Kettering University

Digital Commons @ Kettering University

Industrial & Manufacturing Engineering Publications Industrial & Manufacturing Engineering

5-30-2020

A Retrial Queueing Model With Thresholds and Phase Type Retrial

Times
Srinivas R. Chakravarthy
Kettering University, [email protected]

Follow this and additional works at: https://digitalcommons.kettering.edu/

industrialmanuf_eng_facultypubs

Part of the Operations Research, Systems Engineering and Industrial Engineering Commons

Recommended Citation
Chakravarthy, Srinivas R., "A Retrial Queueing Model With Thresholds and Phase Type Retrial Times"
(2020). Industrial & Manufacturing Engineering Publications. 113.
https://digitalcommons.kettering.edu/industrialmanuf_eng_facultypubs/113

This Article is brought to you for free and open access by the Industrial & Manufacturing Engineering at Digital
Commons @ Kettering University. It has been accepted for inclusion in Industrial & Manufacturing Engineering
Publications by an authorized administrator of Digital Commons @ Kettering University. For more information,
please contact [email protected].
J. Appl. Math. & Informatics Vol. 38(2020), No. 3 - 4, pp. 351 - 373
https://doi.org/10.14317/jami.2020.351

A RETRIAL QUEUEING MODEL WITH THRESHOLDS AND

PHASE TYPE RETRIAL TIMES

SRINIVAS R. CHAKRAVARTHY

Abstract. There is an extensive literature on retrial queueing models.

While a majority of the literature on retrial queueing models focuses on
the retrial times to be exponentially distributed (so as to keep the state
space to be of a reasonable size), a few papers deal with nonexponential
retrial times but with some additional restrictions such as constant retrial
rate, only the customer at the head of the retrial queue will attempt to
capture a free server, 2-state phase type distribution, and finite retrial or-
bit. Generally, the retrial queueing models are analyzed as level-dependent
queues and hence one has to use some type of a truncation method in per-
forming the analysis of the model. In this paper we study a retrial queueing
model with threshold-type policy for orbiting customers in the context of
nonexponential retrial times. Using matrix-analytic methods we analyze
the model and compare with the classical retrial queueing model through
a few illustrative numerical examples. We also compare numerically our
threshold retrial queueing model with a previously published retrial queue-
ing model that uses a truncation method.

AMS Mathematics Subject Classification : 60K20, 60J28.

Key words and phrases : Retrial, queueing, phase type distribution,
thresholds.

1. Introduction
Retrial queueing models are queueing models in which an arriving customer
not able to get into service enters into a retrial orbit and try to capture a
free server by competing with other customers, if any, present in the orbit.
These models have been identified to be very useful in many applications no-
tably in telecommunications and service industries. Retrial queues have been
studied extensively in the literature (see e.g., [1, 3, 6, 8, 21, 26]). Much of
the literature deals with exponential retrials with the exception of a few stud-
ies(see e.g., [5, 15, 18, 23, 36, 37, 38, 40]). As pointed out in [15], the few

Received August 1, 2019. Revised December 21, 2019. Accepted February 11, 2020.
c 2020 KSCAM.

351
352 Srinivas R. Chakravarthy

papers dealing with nonexponential retrials, such as phase type distributions

(P H−distribution) for retrials, propose a variety of approximations in their
studies. Such an approach to the study of retrial queueing models with non-
exponential retrials is not by choice but rather due to the inherent complexity
created by the exponential growth of the state space. Realizing the lack of re-
sults in multi-server retrial queueing models with nonexponential retrial times,
Chakravarthy [15] proposed a different approach via simulation and reported
some interesting results. Specifically, it was shown in [15] that assuming expo-
nential retrial times in place of nonexponential ones could lead to under or over
estimating the system performance measures.
In classical queues, a threshold approach is employed to manage any possi-
ble congestions in the system. For example, in a classical multi-server queueing
system, one way to minimize the congestion (in terms of average number of
customers present) in the system is to have threshold-type policies to deter-
mine the allocation of the jobs/customers among the heterogeneous servers (see,
e.g., [19, 28, 29, 32, 35]). Further, in classical multi-server queues (with or with-
out homogeneity assumption on the servers), threshold-type policies have been
identified to minimize the number of servers (see, e.g., [11, 13, 14, 24]) in the
system.
Efrosinin and Breuer [20] established that a threshold-type policy is also op-
timal for retrial queues. It should be pointed out the threshold they employ is
only at the arrival points of new or retrial customers accessing a free server from
a finite retrial buffer. That is, the threshold becomes active only when a new
customer arrives (at which time a decision is made either to route to an idle
server or to send the customer to the retrial orbit provided there is a space) or
a retrial customer arrives (at which time a decision is made either to route to
an idle server or send the customer back to the retrial orbit). However, to the
best of our knowledge there is no literature that employs a threshold-type policy
for the waiting customers in the retrial orbit. Thus, in this paper, we take a
different approach to the study of the retrial queues by introducing the concepts
of threshold based retrial times.
The main motivation for the study of a threshold-type retrial queueing model
in this paper arose out of a need for including nonexponential retrial times and at
the same time not to significantly increase the complexity of the retrial model.
For example, Shin [37] studied an M/M/c-type retrial queueing model with
phase type (restricted to only two phases due to the size of the underlying state
space) retrial times and apply the level-dependent QBD−process approach to
study the model. Relaxing the assumption of a 2−state P H−distribution for
the retrials, Shin and Moon [38] proposed an approximation for the distribution
of the number of busy servers as well as the mean number of customers in retrial
orbit. These two papers, for example, further illustrate the complexity involved
in relaxing the exponential assumption for the retrial times. Our approach using
threshold-type retrial times has an advantage in that one can use this model as
 A retrial queueing model with thresholds and phase type retrial times 353

another approximation (through appropriate choice of the threshold parameters)

to retrial queues with phase type retrial times.
For use in the sequel, we set up the following notation. By (a) e, a column
vector (of appropriate dimension) of 1’s; (b) ei , a unit column vector (of appro-
priate dimension) with 1 in the ith position and 0 elsewhere; (c) I is an identity
matrix (of appropriate dimension); (d) ∆(T1 , · · · , Tr ), a diagonal matrix with
diagonal entries given by Ti , i = 1, · · · , r. Note that these entries can be scalars,
vectors, or matrices, and will be clear from the context. The dimension of the
vectors and matrices should be clear in the context of usage. We will use e(mn)
to show that the column vector is of dimension mn when more clarity is needed.
We also need to use Kronecker product (denoted by ⊗) and Kronecker sum
(denoted by ⊕) of matrices. For details on these, we refer to [22, 30, 39].
The paper is organized as follows. In Section 2, we describe the threshold
retrial queueing model under study in this paper. The QBD-process needed to
study this model and its steady-state analysis are presented in Section 3. The
comparison of our threshold model to the one discussed in [38] is carried out in
Section 4. Illustrative numerical examples are presented in Section 5 and some
concluding remarks including future research work are presented in Section 6.

2. Model description
We consider a multi-server retrial queueing model in which the customers
arrive according to a Poisson process with rate λ. An arriving customer, find-
ing all c servers busy, will enter into a retrial buffer of infinite capacity. The
service times are assumed to be exponentially distributed with parameter µ.
The customers who are in retrial orbit will attempt to capture a free server at
random times. In classical retrial queueing model, it is generally assumed that
the retrial rates are proportional to the number of customers waiting in the re-
trial orbit and that the random times to attempt to capture a free server are
exponentially distributed. There are variations such as constant retrial policy
(see, e.g., [2, 9, 10, 16]), finite retrial orbit (see, e.g., [4, 7]), and attempts to
capture a free server are made only by the customer at the head of the orbit (see,
e.g., [36]). Such variations lead to level-independent QBD−process to analyze
the model. But these restrictions may not be valid or suitable in practice.
We introduce two threshold parameters, say, N, 1 ≤ N < ∞, and K, 1 ≤ K <
∞, such that when the number of retrial customers is between (k − 1)N + 1
to kN , the retrial rate will be θk , 1 ≤ k ≤ K, and once the number in retrial
exceeds KN , the retrial rate is taken to be θ.
In this paper we will assume that the underlying random variable governing
the retrial times is of phase type with rates dictated by the parameters, N
and K. Suppose that the representation (β, S) of order n with β(−S)−1 e =
1. Then the retrial times are modeled using this P H−distribution and the
threshold parameters. That is, when the number of retrial customers is in the
interval [(k − 1)N + 1, kN ], 1 ≤ k ≤ K, the retrial times are of phase type with
354 Srinivas R. Chakravarthy

representation (β, θk S) of order n. Once the number in the retrial buffer exceeds
KN , the retrial times will follow a P H−distribution with representation (β, θS)
of order n. Note that by taking N = 1, we get the classical retrial queueing model
in which the retrial times are of phase type with all customers’ P H−distributions
having a common set of states but possibly with different rates. The value of K
may be chosen similar to how a level dependent QBD−process is truncated for
numerical purposes in the literature (see, e.g., [12]). As can be seen, in this paper
we assume that we have the same number of states for the P H−distribution for
all customers. However, we will elaborate more on how this can be relaxed and
the impact of this relaxation later on.
Suppose that we define J1 (t) to be the number of customers in the retrial orbit;
J2 (t) to be the number of busy servers; and J3 (t) to be the phase of the retrial
times (when J1 (t) > 0) at time t. It is easy to verify that {(J1 (t), J2 (t), J3 (t)) :
t ≥ 0} is a quasi-birth-and-death process (QBD) with state space given by

[
Ω = {(0, j) : 0 ≤ j ≤ c} {(i, j, k) : i ≥ 1, 0 ≤ j ≤ c, 1 ≤ k ≤ n}.

Let 0 = {(0, j), 0 ≤ j ≤ c} denote the set of states corresponding to the system
in which the retrial buffer is empty and that j servers are busy. Note that in this
case there is no need to keep track of the phase of underlying P H−distribution;
let i = {(i, j, k), 0 ≤ j ≤ c, 1 ≤ k ≤ n}, denote the set of states wherein
exactly i customers are in the retrial orbit with j servers being busy, and the
underlying P H−distribution is in phase k. For the sake of not complicating the
notation, we group the set of levels 1 through KN into K groups of N levels
when displaying the generator Q below. Thus, the matrices Qi i, 1 ≤ i ≤ K are
of dimension N (c + 1)n, whereas the matrices A0 , A1 , and A2 are of dimension
(c + 1)m.
The generator, Q, of the QBD process under consideration is of the form

B̃0,1 B̂0,0
 
 B̂1,2 Q1,1 Ã0 
 

 Q2,1 Q2,2 Ã0 


 Q3,2 Q3,3 Ã0 

 .. .. .. 
 . . . 
Q=  , (1)
 

 QK,K−1 QK,K Â0 


 Â2 A1 A0 


 A2 A1 A0 


 A2 A1 A0 

.. .. ..
. . .
 A retrial queueing model with thresholds and phase type retrial times 355

where the (block) matrices appearing in Q are as follows.

 
−λ λ
 µ −(λ + µ) λ 
 
 2µ −(λ + 2µ) λ 
B̃0,1 =  , (2)
 
. . . . . ..

 . . 

 (c − 1)µ −(λ + (c − 1)µ) λ 
cµ −(λ + cµ)

B̂0,0 = e01 (N ) ⊗ B̃0,0 , B̂1,2 = e1 (N ) ⊗ B̃1,2 , (3)

 
Br,1 A0
 Br,2 Br,1 A0 
 
 Br,2 Br,1 A0 
Qr,r =  , 1 ≤ r ≤ K, (4)
 
.. ..

 . . 

 Br,2 Br,1 A0 
Br,2 Br,1

Qr,r−1 = e1 (N ) ⊗ e0N (N ) ⊗ Br,2 , 2 ≤ r ≤ K,

Ã0 = eN (N ) ⊗ e01 (N ) ⊗ A0 , (5)

Â2 = e0N (N ) ⊗ A2 , Â0 = eN (N ) ⊗ A0 ,

B̃0,0 = λec+1 (c + 1) ⊗ e0c+1 (c + 1) ⊗ β, (6)

0 S0
 

 S0


B̃1,2 = θ1 
 .. ..
..  ,
 . . 
 . (7)
 S0 
0
A0 = λec+1 (c + 1) ⊗ e0c+1 (c + 1) ⊗ I,

S0β
 
0

 S0β 

Br,2 = θr 
 .. .. 
,..
. . . 
(8)
 
0
 S β 
0
Br,1 = ∆(θr S, · · · , θr S, θr (S + S 0 β)) + B̃0,1 ⊗ I,
2 ≤ r ≤ K,
356 Srinivas R. Chakravarthy

S0β
 
0

 S0β 

A2 = θ 
 .. ....  ,
 . ..  (9)
 S0β 
0
A1 = ∆(θS, · · · , θS, θ(S + S 0 β)) + B̃0,1 ⊗ I.

3. Steady-state analysis
In this section the model described in Section 2 will be studied in steady-
state. First, we recall that the stability condition of the classical retrial queueing
model is same as that of the classical queueing model [8]. For example, a retrial
queueing model of the type M/M/c in which all the customers waiting in the
retrial orbit (of infinite size) independently attempt to capture a free server is
stable if and only if λ < cµ. However, the moment one puts a restriction in the
way the retrials are modeled, the stability condition needs to be modified. This
is also the case in our model studied in this paper.

3.1. Stability condition. Suppose that A = A0 + A1 + A2 (note that A is

irreducible) and that π is its steady-state probability vector. That is,
πA = 0, πe = 1. (10)
On noting that the model described in Section 2 is a QBD−process, one can
apply the classical result due to Neuts (see, e.g., [31]) to arrive at the stability
condition of the retrial queueing model under study. This is given in the follow-
ing theorem.

Theorem 1: The threshold retrial queueing model under study with the gen-
erator given in (1) is stable if and only if
c
X
λ<µ jπ j e. (11)
j=1

Proof: First note that the equations given in (10) can be rewritten as
π 0 (θS − λI) + µπ 1 = 0,

θπ j−1 S 0 β + λπ j−1 + π j (θS − (λ + jµI) + (j + 1)µπ j+1 = 0, 1 ≤ j ≤ c − 1,

θπ c−1 S 0 β + λπ c−1 + θπ c (S + S 0 β) − cµπ c = 0,

Pc
j=0 πj e = 1,
(12)
from which it is easy to verify that
θπ j S 0 + λπ j e = (j + 1)µπ j+1 e, 0 ≤ j ≤ c − 1. (13)
 A retrial queueing model with thresholds and phase type retrial times 357

Pc−1
The proof follows immediately on noting (a) πA2 e = θ j=0 π j S 0 ; (b) πA0 e =
Pc−1 Pc
λπ c e; (c) θ j=0 π j S 0 + λ(1 − π c e) = µ j=0 jπ j e and (d) applying the sta-
bility condition for the QBD− process (see, e.g., [31]), which is πA0 e < πA2 e.

Note: (1) It is worth pointing out that the vector π is independent of θk , 1 ≤

k ≤ K. This is as is to be expected as the matrices A0 , A1 and A2 govern tran-
sitions away from the boundary states.
(2) As θ → ∞, π j → 0, 0 ≤ j ≤ c − 1, π c → β(−S)−1 , and hence the condition
λ
cµ < 1 becomes a necessary and sufficient condition for the stability. This is
also intuitive as the retrial queueing model approaches the corresponding clas-
sical queueing model when θ → ∞. It is worth mentioning that in the classical
retrial queue, as the number of customers in the orbit grows without bound (by
fixing the retrial rate), the total (linear) retrial rate will approach ∞; thus, the
retrial queue approaches the corresponding classical queue.

3.2. Steady-state probability vector. Suppose that x = (x(0), x(1), · · · )

denote the steady-state probability vector of the generator Q given in 1. That
is, x satisfies

xQ = 0, xe = 1. (14)

We partition the vectors, x(i), for i ≥ 0, as x(0) = (x0 (0), · · · , xc (0)) and
x(i) = (x0 (i), · · · , xc (i)). First, note that x(0) is of dimension c + 1, while
x(i), i ≥ 1 is of dimension (c + 1)n. Secondly, the j th component, for 0 ≤
j ≤ c, gives the steady-state probability vector that the retrial orbit is empty
with exactly j servers busy serving customers; the k th component of the vector
xj (i), 0 ≤ j ≤ c gives the steady-state probability vector that the retrial orbit
has i customers with exactly j servers busy serving customers, and that the
underlying P H−distribution is in phase k, 1 ≤ k ≤ n. It is worth pointing out
that there can be at least one free server with at least one customer in the retrial
orbit.
Under the stability condition given in (11) the steady state probability vector
x is obtained (see, e.g., [31]) as shown in the following theorem.

Theorem 2: Assume that the stability condition as given in (11) holds good.
Then the steady-state vector, x, of the threshold retrial queueing model under
study with the generator given in (1) is obtained by solving the following system
358 Srinivas R. Chakravarthy

of linear equations:

x(0)B̃0,1 + x(1)B̃1,2 = 0,

x(0)B̃0,0 + x(1)B1,1 + x(2)B1,2 = 0,

x(i − 1)A0 + x(i)B1,1 + x(i + 1)B1,2 = 0, 2 ≤ i ≤ N − 1,

x(N − 1)A0 + x(N )B1,1 + x(N + 1)B2,2 = 0,

(15)
x(i − 1)A0 + x(i)Bj+1,1 + x(i + 1)Bj+1,2 = 0,
jN + 1 ≤ i ≤ (j + 1)N − 1, 1 ≤ j ≤ K − 1,

x(i − 1)A0 + x(i)Bj,1 + x(i + 1)Bj+1,2 = 0, i = jN, 2 ≤ j ≤ K − 1,

x(KN − 1)A0 + x(KN )[BK,1 + RA2 ] = 0,

PKN −1
i=0 x(i)e + x(KN )(I − R)−1 e = 1,
where the matrix R is the minimal nonnegative solution to the matrix-quadratic
equation:
R2 A2 + RA1 + A0 = 0. (16)
Proof: First note that the QBD−structure of the generator given in (1), un-
der the stability condition, yields (see, e.g., [31]) a modified matrix-geometric
solution. Thus, the non-boundary states, namely, for i ≥ KN , are given by

x(i + KN ) = x(KN )Ri , i ≥ 0, (17)

where R satisfies the matrix-quadratic equation given in (16). With regard to

the non-boundary states, the stated equations follow once we rewrite the equa-
tions given in (14).

The following lemma, which is intuitively obvious, will serve as another accuracy
check in numerical implementation of the steady-state probability vector.

Lemma 1: We have
K
hX jN
X i
θj x(i) + θx(KN )R(I − R)−1 (e ⊗ I) = dβ(−S)−1 , (18)
j=1 i=(j−1)N +1

where d is the normalizing constant and is given by

K
hX jN
X i
d= θj x(i)e + θx(KN )R(I − R)−1 e . (19)
j=1 i=(j−1)N +1
 A retrial queueing model with thresholds and phase type retrial times 359

Proof: Post-multiplying the first equation in (15) by e we get

−λx(0)ec+1 (c + 1) + θ1 x(1)((e − ec+1 (c + 1)) ⊗ S 0 ) = 0. (20)

Post-multiplying all but the first and last equations in (15) by e ⊗ I (note that
we replace the last but one equation by x(KN − 1)A2 + x(KN )BK,1 + x(KN +
1)A2 = 0 and x(i − 1)A0 + x(i)A1 + x(i + 1)A2 = 0, i ≥ KN + 1), and adding
the resulting equations we get
h
−λ(x(0)ec+1 (c + 1) ⊗ β) + θ1 x(1)[(e ⊗ S) + (ec+1 (c + 1) ⊗ S 0 β)]
i (21)
PK−1 P(j+1)N P∞
j=1 θj+1 i=jN +1 x(i) + θ i=KN +1 x(i) (e ⊗ I) = 0.

Now post-multiplying equation (20) by β and adding the resulting equation with
(21), we get

K
hX jN
X ∞
X ih i
θj x(i) + θ x(i) e ⊗ (S + S 0 β) = 0, (22)
j=1 i=(j−1)N +1 i=KN +1

from which using the uniqueness of the stationary vector of the generator S+S 0 β
the stated result follows.

The following lemma, which again can be used as another accuracy check in
the numerical computations, is similar to the well-known result in the classical
queue. That is, the average number of busy servers is given by µλ . Towards this
end, we define
P∞
ỹ = i=0 y i , where y i = (yi,0 , · · · , yi,c ), i ≥ 0,

x̃ = (x̃0 , · · · , x̃c ), (23)

PK PN P∞
where x̃j = k=1 θk r=1 x(k−1)N +r,j + θ i=KN +1 xk,j , 0 ≤ j ≤ c.

Note that y 0 = x(0) and y i = (x0 (i)e, · · · , xc (i)e).

λ
Lemma 2: The fraction, µ, gives the mean number of busy servers in the
system. That is,
c
X λ
j ỹj = . (24)
j=1
µ
360 Srinivas R. Chakravarthy

Proof: First, from the steady-state equations given in (15), verify that
 
−λ λ
 µ −(λ + µ) λ 
 
 2µ −(λ + 2µ) λ 
ỹ 
 
.. .. .. 

 . . . 

 (c − 1)µ −(λ + (c − 1)µ) λ 
cµ −cµ

−S 0 S 0
 
 −S 0 S 0 
0 0
 
 −S S 
+ x̃  = 0. (25)
 
.. .. 

 . . 

 −S 0 S 0 
0
From equation (25) it is easy to verify
x̃j S 0 + λỹj = (j + 1)µỹj+1 , 0 ≤ j ≤ c − 1, (26)
from which we get
c−1
X c−1
X c
X
x̃j S 0 + λ ỹj = jµỹj . (27)
j=0 j=0 j=1
Pc
The stated result follows from the above equation on noting that j=0 ỹj = 1
Pc−1
and λỹc = j=0 x̃j S 0 . Note that the latter equation is due to the fact that in
steady-state the rate at which a customer enters a retrial orbit should be equal
to the rate at which a retrial customer captures a free server.

3.3. Computational procedures. In this section we will briefly outline the

computational procedures involved in obtaining the rate matrix R (of dimension
(c + 1)n) and the steady-state probability vector, x̃ which are key ingredients to
system performance measures to qualitatively study the model. First, we look at
the R matrix. One can use a number of well-known methods such as logarithmic
reduction [27] to compute R, especially when the dimension of R is of reasonable
size. If the dimension is prohibitively large, say, when c or n or both large, then
one should use (block) Gauss-Siedel iteration by exploiting the special structure
of the coefficient matrices A0 , A1 , and A2 .
Due to structure of the coefficient matrices in (16), it is easy to verify that R
is of the form  
0 0 ··· 0
 .. .. .. 
R= .
 . ··· . 
. (28)
 0 0 ··· 0 
R0 R1 · · · Rc
 A retrial queueing model with thresholds and phase type retrial times 361

The above form of R along with the structure of the coefficient matrices can be
exploited to rewrite the matrix-quadratic equation given in (16) as follows.
R0 = µR1 (λI − θS)−1 ,

Rk = [λRk−1 + (k + 1)µRk+1 + θRc Rk−1 S 0 β][(λ + kµ)I − θS]−1 , 1 ≤ k ≤ c − 1,

Rc = [λRc−1 + θRc Rc−1 S 0 β + λI][(λ + kµ)I − θ(S + S 0 β)]−1 .

(29)
Similar to exploiting the structure in computing R, one can do the same for
computing the steady-state probability vector, x. We will briefly display two
sets of such equations.
µ
x0 (0) = λ x1 (0),
h i
1
xk (0) = λ+kµ λxk−1 (0) + (k + 1)µxk+1 (0) + θ1 xk−1 (1)S 0 , 1 ≤ k ≤ c − 1,
h i
1
xc (0) = λ+kµ λxc−1 (0) + θ1 xc−1 (1)S 0 ,

x0 (1) = µx1 (1)(λI − θ1 S)−1 ,

xk (1) = [λxk−1 (1) + (k + 1)µxk+1 (1) + θ1 xk−1 (2)S 0 β][(λ + kµ)I − θ1 S]−1 ,
1 ≤ k ≤ c − 1,

xc (1) = [λxc (0)β + λxc−1 (1) + θ1 xc−1 (2)S 0 β][(λ + cµ)I − θ1 (S + S 0 β)]−1 .

3.4. System performance measures. In order to analyze the model qual-

itatively, we need to look into some key system performance measures. Here,
we will list a few performance measures along with the expressions for their
computations.
(1) Probability that the system is idle. The probability, Pidle , that the system
is idle (i.e. all servers are idle and the retrial buffer is empty) at an
arbitrary time is given by
Pidle = x0,0 .
(2) Probability that the orbit is empty. The probability, PO−empty , that the
retrial buffer is empty at an arbitrary time is given by
PO−empty = x(0)e.
(3) Probability that the system is in various mode. The probability, P SMr ,
that the system is operating in mode r, 0 ≤ r ≤ K + 1, is of interest.
Note that the system is said to be in (a) mode 0 when there is no one
in the retrial orbit; (b) in mode i, 1 ≤ i ≤ K when the rate of retrial is
362 Srinivas R. Chakravarthy

given by θi ; and (c) in mode K + 1 when the rate of retrials is given by

θ. 

 x(0)e, r = 0,


 P
N
P SMr = i=1 x((r − 1)N + i)e, 1 ≤ r ≤ K,




x(KN )R(I − R)−1 e, r = K + 1.


(4) Probability mass function of the number of busy servers. The probability
mass function of the number of busy servers is given by {ỹj , 0 ≤ j ≤ c}
(see equation (23)). From this probability mass function, we can get the
mean, µ0 , number of busy servers and the standard deviation, σ0 of the
number of busy servers.
(5) Mean number of customers in retrial orbit. The mean, µRO , is calculated
as
PKN −1 Pc−1
µRO = i=1 ix(i)e + KN k=0 xk (KN )e
c−1
− Rc )−1 k=0 Rk e + KN xc (KN )(I − Rc )−1 e
P
+KN xc (KN )(I 
c−1
+xc (KN )(I − Rc )−2 e + k=0 Rk e + (KN − 1)xc (KN )(I − Rc )−1 e.
P

(6) Probability of blocking The probability, Pblock , that an arriving customer

finds all servers busy is obtained as
∞
X
Pblock = x0 (c) + xc (i)e.
i=1

(7) Rate of successful capture of a free server by a customer from retrial

orbit. The rate, ξr , 1 ≤ r ≤ K + 1, of successful capture of a free server
from the retrial orbit is calculated as
 PN Pc−1
 θr j=1 k=0 xk ((r − 1)N + j)S 0 , 1 ≤ r ≤ K,
ξr =
θxc (KN )R(I − R)−1 [(e(c) − ec (c)) ⊗ S 0 ], r = K + 1.


4. Comparison with Shin-Moon model [38]

In this section, we compare our threshold model with the retrial model studied
in ([38]) through a few numerical examples. In [38], the authors report a num-
ber of numerical examples dealing with Erlang, hyperexponential, mixtures of
Erlang, and compositions of mixtures of two Erlangs. In getting these numerical
results, the authors in [38] recourse to truncation method (to find a cut-off point
for the level-dependent QBD−process that is used to study the model) proposed
in [37] and an approximation method for reporting a few system performance
measures.
For the sake of completeness and for continuity of discussion, we will use the
same notation for the distributions considered for the numerical examples as
in [38]. In the following m1 denote the mean of retrial times of each customer.
 A retrial queueing model with thresholds and phase type retrial times 363

That is m1 denotes the mean of the P H−distribution governing the retrials.

1. Erlang (Ek (ν)): This is Erlang of order k with parameter ν.
2. Hyperexponential (H2 ): This is hyperexponential with two states with
probability density function, f (t), given by f (t) = pν1 e−ν1 t + (1 − p)ν2 e−ν2 t , t ≥
2p
, ν2 = 2(1−p)
p
0, where p = 0.5(1 + ( 23 )), ν1 = m 1 m1 .
3. Mixture of two Erlangs (M ERk (p; ν1 , ν2 )): This is a mixture of two
Erlangs of order k and is obtained as M ERk (p; ν1 , ν2 ) = pEk (ν1 )+(1−p)Ek (ν2 ).
4. Composition of the mixture of Erlangs (CEk,j (p; ν1 , ν2 )): This is a
composition of the mixture of two Erlangs of orders k and j, whose Laplace
 k  j  j
transform is given by f ∗ (s) = p ν1ν+s
1 ν2
ν2 +s + (1 − p) ν2
ν2 +s , Re(s) ≥ 0.

The parameter values chosen for comparison purposes in this section are (see [38])
as follows. λ = cµ, c = 5, and µ = 1. The traffic intensity, ρ, and the mean re-
trial times, m1 are varied and the values are displayed in the comparison tables
below. In the following, theparameters for the  mixture of two Erlangs, de-
4 10
noted by M ER3 , are M ER3 0.0740741; 3m ,
1 3m1
; the two distributions cor-
(1) (2)
responding to the composition of mixtures of Erlangs, CE3,1 and CE3,1 have
 
(1)
parameters, respectively, given by CE3,1 = CE3,1 0.007773; 0.146991
m1 , 1.188568
m1
 
(2) 0.61971 9.79811
and CE3,1 = CE3,1 0.185487; m1 , m1 .
While we used the same set of values for the common parameters such as
λ, µ, c, and ρ, some other additional parameters that are in our model are taken
as follows. We fixed K = 100, N = 1, θ1 = m11 , and θk = kθ1 . The value of θ is
obtained as the minimum value for which µ Pc λ j π j e is close to a given value
j=1
of ρ. That is, θ is such that

 P λ
 
− ρ < 10−3 . (30)

c
µ j=1 jπ j e


A few comments about the choice of K and N . One can think of K as the trun-
cation point similar to the one chosen in [37] to truncate the level-dependent
QBD−process. The idea of fixing N = 1 is to sort of resemble the classical
retrial queue with phase type retrial times. Unlike the models in [37], wherein
the number of phases is fixed at 2 in order to minimize the dimension of the
problem on hand, and in [38], where no such restriction is placed but approx-
imation methods are proposed due to dimensionality issues, we use a common
P H−distribution whose mean depends on the threshold parameters, K and N .
In Tables 1 through 5 we display the difference percentages, which are cal-
culated as 100| T hreshold−ShinM
ShinM oon
oon
|%. The results due to Shin and Moon [38]
correspond to the ones displayed in Tables 1 through 5 of their paper. Specif-
ically, we use their exact results as presented in their Tables 1 through 4, the
approximation denoted by L̂ in their Table 5.
364 Srinivas R. Chakravarthy

Table 1: Difference percentages for P (block)

ρ = 0.3 ρ = 0.5 ρ = 0.8
θ1 10 1 0.2 0.1 10 1 0.2 0.1 10 1 0.2 0.1 0.05
E2 0.37 0.18 0.07 0.40 0.01 0.05 0.10 0.12 0.21 0.39 0.40 0.85 1.06
H2 0.11 0.36 0.32 0.13 0.28 0.30 0.33 0.55 1.17 0.34 3.53 4.65 6.30

Table 2: Difference percentages for σ0

ρ = 0.3 ρ = 0.5 ρ = 0.8
θ1 10 1 0.2 0.1 10 1 0.2 0.1 10 1 0.2 0.1 0.05
E2 0.02 0.01 0.02 0.03 0.03 0.04 0.02 0.14 0.14 0.21 0.33 0.77 0.98
H2 0.00 0.03 0.00 0.02 0.03 0.02 0.06 0.13 0.66 0.14 3.02 4.13 5.46

Table 3: Difference percentages for µRO

θ1 106 10 1 0.2 0.1 0.05
E2 0.02 0.28 1.23 1.24 0.35 0.56
H2 0.02 0.66 0.61 0.81 1.40 1.83
Table 4: Difference percentages for µRO
ρ = 0.3 ρ = 0.5 ρ = 0.8
θ1 10 1 0.2 0.1 10 1 0.2 0.1 10 1 0.2 0.1 0.05
E2 0.37 1.30 1.02 0.66 0.28 1.23 1.16 0.35 0.95 3.14 6.71 8.18 8.81
H2 0.88 1.13 0.37 0.22 0.66 0.61 0.81 1.40 7.05 18.91 30.19 33.88 31.14

Table 5: Difference percentages for µRO

ρ = 0.5 ρ = 0.8
θ1 10 1 0.2 0.1 0.05 10 1 0.2 0.1 0.05
E4 4.66 2.27 0.26 0.95 1.80 5.28 4.37 10.68 12.77 13.72
M ER3 4.79 2.32 0.10 0.45 1.27 10.10 2.48 7.20 9.21 10.14
(1)
CE3,1 3.52 2.21 0.37 0.06 0.34 0.72 11.30 12.76 12.61 11.51
(2)
CE3,1 9.44 11.88 0.19 7.22 29.64 5.98 11.18 63.43 86.73 74.52
The above tables show the difference percentages for all but a few cases to
be reasonably acceptable. Those cases for which the percentages exceed 15%,
we recalculated the values by increasing the value of θ1 . The modified difference
percentages are displayed in Tables 4b and 5b, related, respectively, to Tables 4
and 5.
Table 4b: Modified difference percentages for µRO related to Table 4
ρ = 0.3 ρ = 0.5 ρ = 0.8
θ1 10 1 0.2 0.1 10 1 0.2 0.1 10 1 0.2 0.1 0.05
E2 0.37 1.30 1.02 0.66 0.28 1.23 1.16 0.35 0.95 3.14 6.71 8.18 8.81
∗ †
H2 0.88 1.13 0.37 0.22 0.66 0.61 0.81 1.40 7.05 7.13 4.24 5.23‡ 1.60§
∗
θ1 = 1.2;† θ1 = 0.26;‡ θ1 = 0.13;§ θ1 = 0.07.

Table 5b: Modified difference percentages for µRO related to Table 5

ρ = 0.5 ρ = 0.8
θ1 10 1 0.2 0.1 0.05 10 1 0.2 0.1 0.05
E4 4.66 2.27 0.26 0.95 1.80 5.28 4.37 10.68 12.77 13.72
M ER3 4.79 2.32 0.10 0.45 1.27 10.10 2.48 7.20 9.21 10.14
(1)
CE3,1 3.52 2.21 0.37 0.06 0.34 0.72 11.30 12.76 12.61 11.51
(2)
CE3,1 9.44 11.88 0.19 7.22 10.71∗ 5.98 11.18 12.10† 3.24‡ 14.15§
∗
θ1 = 0.075;† θ1 = 0.4;‡ θ1 = 0.2;§ θ1 = 0.09.
 A retrial queueing model with thresholds and phase type retrial times 365

It is clear from the above tables that our threshold model with a special way
of modeling the retrial times with P H−distribution appears to perform well. In
the next section, we will discuss additional examples by varying N .

5. Illustrative numerical examples

In this section we will discuss a few illustrative examples to bring out the
qualitative nature of our threshold model. Towards this end, we consider three
distributions for the retrial times. It should be pointed out these are special
types of P H−distributions.

A. E10 : This is an Erlang of order 10. The parameter will be chosen so that
the rate will be θk or θ, depending on the context. Thus, if the rate is fixed at
θ, then we have an Erlang of order 10 with parameter 10θ so that the mean of
this Erlang will be θ1 .
B. E1 : This is an exponential distribution whose parameter will be θk or θ de-
pending on the context.
C. H5 : This is hyperexponential with 5 states. The mixing probabilities and
the rates in each of those states are, respectively, (0.60, 0.24, 0.10, 0.05, 0.01) and
a(1000, 100, 10, 1, 0.1), where a = θk or a = θ depending on the context.
Note that the above three distributions are qualitatively different in that
they have different structure. While we will keep the means of these three
distributions to be the same (so as to compare properly), they have different
variance structure. The ratio of the standard deviations of E1 and H5 with
respect to E10 are, respectively, 3.1623 and 27.9492.
One can choose the values of θk , 1 ≤ k ≤ K, in a variety of ways. In our
examples below, we take θk = kθ1 , 2 ≤ k ≤ K, where θ1 will be is specified along
with θ.
EXAMPLE 1: In this example, we investigate the effect of the traffic intensity,
c, and the type of retrial distribution on the minimum value of θ for which the
λ
retrial queue traffic load will be close to a given value of ρ = cµ (see (30)).
1
Towards this end, we fix λ = 1, µ = cρ , and vary c and ρ. It is worth pointing
out that θ does not depend on N, K, and θk , 1 ≤ k ≤ K. The graph of the plot
of ln(θ) as a function of c and ρ under the three distributions for the retrial
times is given in Figure 1 below.
A quick look at Figure 1 reveals the following observations.
(1) Generally speaking, it appears that as c is increased (for fixed ρ), ln(θ)
(and hence θ) approaches a constant value for all ρ up to 0.8. However,
for ρ > 0.8, we see a decreasing trend in ln(θ) as c is increased. This
latter phenomenon can be explained intuitively as follows. When ρ is
large, it is more likely that an arriving customer will enter into the retrial
orbit and hence even a moderate value of ln(θ) will be sufficient for a
retrial customer to capture a free server.
366 Srinivas R. Chakravarthy

Figure 1. Ln(θ) under various scenarios

(2) Again, generally speaking, H5 retrial time with a high variability appears
to need a large ln(θ) compared to E10 and E1 retrial times, especially
for large values of ρ.
(3) For fixed c, we see an interesting behavior in ln(θ) as a function of ρ in
that this measure appears to decrease up to a certain point (depending
on the type of retrial and the value c) and then increases. The ”dips”,
if any, appear to occur for low to moderate values of ρ for c ≥ 2, while
for c = 1 the measure appears to increase as ρ is increased.
EXAMPLE 2: The purpose of this example is to compare the two measures,
λ
Pblock and µRO , when N is varied. Here we fix K = 100, λ = 1, θ1 = 1.0, µ = cρ ,
and vary c = 1, 2, 5, and ρ = 0.5, 0.95. The value of θ is obtained so as to have
the traffic intensity of the threshold retrial queue to be close to the given value
of ρ (see (30)). By considering the three distributions, E10 , E1 , and H5 , and
looking at N = 1, 2, 3, we display the two measures in Figure 2 below. It is
clear from this figure that while the measure Pblock appears to be not sensitive
to the three values of N considered when ρ = 0.5, we notice sensitivity to N
when ρ = 0.95 only in the case of H5 retrial times. This indicates the role of
variability in the retrial times. With respect to the measure, µRO , we see the
sensitivity to N for all scenarios considered. As is to be expected this measure
decreases as c is increased. Also, as seen in the classical queue, here also we see
the mean number in retrial orbit appears to increase with increasing variability
(in the retrial times).
A natural question that arises here is how should the retrial rate be increased
so that the mean number in retrial orbit for the threshold model under study
 A retrial queueing model with thresholds and phase type retrial times 367

Figure 2. Pblock and Ln(µRO ) under various scenarios

here will be close to the corresponding classical retrial queue. The next example
discusses this interesting question. Since there are no analytical results available
for multi-server retrial queue with phase type retrial times without placing any
restriction on the retrial model, we use the simulated model developed in [15] to
get µRO .

EXAMPLE 3: In this example we explore how the value of θ1 varies under

different scenarios so that the measure, µRO , is within 5% of the corresponding
value in the classical retrial queue. Due to lack of theoretical results for these
cases, we use the simulated values based on the model in [15]. Towards this end
λ
we fix λ = 1, µ = cρ , and vary c = 1, 2, 5, and ρ = 0.5, 0.95. First, we obtain the
value of θ (which does not depend on θ1 ) so as to have the traffic intensity of
the threshold retrial queue to be close to the given value of ρ (see (30)). By con-
sidering the three distributions, E10 , E1 , and H5 , and looking at N = 1, 2, 3, 4, 5
and K = 50, 60, · · · , 100, we display the value of θ1 in Figures 3 and 4 below.
An examination of the above two figures reveal that (a) as N is increased, θ1
appears to increase; (b) the higher the variation in retrial times, the larger the
368 Srinivas R. Chakravarthy

Figure 3. θ1 under various scenarios when ρ = 0.5

value of θ1 is required; (c) While the values of θ1 appear to increase when going
from ρ = 0.5 to ρ = 0.95, the increase is several fold for H5 retrials as compared
to the other two retrial times.
Note: The classical retrial models simulated in ([15]) are such that each or-
biting customer attempts to capture a free server using his/her own (identical)
P H−distribution. Thus, this example illustrates the worthiness of our thresh-
old model (with reasonable size state space) to approximate a classical retrial
queueing model with P H−distribution with proper choice of the retrial rates.
 A retrial queueing model with thresholds and phase type retrial times 369

Figure 4. θ1 under various scenarios when ρ = 0.95

In some applications it is of interest to see where the queueing system is spend-

ing most of the times as well as the state from where the customers have a high
probability of capturing a free server. In the next example we look at the two
measures, P SMr , 0 ≤ r ≤ K + 1 and ξr , 1 ≤ r ≤ K + 1.

EXAMPLE 4: The purpose of this example is to identify the values of K, N,

and r for which the probability that the system will be operating in mode r will
be the largest among K + 1 modes. Similarly, we identify the values for which
370 Srinivas R. Chakravarthy

the customer has the highest probability of capturing a free server. That is,
we look for (K1∗ , r1∗ , N1∗ , ξ ∗ ) and (K2∗ , r2∗ , N2∗ , P SM ∗ ), respectively, for the two
measures, ξ and P SM . These values are displayed in Tables 6 and 7, respec-
tively, under various scenarios. It should be pointed that we searched for these
by fixing λ = 1, θ1 = 1, θ = KN θ1 ; vary N from 1 to 10 and K from 1 to 50.
The value of µ is chosen so that the traffic intensity of the threshold  model is
close to a given value of ρ. That is, choose µ so that µ − ρ c πe  < 10−3 . λ
 P

j=1

Table 6: Mode for ξ

ρ
T oR c=1 c=2 c=5
E10 ( 1, 1, 10, 0.447 ) ( 2, 1, 10, 0.294 ) ( 1, 1, 6, 0.117 )
0.5 E1 ( 1, 1, 10, 0.408 ) ( 3, 1, 10, 0.292 ) ( 1, 1, 5, 0.121 )
H5 ( 1, 1, 10, 0.191 ) ( 1, 1, 10, 0.163 ) ( 4, 1, 10, 0.107 )
E10 ( 1, 2, 9, 0.819 ) ( 1, 2, 7, 0.730 ) ( 1, 2, 3, 0.626 )
0.95 E1 ( 1, 2, 10, 0.805 ) ( 1, 2, 8, 0.724 ) ( 1, 2, 4, 0.620 )
H5 ( 50, 51, 10, 0.753 ) ( 25, 26, 10, 0.676 ) ( 11, 12, 10, 0.546 )
Table 7: Mode for P SM
ρ T oR c=1 c=2 c=5
E10 ( 28, 1, 10, 0.724 ) ( 1, 0, 1, 0.691 ) ( 14, 0, 1, 0.882 )
0.5 E1 ( 1, 1, 10, 0.779 ) ( 1, 0, 1, 0.673 ) ( 24, 0, 1, 0.868 )
H5 ( 1, 1, 10, 0.698 ) ( 1, 1, 10, 0.653 ) ( 5, 0, 1, 0.687 )
E10 ( 1, 2, 1, 0.870 ) ( 1, 2, 1, 0.852 ) ( 1, 2, 1, 0.808 )
0.95 E1 ( 1, 2, 1, 0.875 ) ( 1, 2, 1, 0.850 ) ( 1, 2, 1, 0.792 )
H5 ( 1, 2, 1, 0.925 ) ( 1, 2, 1, 0.915 ) ( 1, 2, 1, 0.905 )
An examination of the above tables reveals the following.
• With respect to the probability of capturing a free server, we notice that
in the case of ρ = 0.5, both E10 and E1 retrial times have the largest
probability for a customer to capture a server. But for all the three
retrial times and for c = 1, 2, and 5, we notice that their respective
largest probability of capturing a free server occurs for customers from
the orbit. This is somewhat counter intuitive at first as the traffic load is
only 0.5 and so one would expect an arriving (new) customer to capture
a free server, especially, in multiple server case. However, the retrial
rate, θ, chosen for this example might be large enough that the retrial
customers capture a server more often than a new arriving customer.
• When ρ = 0.95, we notice a similar behavior for the probability of cap-
turing a free server; however, the values are significantly larger compared
to ρ = 0.5. Another interesting observation is that for all the three re-
trial times we see the largest probability of capturing a free server occurs
when the customers are seeking a free server when the system is operat-
ing in its last mode (i.e., namely when r = K + 1).
• The measure, P SM , indicates that the system seems to operate most of
the times in extreme modes.
 A retrial queueing model with thresholds and phase type retrial times 371

6. Concluding remarks
In this paper we introduced a new type threshold model in the context of
the classical retrial queue. Assuming the retrial times to be of phase type but
with the rates driven by the threshold parameters, we analyzed the model in
steady-state. Some illustrative numerical examples to bring out the qualitative
nature of the model were presented. We assumed that all the P H−distributions
governing the orbiting customers’ retrial attempts to have the same number of
phases but with different rates that depend on the threshold parameters. How-
ever, this assumption can be relaxed to have different orders (essentially different
P H−distributions). The needed modifications are in the matrices governing the
transitions from one threshold interval to the next (adjacent) one and all other
structure will still be preserved and one can analyze the model along the lines
described here. We also showed how our threshold model can be used to ap-
proximate a classical retrial queueing model with phase type retrials. The model
considered in this paper can be extended to include more versatile arrival pro-
cess as well as more robust service times. These will be topics for future research.

Acknowledgments: The author is grateful to the editor and the anonymous

referees for their constructive suggestions that improved the presentation of the
paper.

References
1. J.R. Artalejo, Accessible Bibliography on Retrial Queues, Mathematical and Computer
Modelling, 30 (1999), 1-6.
2. J.R. Artalejo, M.J. Lopez-Herrero, On the Busy Period of the M/G/1 Retrial Queue, Naval
Research Logistics 47 (2000), 115-127.
3. J.R. Artalejo, A. Gomez-Corral, M.F. Neuts, Analysis of multiserver queues with constant
retrial rate, Euro. J. Oper. Res. 135 (2002), 569-581.
4. J.R. Artalejo, S.R. Chakravarthy, Algorithmic Analysis of a M AP/P H/1 Retrial Queue,
TOP 14 (2006), 293-332.
5. J.R. Artalejo, A. Economou, M.J. Lopez-Herrero, Algorithmic approximations for the busy
period distribution of the M/M/c retrial queue, Euro. J. Oper. Res. 176 (2007), 1687-1702.
6. J.R. Artalejo, A. Gomez-Corral, Modelling communication systems with phase type service
and retrial times, IEEE Communications Letters 11 (2007), 955-957.
7. J.R. Artalejo, S.R. Chakravarthy, M.J. Lopez-Herrero, The busy period and the waiting
time analysis of a M AP/M/c queue with finite retrial group, Stochastic Analysis and Ap-
plications 25 (2007), 445-469.
8. J.R. Artalejo, A. Gomez-Corral, Retrial Queueing Systems: A Computational Approach,
Springer-Verlag, Berlin, Heidelberg, 2008.
9. K. Avrachenko, U. Yechiali, Retrail networks with finite buffers and their application to
internet data traffic, Probability in the Engineering and Informational Sciences 22 (2008),
519-536.
10. K. Avrachenko, U. Yechiali, On tandem blocking queues with a common retrial queue,
Computers and Operations Research 37 (2010), 1174-1180.
11. N. Baer, R.J. Boucherie, J-K. Ommeren, The P H/P H/1 multi-threshold queues, In: B.
Sericola, M. Telek, and G. Horvath (Eds.): ASMTA 2014, LNCS 8499, 95-109. Springer
International Publishing Switzerland, 2014.
372 Srinivas R. Chakravarthy

12. S.R. Chakravarthy, A. Krishnamoorthy and V.C. Joshua, Analysis of a Multi-server Re-
trial Queue with Search of Customers from the Orbit, Performance Evaluation 63 (2006),
776-798.
13. S.R. Chakravarthy, A Multi-server Queueing Model with Markovian Arrivals and Multiple
Thresholds, Asia-Pacific Journal of Operational Research 24 (2007), 223-243.
14. S.R. Chakravarthy, A multi-server synchronous vacation model with thresholds and a
probabilistic decision rule, European Journal of Operational Research 182 (2007), 305-320.
15. S.R. Chakravarthy, Analysis of M AP/P H/c retrial queue with phase type retrials - Simu-
lation approach, Communications in Computer and Information Science 356 (2013), 37-49.
16. B.D. Choi, Y.W. Shin, W.C. Ahn, Retrial queues with collision arising from unslotted
CSMA/CD protocol, Queueing Systems 11 (1992), 335-356.
17. D.I. Choi, T.S. Kim, S. Lee, Analysis of an MMPP/G/1/K Queue with Queue Length De-
pendent Arrival Rates, and its Application to Preventive Congestion Control in Telecom-
munication Networks, European Journal of Operational Research 187 (2008), 652-659.
18. J.E. Diamond, A.S. Alfa, Approximation method for M/PH/1 retrial queues with phase
type inter-retrial times, European Journal of Operational Research 113 (1999), 620-631.
19. D.V. Efrosinin, Controlled Queueing Systems with Heterogeneous Servers, Ph.D. Disser-
tation, Trier University, Germany, 2004.
20. D. Efrosinin, L. Breuer, Threshold policies for controlled retrial queues with heterogeneous
servers, Ann Oper Res 141 (2006), 139-162.
21. G.I. Falin, J.G.C. Templeton, Retrial Queues, Chapman and Hall, London, 1997.
22. A. Graham, Kronecker Products and Matrix Calculus with Applications, Ellis Horwood,
Chichester, UK, 1981.
23. Q.M. He, H. Li, Y.Q. Zhao, Ergodicity of the BM AP/P H/s/s + K retrial queue with
PH-retrial times, Queueing Systems 35 (2000), 323-347.
24. O.C. Ibe, J. Keilson, Multi-server threshold queues with hysteresis, Performance Evalua-
tion 21 (1995), 185-2135.
25. S.B. Khodadadi, F. Jolai, A fuzzy based threshold policy for a single server retrial queue
with vacations, Central European Journal of Operations Research 20 (2012), 281-297.
26. J. Kim, B. Kim, A survey of retrial queueing systems, Ann. Oper. Res. 247 (2016), 3-36.
27. G. Latouche, V. Ramaswami, Introduction to matrix analytic methods in stochastic mod-
eling, SIAM, 1999.
28. W. Lin, P.R. Kumar, Optimal Control of a Queueing System with two Heterogeneous
Servers IEEE Trans. on Autom. Control 29 (1984), 696-703.
29. H. Luh, I. Viniotis, Threshold control policies for heterogeneous server systems, Mathe-
matical Methods of OR 55 (2002), 121-142.
30. M. Marcus, H. Minc, A Survey of Matrix Theory and Matrix Inequalities, Allyn and
Bacon, Boston, MA, 1964.
31. M.F. Neuts, Matrix-geometric solutions in stochastic models: An algorithmic approach,
The Johns Hopkins University Press, Baltimore, MD. [1994 version is Dover Edition], 1981.
32. R. Nobel, H.C. Tijms, Optimal Control of a Queueing System with Heterogeneous Servers,
IEEE Transactions on Autom. Control 45 (2002), 780-784.
33. V. Ponomarov, E. Lebedev, Finite Source Retrial Queues with State-Dependent Service
Rate, Communications in Computer and Information Science 356 (2013), 140146.
34. V. Ponomarov, E. Lebedev, Optimal Control of Retrial Queues with Finite Population
and State-Dependent Service Rate, Advances in Intelligent Systems and Computing 754
(2018), 359-369.
35. V.V. Rykov, D.V. Efrosinin, Numerical Analysis of Optimal Control Polices for Queueing
Systems with Heterogeneous Servers, Information Processes 2 (2002), 252-256.
36. M. Senthilkumar, K. Sohraby, K. Kim, On a multiserver retrial queue with phase type
retrial time, Mathematical and Computational Models. Eds: R. Nadarajan et al., 65-78,
2012, Narosa Publishing House, New Delhi, India.
 A retrial queueing model with thresholds and phase type retrial times 373

37. Y.W. Shin, Algorithmic solutions for M/M/c retrial queue with PH2 retrial time, Journal
of Applied Mathematics and Informatics 29 (2011), 803-811.
38. Y.W. Shin, D.H. Moon, Approximation of M/M/c retrial queue with PH-retrial times,
European Journal of Operational Research 213 (2011), 205-209.
39. W-H. Steeb and Y. Hardy, Matrix Calculus and Kronecker Product, World Scientific Pub-
lishing, Singapore, 2011.
40. T. Yang, M.J.M. Posner, J.G.C. Templeton, H. Li, An approximation method for the
M/G/1 retrial queues with general retrial times, European Journal of Operational Research
76 (1994), 552-562.

Srinivas R. Chakravarthy received B.Sc. in Mathematics, from the University of Madras,

India; M.Sc. in Statistics from the University of Madras, India; and Ph.D. in Operations
Research, from the University of Delaware, Newark, US. Since 1983, he has been in the
Departments of Industrial and Manufacturing Engineering & Mathematics, Kettering Uni-
versity , Flint, Michigan, USA. His research interests include queueing, inventory, reliability,
matrix-analytical methods, algorithmic probability, and simulation.
Departments of Industrial and Manufacturing Engineering & Mathematics, Kettering Uni-
versity , Flint, Michigan, USA.
e-mail: [email protected]