∈ A \ {Ak } is given by: R ECURSIVE A LGORITHMS
P (Ai ∩ Ack ) Algorithms 1 to
P (Ai |Ack ) = (28)
3 present a recursive
procedure to calculate
P (Ack )


P H , P L H and P X H, L, Φ , respectively.
where Ack is the complement of Ak with P (Ack ) = 1 − πk = Notice that in Algorithm 3 we store values for all the
PM
i=1,i6=k πi . The intersection of each remaining event Ai possible Njh , Njl , p̂hXj , and p̂lXj ; instead, only the values for
with Ack is the event Ai itself since all remaining events are the current iteration are calculated then overwritten with the
contained within Ack : values of the next iteration. Additionally, we use two helper
P (Ai ∩ Ack ) = P (Ai ). functions, F IND C OMBINATIONS and F IND C OMBINATIONS H,
which find all feasible combinations of the total number of
Hence, the probability of each remaining event Ai given event
UEs and H-UEs that can occupy the RBs in X , respectively.
Ak does not occur is:

Algorithm 1 P H
π̂i = P (Ai |Ack ) = P (Ai ∩ Ack )/P (Ack )
n 1: Initialization: z = 1, N h = nh , p̂h = ph , and p̂l = pl
2: for each i in H do
X
= P (Ai )/P (Ack ) = πi / πi , ∀Ai ∈ A \ {Ak } (29) 3:
h
z ← z × N h p̂hi (1 − phi )N −1 (1 − p̂li )n
l

i=1 h h
i6=k 4: N ←N −1
Pn 5: p̂hi ← 0, p̂li ← 0
such that i=1 π̂i = 1. p̂h p̂lj
6: p̂hj ← PM
j
p̂h
, p̂lj ← PM
p̂l
for j 6= i
k=1 k k=1 k
We can now begin the proof of Proposition 1, as shown k6=i k6=i


below. The first factor in (4) is the probability that only one 7: At the end of the ith iteration z holds P H .
H-UE occupies RB h1 , and no other UEs are in RB h1 :
 h  l
n h h nh −1 n l
P (Hh1 ) = ph1 (1 − ph1 ) (plh1 )0 (1 − plh1 )n , R EFERENCES
1 0
(30) [1] I. Leyva-Mayorga, L. Tello-Oquendo, V. Pla, J. Martinez-Bauset, and
which when simplified results in (5). The second factor in (4) V. Casares-Giner, “Performance analysis of access class barring for
handling massive M2M traffic in LTE-A networks,” in 2016 IEEE
is the probability that only one H-UE occupies RB h2 , and no international conference on communications (icc). IEEE, 2016, pp.
other UEs are in RB h2 , given that one H-UE is in RB h1 . 1–6.


Algorithm 2 P L H (PIMRC). IEEE, 2018, pp. 1972–1976.
h h l l [8] S. Duan, V. Shah-Mansouri, Z. Wang, and V. W. Wong, “D-ACB:
1: Initialization:
 z = 1, N = n − H, N = n Adaptive congestion control algorithm for bursty M2M traffic in LTE
0 if j ∈ H networks,” IEEE Transactions on Vehicular Technology, vol. 65, no. 12,
2: p̂h
j = ph pp. 9847–9861, 2016.
 P j ph otherwise [9] Evolved Universal Terrestrial Radio Access, “Medium access control
 i∈H
/ i
(MAC) protocol specification (3GPP TS 36.321 version 8.5. 0 release
0 if j ∈ H
8),” ETSI TS, vol. V16, no. 321, 2021.
3: p̂lj = plj
P l otherwise [10] H. D. Althumali, M. Othman, N. K. Noordin, and Z. M. Hanapi,
/ pi
i∈H “Dynamic backoff collision resolution for massive M2M random access
4: for each i in L do l h in cellular IoT networks,” IEEE Access, vol. 8, pp. 201 345–201 359,
5: z ← z × N l p̂li (1 − p̂li )N −1 (1 − p̂hi )N 2020.
l l
6: N ←N −1 [11] M. R. Chowdhury and S. De, “Queue-aware access prioritization for
7: p̂hi ← 0, p̂li ← 0 massive machine-type communication,” IEEE Internet of Things Jour-
p̂h p̂lj nal, vol. 9, no. 17, pp. 15 858–15 873, 2022.
8: p̂hj ← PM
j
p̂h
, p̂lj ← PM
p̂l
for j 6= i [12] H. Althumali, M. Othman, N. K. Noordin, and Z. M. Hanapi, “Priority-
k=1 k k=1 k
k6=i k6=i based load-adaptive preamble separation random access for QoS-
differentiated services in 5G networks,” Journal of Network and Com-


9: At the end of the ith iteration z holds P L H .
puter Applications, vol. 203, p. 103396, 2022.
[13] L. Tello-Oquendo, I. Leyva-Mayorga, V. Pla, J. Martinez-Bauset, J.-


Algorithm 3 P X H, L, Φ
R. Vidal, V. Casares-Giner, and L. Guijarro, “Performance analysis and
1: Initialization: z = 0, N h = nh −H, N l = nl −L, N = N h +N l optimal access class barring parameter configuration in LTE-A networks
 with massive M2M traffic,” IEEE Transactions on Vehicular Technology,
0 if j ∈ H, L, Φ vol. 67, no. 4, pp. 3505–3520, 2017.
2: p̂h
j = ph
j
[14] M. Alvi, K. M. Abualnaja, W. T. Toor, and M. Saadi, “Performance
P otherwise analysis of access class barring for next generation IoT devices,”
 i∈X phi Alexandria Engineering Journal, vol. 60, no. 1, pp. 615–627, 2021.
0 if j ∈ H, L, Φ [15] S. Duan, V. Shah-Mansouri, and V. W. Wong, “Dynamic access class
3: p̂lj = plj barring for M2M communications in LTE networks,” in 2013 IEEE
P l otherwise Global Communications Conference (GLOBECOM). IEEE, 2013, pp.
i∈X pi
4: K ← F IND C OMBINATIONS(X, N, empty list) 4747–4752.
5: for each k in K do [16] H. Zhang, M. Jiang, X. Liu, X. Wen, N. Wang, and K. Long, “PPO-
6: I ← F IND C OMBINATIONS H(X, N h , k, empty list) based PDACB traffic control scheme for massive IoV communications,”
IEEE Transactions on Intelligent Transportation Systems, 2022.
7: for each i in N do [17] H. Wu, C. Zhu, R. J. La, X. Liu, and Y. Zhang, “Fast adaptive S-ALOHA
8: Initialization: z ′ = 1, p = 1, N h , N l , p̂hj , p̂lj scheme for event-driven machine-to-machine communications,” in 2012
9: for each j in X do IEEE vehicular technology conference (VTC Fall). IEEE, 2012, pp.
h
h
10: z′ ← z ′ × Ni (p̂hj )ip (1 − p̂hj )N −ip ×
p
1–5.
Nl

l kp −ip l [18] O. Galinina, A. Turlikov, S. Andreev, and Y. Koucheryavy, “Stabilizing
kp −ip
(p̂j ) (1 − p̂lj )N −kp +ip multi-channel slotted aloha for machine-type communications,” in 2013
11: N h ← N h − ip , N l ← N l − (kp − ip ) IEEE International Symposium on Information Theory. IEEE, 2013,
12: h
p̂j ← 0, l
p̂j ← 0 pp. 2119–2123.
p̂h p̂l [19] Z. Wang and V. W. Wong, “Optimal access class barring for stationary
13: p̂hm ← PM m h ,
p̂
p̂lm ← PM m l
p̂
for m 6= j machine type communication devices with timing advance information,”
k=1 k k=1 k
k6=j k6=j IEEE Transactions on Wireless communications, vol. 14, no. 10, pp.
14: p←p+1 5374–5387, 2015.
15: z ← z + z′ [20] M. Polese, M. Centenaro, A. Zanella, and M. Zorzi, “M2M massive

access in LTE: RACH performance evaluation in a smart city scenario,”
16: At the end of the kth iteration z holds P X H, L, Φ .
in 2016 IEEE International Conference on Communications (ICC).
ieee, 2016, pp. 1–6.
[2] M. Condoluci, M. Dohler, G. Araniti, A. Molinaro, and K. Zheng, [21] D. H. Morais, “5G NR overview and physical layer,” in Key 5G/5G-
“Toward 5G densenets: architectural advances for effective machine-type Advanced Physical Layer Technologies: Enabling Mobile and Fixed
communications over femtocells,” IEEE Communications Magazine, Wireless Access. Springer, 2024, pp. 233–321.
vol. 53, no. 1, pp. 134–141, 2015. [22] J.-P. Cheng, C.-h. Lee, and T.-M. Lin, “Prioritized random access with
[3] D. T. Wiriaatmadja and K. W. Choi, “Hybrid random access and dynamic access barring for RAN overload in 3GPP LTE-A networks,”
data transmission protocol for machine-to-machine communications in in 2011 IEEE GLOBECOM Workshops (GC Wkshps). IEEE, 2011, pp.
cellular networks,” IEEE Transactions on Wireless Communications, 368–372.
vol. 14, no. 1, pp. 33–46, 2014. [23] T.-M. Lin, C.-H. Lee, J.-P. Cheng, and W.-T. Chen, “PRADA: Prioritized
[4] C.-H. Wei, G. Bianchi, and R.-G. Cheng, “Modeling and analysis random access with dynamic access barring for MTC in 3GPP LTE-A
of random access channels with bursty arrivals in OFDMA wireless networks,” IEEE Transactions on Vehicular Technology, vol. 63, no. 5,
networks,” IEEE Transactions on Wireless communications, vol. 14, pp. 2467–2472, 2014.
no. 4, pp. 1940–1953, 2014. [24] 3GPP, “Study on RAN improvements for machine type communica-
[5] O. Arouk, A. Ksentini, and T. Taleb, “How accurate is the RACH tions,” TR 37.868, Tech. Rep., 2011.
procedure model in LTE and LTE-A?” in 2016 International Wireless
Communications and Mobile Computing Conference (IWCMC). IEEE,
2016, pp. 61–66.
[6] L. Tello-Oquendo, V. Pla, I. Leyva-Mayorga, J. Martinez-Bauset,
V. Casares-Giner, and L. Guijarro, “Efficient random access channel
evaluation and load estimation in LTE-A with massive MTC,” IEEE
Transactions on Vehicular Technology, vol. 68, no. 2, pp. 1998–2002,
2018.
[7] J. Sun, R. Liu, and E. Paolini, “Detecting the number of active users
in IRSA access protocols,” in 2018 IEEE 29th Annual International
Symposium on Personal, Indoor and Mobile Radio Communications