European Journal of Operational Research 256 (2017) 187195

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Stochastics and Statistics

Multi-server tandem queue with Markovian arrival process,

phase-type service times, and nite buffers
Hendrik Baumann a, Werner Sandmann b,
a
Department of Applied Stochastics and Operations Research, Clausthal University of Technology, Erzstr. 1, Clausthal-Zellerfeld 38678, Germany
b
Department of Computer Science, Saarland University, Campus E1 3, Saarbrcken 66123, Germany

a r t i c l e i n f o a b s t r a c t

Article history: We consider multi-server tandem queues where both stations have a nite buffer and all services times
Received 21 September 2015 are phase-type distributed. Arriving customers enter the rst queueing station if buffer space is avail-
Accepted 16 July 2016
able or get lost otherwise. After completing service in the rst station customers proceed to the second
Available online 21 July 2016
station if buffer space is available, otherwise a server at the rst station is blocked until buffer space
Keywords: becomes available at the second station. We provide an exact computational analysis of various steady-
Queueing state performance measures such as loss and blocking probabilities, expectations and higher moments
Multi-server tandem queue of numbers of customers in the queues and in the whole system by modeling the tandem queue as a
Markovian arrival process level-dependent quasi-birth-and-death process and applying suitable matrix-analytic methods. Numerical
Phase-type service time distributions results are presented for selected representative examples.
Matrix-analytic method
2016 Elsevier B.V. All rights reserved.

1. Introduction exponentially distributed and interarrival times are often neither

Tandem queueing networks, in short: tandem queues, are
widely used to model systems and scenarios where services are
delivered in successive stages, that is generic customers enter a
multi-stage service facility and successively proceed through the
service stages. This applies to such diverse elds as manufactur-
ing, transportation, logistics, computer networking, telecommuni-
cations, and many daily-life service operations, amongst others.
Consider for example production lines where different steps are
processed by different machines, airplane maintenance and refu-
eling, ordering and delivery of goods or services, the Internet or
other communication networks where messages or data packets
are routed from a source to a destination via multiple hops, or su-
permarkets where the rst stage is the self service of customers
collecting their items and the second stage is the payment at the
check out.
Queueing theory has extensively dealt with tandem queues
for more than ve decades. However, standard assumptions of
classical queueing theory, especially Poisson arrival processes and
independent exponentially distributed service times, do not ap-
propriately reect the trac characteristics, service or processing
times in many modern applications. Service times are often not
Corresponding author.
E-mail addresses: hendrik.baumann@tu-clausthal.de (H. Baumann), werner.
sandmann@uni-saarland.de (W. Sandmann).
independent, identically distributed nor exponentially distributed.
Thus, the arrival process is not even a renewal process. Therefore,
we consider phase-type distributed service times and Markovian
arrival processes.
Phase-type (PH) distributions Neuts (1981), OCinneide
(1990) are extensions of the exponential distribution with the
particularly nice property that any probability distribution on
the nonnegative real numbers can be approximated, in principle
arbitrarily accurately, by a PH distribution. Markovian arrival pro-
cesses (MAPs) Neuts (1979), Lucantoni, Meier-Hellstern, and Neuts
(1990) can properly incorporate correlations between successive
interarrival times. They include Poisson processes, PH renewal
processes, interrupted and Markov modulated Poisson processes
as special cases and the whole class of MAPs is dense within the
class of marked point processes Asmussen and Koole (1993). For
monographs with a special focus on PH distributions and MAPs we
refer to Breuer and Baum (2005) and Buchholz, Kriege, and Felko
(2014), where the latter also addresses how data can be tted to
PH distributions and MAPs, respectively.
PH distributions and MAPs provide versatile means of model-
ing at the cost of a signicantly increased complexity in that the
state space of the associated Markov chain model becomes ex-
tremely huge. In particular, using PH distributed service times in
multi-server models considerably increases the dimension of the
state space. This curse of dimensionality poses a great challenge
on model analysis, which is often tackled via approximations. For
instance, Brandwajn and Begin (2014) have recently approached

http://dx.doi.org/10.1016/j.ejor.2016.07.035
0377-2217/ 2016 Elsevier B.V. All rights reserved.
188 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195
the seemingly simple case of a multi-server queue with Poisson
arrivals, PH distributed service times, and nite buffer by using a
reduced state description in which the state of only one server is
represented explicitly, while the other servers are accounted for
through their rate of completions. In fact, queueing models with
both MAP and PH distributed service times, even for single sta-
tion single-server queues, are typically analyzed by approxima-
tions using state space reduction approaches and/or sophisticated
matrix-analytic methods. Different variants of single-server queues
with MAP and PH distributed service times have been studied in,
e.g., Artalejo and Chakravarthy (2006), Krishnamoorthy, Babu, and
Narayanan (2009), Sreenivasan, Chakravarthy, and Krishnamoorthy
(2013). Multi-server queues with MAP and PH distributed service
times are addressed in Chakravarthy (2013) where the MAP/PH/c
retrial queue with PH distributed retrials is studied via simu-
lation, and in Kim, Dudin, Taramin, and Baek (2013b) where a
MAP/PH/c/c+K queue with two customer classes is analyzed by
combining specic state space reduction techniques for queues
with PH distributed service times and matrix-analytic methods.
Only relatively few works have investigated tandem queues
with MAP and PH distributed service times. Gomez-Corral
(2002) analyzes a tandem queue with MAP and two single-server
stations, innite buffer at the rst station and no buffer at the sec-
ond station. Service times at the rst station are PH distributed and
the service times at the second station are general. Van Houdt and
Alfa (2005) consider a (discrete-time) tandem queue with MAP and
two single-server stations where the service times at both stations
are PH distributed, the queue at the rst station is innite and the
queue at the second station is nite. Lian and Liu (2008) deal with
a tandem queue with MAP, two single-server stations with queues
of innite capacity and exponentially distributed service times at
both stations. Baumann and Sandmann (2013) consider a tandem
queue with MAP, two single-server stations with queues of innite
capacity and PH distributed service times at both stations where
customers may leave the system after completing service at the
rst station. Kim, Dudin, Dudin, and Dudina (2013a) model a call
center as a tandem queue with MAP, multiple servers at both sta-
tions, exponentially distributed service times and no buffer at the
rst station, PH distributed service times and a nite buffer at the
second station. Kim, Dudin, Dudina, and Dudin (2014) investigate
a tandem queue with two types of customers, marked MAP, multi-
ple servers at both stations, exponentially distributed service times
and no buffer at the rst station, PH distributed service times, one
nite buffer and one innite buffer at the second station.
In this paper, we consider tandem queues with MAP, two multi-
server stations with nite buffers and PH distributed service times
at both stations where at the rst station losses of arriving cus-
tomers and blocking after service can occur. We compute steady-
state performance measures without introducing any approxima-
tion. For this purpose we model the class of tandem queues under
consideration by level-dependent quasi-birth-and death (LDQBD)
processes and use a matrix-analytic method according to Baumann
and Sandmann (2013) for obtaining the desired steady-state perfor-
mance measures without explicitly computing the stationary dis-
tribution. In the next section, we formally introduce the class of
tandem queues considered, in particular we give general formal
descriptions of the arrival process and the service time distribu-
tions. In Section 3 we describe the modeling of this class of tan-
dem queues as LDQBD processes, where we introduce the structur-
ing of the multi-dimensional state space as well as the state transi-
tions and transition rates. Subsequently, in Section 4 we show how
to express relevant performance measures and in Section 5 we
present a matrix-analytic algorithm for the ecient computation
of these performance measures. Numerical results for specic tan-
dem queues of the considered type are presented in Section 6.
In particular, for certain choices of the arrival process, the ser-
vice time distributions and varying parameter values we study the
loss and blocking probabilities as well as the expected numbers of
customers in the two stations of the tandem queueing network.
Section 7 concludes the paper and outlines further research direc-
tions.
2. Tandem queue specication
The class of the tandem queue we investigate can be specied
in Kendall notation by MAP/PH/c1 /c1 + K1 /PH/c2 /c2 + K2 . The
rst station consists of c1 identical servers and a buffer of nite
capacity K1 . The second station consists of c2 identical servers and
a buffer of nite capacity K2 . Customers arrive at the rst queue-
ing station. If upon arrival the buffer is fully occupied, then the
customer is lost. Customers having completed service in the rst
station proceed to the second station. If upon service completion
in the rst station the buffer at the second station is fully occu-
pied, then the customer blocks a server at the rst station until
buffer space becomes available at the second station.
The service times at both stations are PH distributed. A prob-
ability distribution on R+ = [0, ) is a (continuous) PH distribu-
tion, iff it is the distribution of the time until absorption in a -
nite time-homogeneous continuous-time Markov chain (CTMC). For
such a CTMC with state space {1, . . . , n + 1} and initial distribution
= (1 , . . . , n+1 ), where the states 1, . . . , n are transient and the
state n + 1 is absorbing, the generator matrix A has the form
 
B b
A= , B Rnn , b Rn (1)
0 0
and (1 , . . . , n , B ) is a representation of the PH distribution, in
short PH( , B), with = (1 , . . . , n ). The number n of transient
states is called the order of the PH distribution or the number of
phases. In our model we denote the service time distributions at
the rst and the second station by PH( , S) and PH( , T), respec-
tively, where PH( , S) is of order V and PH( , T) is of order W.
Hence, S = (si j )Vi, j=1 and T = (ti j )W i, j=1
. Besides, we dene the vec-
tors s := S1 and t := T 1. The moments of a PH( , B) distributed
random variable Z are given by E[Z k ] = k! (B )k 1. Hence, the
expected service times in our model are S1 1 at the rst sta-
tion and T 1 1 at the second station.
The arrival process is a continuous-time MAP of order U de-
termined by the pair of (U U)-matrices (D0 , D1 ) where D0 =
(i j )Ui, j=1 , D1 = (i j )Ui, j=1 with ij 0 for i = j, ij < 0 for i =
j, ij 0 for all i, j {1, 2 . . . , U }, D1 = 0, and D0 1 + D1 1 = 0
such that D = D0 + D1 is the generator matrix of an irreducible
CTMC on the state space {1, 2, . . . , U }. This CTMC is commonly
referred to as environmental process, phase process, or back-
ground CTMC. When it changes its state from state i to state
j = i, with probability i j /(i j + i j ) it triggers an arrival and
with probability i j /(i j + i j ) it does not. The average arrival
rate of the MAP is given by = D1 1 where denotes the
stationary distribution of the environmental process, i.e. D =
0, 1 = 1. The stationary moments of the interarrival time Y are
E[Y k ] = k!1 (D0 )1k 1. In particular, E[Y ] = 1 , and Var[Y ] =
(2 (D0 )1 1 1 )/2 , from which we get the squared coecient
of variation c2 = 2 (D0 )1 1 1. The squared coecient of cor-
relation is given by = ( (D0 )1 D1 D1 1 1 )/c2 .
0
3. Markov chain model
We model the tandem queue as a CTMC (X (t ))t0 =
(X1 (t ), . . . , XV +W +6 (t ))t0 whose states are integer-valued vectors
of the form
(n1 , n2 , u, v0 , v1 , . . . , vV , vV +1 , w0 , w1 , . . . , wW )
where
 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195 189

Table 1
Transitions in the CTMC model of the tandem queue.

Event Condition State changes Rate

Arrival v0 > 0, u = 1, . . . , U, i = 1, . . . , V u u , v0 v0 1, v i vi + 1 u,u i

v = 0, n < K , u = 1, . . . , U
0 1 1 u u , n1 n1 + 1 u,u
n1 = K1 , u = 1, . . . , U u u u,u
Change of MAP state u = 1, . . . , U u u u,u
First station Service completion n1 = 0, w0 > 0, i = 1, . . . , V, k = 1, . . . , W v 0 v0 + 1, v i vi 1, w0 w0 1, wk wk + 1 vi si k
n1 > 0, w0 > 0, i, j = 1, . . . , V, k = 1, . . . , W n1 n1 1, (vi , v j ) (vi 1, v j + 1 ) for j = i, w0 w0 1, wk wk + 1 vi si j k
n1 = 0, n2 < K2 , w0 = 0, i = 1, . . . , V v 0 v 0 + 1, v i v i 1, n2 n2 + 1 vi si
n1 > 0, n2 < K2 , w0 = 0, i, j = 1, . . . , V n1 n1 1, n2 n2 + 1, (vi , v j ) (vi 1, v j + 1 ) for j = i vi si j
n2 = K2 , i = 1, . . . , V vi vi 1, vk+1 vk+1 + 1 vi si
Change of phase i, j = 1, . . . , V, j = i v i vi 1, v j v j + 1 vi si j
Second station Service completion n2 = 0, k = 1, . . . , W w0 w0 + 1, wk wk 1 wk tk
n2 > 0, vV +1 = 0, k, = 1, . . . , W n2 n2 1, (wk , w ) (wk 1, w + 1 ) for = k wk tk
vV +1 > 0, n1 = 0, k, = 1, . . . , W w0 w0 + 1, vV +1 vV +1 1, (wk , w ) (wk 1, w + 1 ) for = k wk tk
vV +1 > 0, n1 > 0, i = 1, . . . , V, k, = 1, . . . , W n1 n1 1, vi vi + 1, vV +1 vV +1 1, (wk , w ) (wk 1, w + 1 ) for = k wk tk i
Change of phase k, = 1, . . . , W, = k w k wk 1, w w + 1 wk tk

n1 , n2 are the numbers of customers waiting at the rst and Changes of phases are quite clear. The factor vi is explained by
second station, respectively, i.e. the numbers of customers in the fact that there are vi servers operating in phase i. The factor
the buffers, wk is explained analogously.
u is the environmental state of the MAP, With rate u,u , an arrival occurs and the MAP goes to state u .
v0 , w0 are the numbers of free servers at the rst and second For v0 > 0, there is at least one idle server at the rst station. An
station, respectively, arriving customer will head for one of the idle servers and there-
vi , w j for 1 i V, j = 1, . . . , W are the number of servers op- fore decrease the number of idle servers by 1. With probability
erating in phase i at the rst station and the number of servers i , the new service starts in phase i. For v0 = 0 and n1 < K1 , all
operating in phase j at the second station, respectively, servers are busy, and arriving customers are forced to wait, that
vV +1 is the number of blocked servers at the rst station due is they enter the buffer. Hence, the number of waiting customers
to lack of buffer space at the second station. increases by 1. For n1 = K1 , all servers are busy, and there is no
buffer space available. Thus, arriving customers are lost.
The (V + W + 6 )-dimensional state space S NV +W +6 of the
If a server at the rst station operates in phase i, the service
CTMC is given by the set of all integer-valued vectors of this form
completion rate is si . The number of servers operating in phase i
subject to the conditions
is vi . Hence, the rate for service completion of one of the servers
1 u U, operating in phase i is given by vi si .
0 n1 K1 , 0 n2 K2 , If no customer waits at the rst station (n1 = 0), the number
n1 = 0 or v0 = 0, of idle servers increases by 1. Otherwise, the server starts imme-
n2 = 0 or w0 = 0, diately with the next service, the number n1 of waiting customers
vi 0 for 0 i V + 1 and wi 0 for 0 i W, decreases by 1, and the server starts in phase j with probability
v0 + v1 + + vV + vV +1 = c1 and w0 + w1 + + wW = c2 , j . For j = i, i is decreased by 1 and j is increased by 1. If
vV +1 = 0 or n2 = K2 . j = i, then i does not change. For w0 > 0, there is at least one
For computing the number of states, we remark that accord- idle server at the second station. After service completion at the
c +k rst station, customers will head for one of the idle servers at
ing to simple elementary combinatorics there are 1k partitions
the second station and the service starts in phase k with proba-
c1 = 0 + + k with 0 , . . . , k 0. By distinguishing for the rst
bility k . For w0 = 0 and n2 < K2 , all servers at the second sta-
station between n1 = 0 and n1 > 0 and for the second station be-
tion are busy and the number n2 of waiting customers at the sec-
tween n2 = 0 and 0 < n2 < K2 and n2 = K2 , we obtain the size of
ond station is increased by 1. For n2 = K2 , there is no buffer space
the state space
    available at the second station. Hence, after service completion at
c1 + V 1 c2 + W 1 the rst station, customers cannot leave the server and will block
|S | = U
V 1 W 1 it. In this case, the number vV +1 of blocked servers is increased
 (c + V )(c + W ) c1 + V by 1.
+ (K2 1 )
1 2
If a server at the second station operates in phase k, the service
VW V
completion rate is tk . The number of servers operating in phase
(c1 + V )(c1 + V + 1 ) c2 + W k is wk , and hence, the rate for service completion of one of the
+ + K1
V (V + 1 ) W servers operating in phase k is wk tk .
c1 + V
 For n2 = 0, no customer waits at the second station. Hence, the
+ K1 (K2 1 ) + K1
V number of idle servers is increased by 1. Otherwise, the number
    of waiting customers decreases by 1 and the new service starts in
c1 + V c2 + W
U K1 K2 . phase with probability . For vV +1 > 0, implying n2 = K2 , ser-
V W vice completion at the second station has impact on the rst sta-
All transitions from state (u, n1 , v0 , v1 , . . . , vV , vV +1 , n2 , w0 , w1 , tion: A customer leaves the system, service for another customer
. . . , wW ) that change this state, their effects, rates, and conditions (waiting at the second station) starts, and thus, there is a buffer
under which the transitions occur are given in Table 1. Note that space available at the second station. vV +1 > 0 implies that there
formally, all transitions in the second (third) block require vi > 0 are customers blocking servers at the rst station after service
(wk > 0). We can omit these conditions here due to the fact that completion. One of these customers will join the buffer at the sec-
the transition rates have the value 0 for vi = 0 or wk = 0. In the ond station so that still n2 = K2 . The corresponding server will be-
following we explain the transitions. come idle if there are no customers waiting at the rst station or
190 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195

will start a new service, which then starts in phase i with proba-
bility i .
4. Performance measures
In order to describe relevant steady-state queueing network
performance measures we consider the limit X (t ) X ( ) =: X =
(X1 , . . . , XV +W +6 ) and the corresponding limiting probability distri-
bution that coincides for any ergodic CTMC with the unique sta-
tionary probability distribution.
Let (n1 , n2 , u, v0 , v1 , . . . , vV , vV +1 , n2 , w0 , w1 , . . . , wW ) denote
the stationary probability of the state (n1 , n2 , u, v0 , v1 , . . . , vV , vV +1 ,
w0 , w1 , . . . , wW ) and u := u1 + u2 + + uU , u = 1, . . . , U.
For notational convenience we dene (n1 , n2 , u, v, w ) :=
(n1 , n2 , u, v0 , v1 , . . . , vV , vV +1 , w0 , w1 , . . . , wW ). The stationary
k-th moments  of the numbers of customers in the buffers are
E[X1k ] and E X2k , respectively, which are given by

the higher moments. Furthermore, according to (5) and (7) also
loss probabilities and blocking probabilities can be expressed by
expectations. As we shall see, expressing and computing perfor-
mance measures via expectations is in many regards much nicer
than doing so via the stationary probabilities.
An obvious approach to obtaining the performance measures is
by rst computing the stationary probability distribution (or an ap-
proximation thereof) and then using the stationary probabilities for
calculating the desired expectations according to (2), (3), (4) and
(6). However, in many cases, in particular for our tandem queue
with MAP and PH distributed service times, the whole stationary
distribution is not relevant. More specically, the stationary prob-
abilities of each state are typically not of practical interest, e.g.,
the stationary distribution includes the probabilities that services
are in specic phases and the like. In fact, the practical interest is
in marginal probabilities, loss and blocking probabilities, expecta-
tions, higher moments as dened above, and in cumulants such as
variances, skewness, or kurtosis, all of which can be composed of

K1

K2

c1

c1

c2

c2
higher moments. Moreover, from a computational point of view, of
E X1k = nk1 (n1 , n2 , u, v, w ),
course, obtaining the whole stationary distribution requires a lot
n1 =0 n2 =0 u=1 v0 =0 vV +1 =0 w0 =0 wW =0
of storage that can easily exceed the memory resources even of
(2) todays computers.

K1

K2

c1

c1

c2

E X2k =
n1 =0 n2 =0 u=1 v0 =0 vV +1 =0 w0 =0
where (n1 , n2 , u, v, w ) := 0 for (n1 , n2 , u, v, w )
/ S. Note that
these moments in particular include the marginal probabilities (for
k = 0) and the expectations (for k = 1) of the numbers of cus-
tomers in the buffers. Furthermore, cumulants such as variances
and skewnesses can be obtained easily from the higher moments.
The stationary probability that an arriving customer is lost can
be expressed as the proportion of the rate of arrivals while Xt(1 ) =
K1 and the total arrival rate. Hence, it is given by
c2
nk2 (n1 , n2 , u, v, w ), 5. Computational solution technique
wW =0
Two key points in our following analysis are:
(3)
1. We partition the multidimensional state space such that the
model can be expressed as a level-dependent quasi-birth-and-
death (LDQBD) process, that is, a CTMC with block tridiagonal
generator matrix where the matrix structure can be exploited
for ecient computational analysis.
2. We compute the expectations that describe the relevant per-
formance measures by a memory-ecient matrix-analytic al-
gorithm without explicitly computing the stationary distribu-
  tion or a huge set of stationary probabilities. In this manner
1

K2

c1

c1

c2

c2
we avoid huge storage requirements for stationary probabilities
ploss = (K1 , n2 , u, v, w )u
n2 =0 u=1 v0 =0 vV +1 =0 w0 =0 wW =0
so that it becomes indeed possible to compute the performance
measures eciently without approximation errors.
(4)
  5.1. State space partitioning
1

U
u
= E I{X1 =K1 ,X3 =u} u =E I{X1 =K1 ,X3 =u} (5)
u=1 u=1
We partition the multidimensional state space S of the CTMC
according to the number of customers in the system, that is, the
where I{ } denotes an indicator function. Similarly, the probability number of customers in both queues and all servers, which is n1 +
that a customer completing service at the rst station blocks this n2 + c1 v0 + c2 w0 . Denote by N the maximum possible number
station, because there is no free space at the second station, can of customers in the system and for all n {0, 1, . . . , N} by S (n ) the
be expressed as the proportion of the rate of service completions
set of all states in which exactly n customers are in the system.
at the rst station while Xt(2 ) = K2 and the total service completion
Hence, N = c1 + K1 + c2 + K2 and
rate at the rst station. Hence, this probability is given by
S (n ) : = {(n1 , n2 , u, v0 , v1 , . . . , vV , vV +1 , w0 , w1 , . . . , wW ) S : n1
pblock
K1 U
c1 c1 c2 c2 V + n 2 + c1 v0 + c2 w 0 = n }, (8)
n1 =0 v0 =0 . . . vV +1 =0 w0 =0 . . . wW =0 (n1 , K2 , u, v, w ) i=1 si vi
u=1
=
K1 K2 U c1 c1 c2 c2 V which implies S (i ) S ( j ) = for i = j and S = S (0 ) S (1 )
n1 =0 n2 =0 u=1 v0 =0 . . . vV +1 =0 w0 =0 . . . wW =0 (n1 , n2 , u, v, w ) i=1 si vi
S (N ) . We refer to the number of customers in the system in this
(6) context as the level and call S (n ) the nth level set. It is clear from
 1 the description of the tandem queue and in particular from the

V
= E I{X2 =K2 } si Xi+4 E si Xi+4
i=1 i=1
An important observation with regard to the eciency of our
following analysis is that all these performance measures can be
expressed conveniently in terms of expectations with respect to
the stationary probability distribution, i.e. the probability vector
containing all stationary probabilities of the states. The mean num-
bers of customers in the queues are obviously expectations, as are
possible transitions of the corresponding CTMC, given in Table 1,
. (7)
that any event in the tandem queue changes the number of cus-
tomers in the system (i.e. the level) by at most one and that cor-
respondingly any direct state transition in the CTMC is either be-
tween two states within the same level set or between two states
in adjacent (with regard to the level number) level sets. Thus, any
state transition is from a state in S (i ) to a state in S ( j ) , |i j| 1.
This means, according to the state space partitioning the states can
be arranged such that the generator matrix of the CTMC is of block
 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195 191

tridiagonal form Since the state space S is discrete (countable) we can inter-
pret f() for any {1, . . . , m} as a column vector, partitioned
Q00 Q01
conformally with the state space partitioning, that is, f ( ) =
Q10 Q11 Q12
( f1() , . . . , fN() )T with fn() = ( fn,(1) , . . . , fn,d
( ) T
) for n = 0, 1, . . . , N,
Q =

Q21 Q22 Q23 (9) n

..
( )
.. .. where fn,i for i = 1, . . . , dn denotes the value of the function f()
. . . applied to the ith state (according to the numbering in lexico-
QN,N1 QNN graphical order) in the nth level set S (n ) . Finally, we use the nota-
di d j tion fn := ( fn(1 ) , . . . , fn(m ) ) Rdn m for n = 0, 1, . . . , N. Similarly, we
where Qi j R , i, j = 0, 1, . . . , N, contains the transition rates
consider the stationary probability distribution of the LDQBD
from states in level set i to states in level set j, and dn := |S (n ) | process partitioned conformally with the state space partitioning,
for all n {0, 1, . . . , N}. A CTMC with a block tridiagonal generator that is, = (0 , 1 , . . . , N ) with n = (n,1 , n,2 , . . . , n,dn ), n =
matrix is called a level-dependent quasi-birth-and-death (LDQBD) 0, 1, . . . , N, where n, i for i = 1, . . . , dn denotes the stationary
process. Hence, with our state space partitioning we have modeled probability of the ith state (according to the numbering in lex-
the multi-server tandem queue under consideration as an LDQBD icographical order) in the nth level set S (n ) . Then, as shown by
process. Although with regard to the block tridiagonality of the Baumann and Sandmann (2013),
generator matrix the numbering of states within the level sets can  
be arbitrary, we assume that they are numbered in lexicographical

N 
n 1
order. Since all transition rates are properly included in Table 1 we E[ f ] = f = n f n = 0 Rk fn (11)
omit to describe the block matrices Qij explicitly. n=0 n=0 k=0

5.2. Computing performance measures via expectations of the LDQBD = 0 ( f0 + R0 ( f1 + R1 ( f2 + R2 (. . . + RN2 ( fN1 + RN1 fN ))))) (12)

As outlined in Section 4, in principle one could compute per- where R0 , . . . , RN1 are nonnegative matrices dened by
formance measures by rst computing the stationary probability
Rn = Qn,n+1 (Qn+1,n+1 + Rn+1 Qn+2,n+1 )1 , n = 0, . . . , N 2, (13)
distribution of the underlying CTMC and then using it for com-
puting the desired performance measures from it. For our case
1
of an LDQBD process fast and numerically stable algorithms for RN1 = QN1,N QNN . (14)
computing stationary probability distributions exist, see Baumann
These equations can be derived from (backward) block-Gaussian
and Sandmann (2010, 2012); Bright and Taylor (1995), Gaver, Ja-
elimination for block-tridiagonal matrices, see Grassmann
cobs, and Latouche (1984), but they have enormous storage re-
and Heyman (1990) for details: In the reduction step, Qnn
quirements for models where the dimension is as high and the size
is replaced by Un , where Un = Qn,n+1 (Un+1 )1 Qn+1,n . By
of the state space is as huge as for the tandem queue under con-
setting Rn = Qn,n+1 (Un+1 )1 , we obtain (13) and (14). Fur-
sideration. Following our argumentation in Section 4 we instead
apply the memory-ecient method developed by Baumann and
thermore, block-Gaussian elimination guarantees n+1 =
Sandmann (2013) for computing stationary expectations without
n Qn,n+1 (Un+1 )1 = n Rn . Since the recursions for Rn and
explicitly computing the stationary probability distribution, which
n are well-established in the context of matrix-analytic methods
(see references above), we omit further details.
is as fast and numerically stable as the aforementioned algorithms
Note that the matrices Rn appear in numerical techniques for
but provides an enormous complexity reduction in that it saves
computing stationary probability distributions of LDQBD processes
a huge amount of storage. It renders possible to compute per-
where all these matrices must be stored, but when computing
formance measures of such huge models as, e.g., a multi-server
family E[f] according to (12) the term
tandem queue with Markovian arrival process, phase-type service
times, nite buffers, losses, and blocking. In fact, we can even com- Z0 := f0 + R0 ( f1 + R1 ( f2 + R2 (. . . + RN1 ( fN1 + RN1 fN )))) (15)
pute multiple performance measures simultaneously.
Assume we want to compute multiple performance measures can be computed in a Horner-type fashion starting with ZN = fN
simultaneously. Then we express all these performance mea- and repeatedly updating
sures together as one expectation family E[f] of a function f =
( f (1) , . . . , f (m) ) : S Rm where the components f (1) , . . . , f (m) are Zn = fn + Rn Zn+1 (16)
the functions whose expectations are taken to express the desired
for n = N 1, N 2, . . . , 0, so that all Rn , and all Zn are required
performance measures. For instance, if we want to compute the
only once and need not be stored for any further operation. Finally,
expected steady-state numbers of customers in the buffers, the loss
0 required in (12) for determining E[ f ] = 0 Z0 can be computed
probability and the blocking probability simultaneously, then we
as solution x to the equation
choose
 

x(Q00 + R0 Q10 ) = 0, (17)
U
u

V
V
f = X1 , X2 , I{X1 =K1 ,X3 =u} , si Xi+4 , I{X2 =K2 } si Xi+4
u=1
i=1 i=1
where the rst three components of family E[f] directly yield the
expected numbers of customers in the rst and second buffer, re-
spectively, and the loss probability. The blocking probability can
be composed by the fourth and fth component according to for-
mula (7). Similarly, if we additionally want to compute the steady-
state variances of the numbers of customers in the buffers we can
extend the function f by the components X12 and X22 , and com-
pose the variances as Var[Xi ] = E[Xi2 ] E[Xi ]2 , i = 1, 2. Of course,
the same principle works for arbitrary cumulants, all of which can
be composed by moments of appropriate orders.
(10)
which is unique up to a multiplicative constant c given by the
normalization condition that the sum of all state probabilities
must be one. In order to obtain the normalization constant with-
out explicitly computing all stationary probabilities the func-
tion f is extended to f = ( f (0 ) , f (1 ) , . . . , f (m ) ) where f(0) is the
constant function that takes the value one for all states such
that the corresponding column vector representation is f n(0 ) =
(1, 1, . . . , 1 )T Rdn . Then the normalization condition can be ex-
pressed as E[ f (0 ) ] = 1 and for f a vector Z = (z(0 ) , z(1 ) , . . . , z(m ) ) =
cE[ f] is obtained. Thus, the expectation of f, that is, the vector con-
taining all expectations needed for the performance measures that
we consider simultaneously, is given by
192 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195

bii = n, bi j = n for j = i + 1, and bi j = 0, otherwise. The ex-

1
pectation is E[Z] = and the squared coecient of variation (cov)
is c2 = 1n . Hence, if specic values of the expectation and/or the
cov are prescribed or assumed, as it is often the case in practice,
it is easy to set the values of n and accordingly. Erlang distri-
butions with a (relatively) large number of phases are often used
to approximate deterministic behavior, because the Erlang distri-
bution is the least variable PH distribution in the sense that for
any n its squared cov is minimal in the class of all PH distributions
with n phases (see Aldous & Shepp (1987)), and obviously with in-
creasing n the cov decreases, approaching zero for n approaching
innity.
A hyperexponential distribution is a PH( , B) distribution where
Fig. 1. Algorithm for computing multiple performance measures simultaneously.
B is a diagonal matrix. We consider hyperexponential distribu-
tions with two phases, we set = (1/2 1/2 ). Then, for B =
  diag(1 , 2 ) with any choice of 1 , 2 > 0, the expectation
z (1 ) z (m )
E[ f ] = ,..., . (18) and squared cov are
z (0 ) z (0 )  
1 1 1 1 1
With thenotation just introduced, the algorithm given in Fig. 1 E[Z] = + =: , c =2 2
+ 1.
2 1 2 2 21 22
computes E f and, thus, all desired performance measures are
obtained without computing the whole stationary distribution. As Similarly as for Erlang distributions, if the expectation or the cov
pointed out above, there is no need for storing all the matrices Rn should be set to specic values, then the values of 1 , 2 can be
and Zn . Instead, we replace Rn+1 and Zn+1 by Rn and Zn , respec- chosen accordingly as
tively, that is, we use matrices R and Z which are updated corre-
sponding to the recursions for Rn and Zn , respectively. 1 =  , 2 =  . (20)
c2 1 c2 1
Furthermore, each of the matrices Qij and fn is used only once. 1 2
1+ 2
These matrices should be constructed when needed for updat-
ing R and Z and are not stored. Therefore, our algorithm is very For the arrivals, as specic MAPs we consider Markov-
memory-ecient and renders possible to analyze multiple perfor- modulated Poisson processes (MMPPs) and PH renewal processes.
mance measures simultaneously. The maximum number of real A MAP is MMPP if D1 is a diagonal matrix, that is, arrivals do not
numbers to be stored at the same time is cause a state change for the environmental process. In our numer-
ical study, we will use MMPPs with U = 2,
max (dn dn+1 + dn (m + 1 ) ). (19)    
n=0,...,N 1 1 1 1 0
D0 = and D1 = .
Especially for a large maximum possible number of customers in 1 2 1 0 2
the system, that is, for large buffer sizes and/or many servers the
For given average arrival rate > 0 and squared cov c2 1 for the
algorithm is much more ecient than similar matrix-analytic al-
interarrival times, we choose
gorithms that rst compute the whole stationary distribution of 
the underlying CTMC and subsequently compute the performance c2 1
measures of the tandem queue. In fact, the algorithm is as fast as 1 = + ( + 2 ) , 2 = 2 1 . (21)
c2 + 1
counterparts for computing stationary distributions while having
far less storage requirements. By setting D0 = B and D1 = b , the MAP becomes a PH( , B) re-
newal process, that is, the interarrival times are iid PH( , B) dis-
tributed.
6. Numerical results

In this section, we present numerical results for performance
measures of specic MAP/PH/c1 /c1 + K1 /PH/c2 /c2 + K2 tandem
queues. We emphasize that with the approach presented in the
previous section, one can conduct detailed studies of any such tan-
dem queue, but we have to restrict the presentation of numerical
results to selected choices of MAPs and PH distributions. In par-
ticular, we study in detail the expectations of the numbers of cus-
tomers in the queues, loss and blocking probabilities, where we
vary the buffer sizes as well as the respective parameters of the
MAPs and the PH distributions considered.
6.1. Specication of the service time distributions and the arrival
process
For the service times, we consider Erlang and hyperexponen-
tial distributions, which are discussed extensively in, e.g., Morse
(1958), Neuts (1981), so that here it is sucient to summarize the
most important properties.
An Erlang distribution with n phases and parameter n is a
PH( , B) distribution with = (1 0 . . . 0 ) and B = (bi j ) where
6.2. Base parameter setting
If not indicated otherwise (by the corresponding variation of
parameters),
the arrival process is a MMPP with U = 2, average arrival rate
= 19 and squared cov 1.2 for the interarrival times.
the service at each server at the rst station is Erlang dis-
tributed with 2 phases and expected service time 1/7.
there are c1 = 3 servers at the rst station.
the waiting capacity at the rst station is K1 = 10.
the service at each server at the second station is Erlang dis-
tributed with 2 phases and expected service time 1/5.
there are c2 = 4 servers at the second station.
the waiting capacity at the second station is K2 = 10.
6.3. Variation of the arrival rate
In our rst variation of parameters, we vary the arrival rate
at the rst station, while keeping the other parameters xed as
specied above.
 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195 193

Fig. 2. Performance measures for various arrival rates.

Fig. 3. Performance measures for various values of K1 .

As can be seen in Fig. 2, unsurprisingly, with increasing arrival
rate, the expected numbers of customers in both stations (includ-
ing those in service) as well as the probabilities of loss and block-
ing increase. However, in particular for the respective probabili-
ties we observe that the slopes of the functions are quite different.
Due to an increasing arrival rate the probability of a full buffer at
the rst station, which causes losses of newly arriving customers,
grows quickly and correspondingly the loss probability does. For
the blocking probability the increase is slower, because blocking
only occurs if upon service completion at the rst station there is
no buffer space available at the second station. This implies that
even if the rst station is highly saturated (i.e. all servers are al-
most always busy), it can, of course, not serve faster than at its
overall (multi-server) service rate. Thus, for the departure rate out
of the rst station there is a maximum and if this is reached then
a further increase in the arrival rate at the rst station still causes
more and more losses but the departure rate out of the rst station
does not increase more and more. Therefore, at a certain value of
the arrival rate the rst station becomes a serious bottleneck im-
plying that the loss probability grows further while the blocking
probability remains almost constant.
Clearly, in practice, for instance when dimensioning a system
with multi-server nite-buffer tandem queue structure this effect
can be taken into account using numerical studies with our algo-
rithm in order to nd, e.g., buffer sizes that yield acceptable block-
ing and loss probabilities.
6.4. Variation of the buffer sizes
While keeping all other parameters xed as specied before, we
vary the buffer size either at the rst or at the second station.
As depicted in Fig. 3, for increasing buffer size K1 at the rst
station, the expected number of customers in the rst station in-
creases, and the probability of loss decreases. Up to some value of
the buffer size K1 at the rst station more customers are served
and intended to proceed to the second station, and hence, the
blocking probability and the expected number of customers at the
second station increase. However, similarly as for the variation in
the arrival rate, when the rst station becomes more and more sat-
urated it cannot serve on a higher rate than its overall service rate
such that the curves showing the number of customers at the sec-
ond station and the blocking probability become atter and at-
ter, which means both the number of customers at the second sta-
tion and the blocking probability approach a respective nite max-
imum.
Fig. 4 shows that for increasing buffer size K2 at the second
station, up to some value of K2 the expected number of customers
in the second station increases and then approaches some nite
194 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195

Fig. 4. Performance measures for various values of K2 .

Fig. 5. Performance measures for various values of the squared cov of the interarrival times.

constant value, and the probability of blocking decreases, nally
approaching zero. More precisely, for K2 there is always
enough buffer space available at the second station such that there
is no blocking of servers at the rst station. Besides, the arrival rate
at the second station cannot exceed the arrival rate at the rst sta-
tion. Thus, for K2 the second station becomes essentially a
stable innite capacity queue such that the expected number of
customers converges to a nite value. As for the rst station, due
to fewer customers blocking the servers there, the expected num-
ber of customers at the rst station and the loss probability de-
crease with increasing K2 . For K2 , the rst station behaves
like a MAP/PH/c1 /c1 + K1 queue without blocking, that is, the ex-
pected number of customers in the rst station and the probability
of loss converge to some nite value.
6.5. Variation of coecients of variation
One great advantage of our model is that we can vary arrival
or service rate and higher moments of the interarrival or ser-
vice times, respectively, separately. We vary the squared cov c2
for the interarrival times, keeping all other parameters, in partic-
ular the arrival rate, xed as specied in Section 6.2 by exploit-
ing that MAPs include as special cases many classes of arrival pro-
cesses with specic interarrival time distributions. In our studies,
for c2 = 1, we use a Poisson arrival process (that is, exponentially
distributed interarrival times), for c2 < 1, we use PH renewal pro-
cesses with Erlang distributed interarrival times, and for c2 > 1, we
use second-order MMPPs as arrival processes. Note that this choice
is not unique but our general framework of MAPs is very exible.
For instance, an alternative choice for obtaining interarrival times
with c2 > 1 would be a PH-type renewal process with hyperexpo-
nentially distributed interarrival times.
Fig. 5 shows the performance measures considered before, now
as a function of the squared cov of the interrarival times accord-
ing to the just described choices. Clearly, for increasing cov, the
arrival process becomes more random, and the probability of loss
at the rst station increases. Up to some value of the (squared) cov
c2 the expected number of customers in the rst station increases,
too, but then slowly decreases again. This observation can be ex-
plained by the fact that the arrival rate is constant and the loss
probability increases with increasing cov, which implies that fewer
customers enter the system. At the second station, for increasing
c2 , the expected number of customers decreases. Again, this is due
to the fact that fewer customers enter the system. Hence, it seems
to make sense that the blocking probability should decrease, too.
However, this is not true for very small values of the cov. Any-
way, as can be seen in the respective curve in the picture on the
left hand side of Fig. 5, the impact of the cov on the blocking
 H. Baumann, W. Sandmann / European Journal of Operational Research 256 (2017) 187195
probability is quite weak, that is the blocking probability remains
almost constant and also for increasing cov decreases only very
slowly such that the effect of the cov is negligible. However, we
emphasize that the inuence of the cov on the expected number
of customers (in particularly, in the second station), and on the loss
probability is quite strong.
6.6. Variation of the service time distributions
In the same way as for the interarrival times we have studied
the impact of varying service rates and varying covs of the service
times. We omit to present the details here and instead, we sum-
marize some results concerning variations at the rst station.
For increasing service rate at the rst station, the expected
number of customers at the rst station and the loss probability
decrease. Subsequently, more customers arrive at the second sta-
tion, and since we keep the service rate at the second station con-
stant, both the expected number of customers in the second sta-
tion and the blocking probability increase.
For increasing cov, the rst station becomes more random.
Hence, the expected number of customers in the rst station and
the loss probability increase. As compared to the impact of the
cov of the interarrival times, the loss probability increases slower.
Therefore, the expected number of customers in the rst station
does not decrease with increasing cov of the service times at the
rst station and thus the blocking probability increases, too. The
increasing loss probability has an impact on the expected number
of customers in the second station in that fewer customers enter
the system which implies that the expected number of customers
in the second station decreases.
It turns out that the impact of the service rates in our exam-
ples is stronger than the impact of the cov of the service times,
but nevertheless, the results clearly indicate that the inuence of
the cov cannot be neglected in practical applications. In particular,
this means that not only the service rates matter but also the spe-
cic service time distributions. It is therefore particularly impor-
tant to have a model that is as exible as our model (being able to
cope with arbitrary PH distributions) and to have an algorithm that
can eciently analyze the model and compute many relevant per-
formance measures for various different parameter settings, arrival
processes, interarrival and service time distributions.
7. Conclusion
We have presented a computational analysis of steady-state
performance measures for multi-server tandem queues with
Markovian arrival process, nite buffers, and phase-type dis-
tributed service times at both queueing nodes, where losses and
blocking can occur. Our approach is based on the structured
modeling of such tandem queues as level-dependent quasi-birth-
and-death processes and using suitable computationally ecient
matrix-analytic algorithms. This approach is conceptually exact,
that is, our modeling does not introduce any approximation such
as, e.g., decomposition approaches, and the computational analysis
using the matrix-analytic algorithm is exact up to numerical errors.
It applies to any tandem queue of the form MAP/PH/c1 /c1 + K1
/PH/c2 /c2 + K2 , that is, to any specic choice of the MAP, the PH
distributions of the service times, and the nite buffer sizes. We
have presented detailed numerical studies of the impact of various
parameters with a focus on the loss probability, the blocking prob-
ability, and the expected numbers of customers in the two stations.
Future research aims at generalizations and extensions of
our computational analysis approach and its applicability. This
means to increase even further the model complexity and the
size/dimension of the state space that can be handled. For instance,
195
in a next step the approach may be extended to series of multi-
server queues with MAP and PH distributed service times and/or
to respective models with innite buffers. In order to cope with
such complex models it is planned to combine the matrix-analytic
algorithm used in the present paper with state space truncation
methods similar to those in Baumann and Sandmann (2015) in or-
der to develop an approximate analysis with a priori error bounds.
Acknowledgments
We like to thank the three anonymous referees for their valu-
able comments and suggestions.
References
Aldous, D., & Shepp, L. (1987). The least variable phase type distribution is Erlang.
Communications in Statistics - Stochastic Models, 3(3), 467473.
Artalejo, J. R., & Chakravarthy, S. R. (2006). Algorithmic analysis of the MAP/PH/1
retrial queue. Top, 14(2), 293332.
Asmussen, S., & Koole, G. (1993). Marked point processes as limits of Markovian
arrival streams. Journal of Applied Probability, 30, 365372.
Baumann, H., & Sandmann, W. (2010). Numerical solution of level dependent
quasi-birth-and-death processes. Procedia Computer Science, 1(1), 15611569.
Baumann, H., & Sandmann, W. (2012). Steady state analysis of level dependent
quasi-birth-and-death processes with catastrophes. Computers & Operations Re-
search, 39(2), 413423.
Baumann, H., & Sandmann, W. (2013). Computing stationary expectations in lev-
el-dependent QBD processes. Journal of Applied Probability, 50(1), 151165.
Baumann, H., & Sandmann, W. (2015). Bounded truncation error for long run av-
erages in innite Markov chains. Journal of Applied Probability, 52(3), 609
621.
Brandwajn, A., & Begin, T. (2014). Reduced complexity in M/Ph/c/N queues. Perfor-
mance Evaluation, 78, 4254.
Breuer, L., & Baum, D. (2005). An introduction to queueing theory and matrix-analytic
methods. Springer.
Bright, L., & Taylor, P. G. (1995). Calculating the equilibrium distribution in level
dependent quasi-birth-and-death processes. Stochastic Models, 11(3), 497525.
Buchholz, P., Kriege, J., & Felko, I. (2014). Input modeling with phase-type distribu-
tions and Markov models theory and applications. Springer briefs in mathe-
matics. Springer.
Chakravarthy, S. R. (2013). Analysis of MAP/PH/c retrial queue with phase type re-
trials simulation approach. In A. Dudin, V. Klimenok, G. Tsarenkov, & S. Dudin
(Eds.), Proceedings of the Belarusian winter workshops in queueing theory, BWWQT
(pp. 3749). Springer-Verlag.
Gaver, D. P., Jacobs, P. A., & Latouche, G. (1984). Finite birth-and-death mod-
els in randomly changing environments. Advances in Applied Probability, 16(4),
715731.
Gomez-Corral, A. (2002). A tandem queue with blocking and Markovian arrival pro-
cess. Queueing Systems, 41(4), 343370.
Grassmann, W. K., & Heyman, D. P. (1990). Equilibrium distribution of block-struc-
tured Markov chains with repeating rows. Journal of Applied Probability, 27(3),
557576.
Kim, C., Dudin, A., Dudin, S., & Dudina, O. (2013a). Tandem queueing system with
impatient customers as a model of call center with interactive voice response.
Performance Evaluation, 70, 440453.
Kim, C., Dudin, A., Dudina, O., & Dudin, S. (2014). Tandem queueing system with
innite and nite intermediate buffers and generalized phase-type service time
distribution. European Journal of Operational Research, 235, 170179.
Kim, C., Dudin, S., Taramin, O., & Baek, J. (2013b). Queueing system MAP/PH/N/N+R
with impatient heterogeneous customers as a model of call center. Applied
Mathematical Modelling, 37, 958976.
Krishnamoorthy, A., Babu, S., & Narayanan, V. C. (2009). The MAP/(PH/PH)/1 queue
with self-generation of priorities and non-preemptive service. European Journal
of Operational Research, 195, 174185.
Lian, Z., & Liu, L. (2008). A tandem network with MAP inputs. Operations Research
Letters, 36, 189195.
Lucantoni, D. M., Meier-Hellstern, K. S., & Neuts, M. F. (1990). A single server queue
with server vacations and a class of non-renewal arrival processes. Advances in
Applied Probability, 22, 676705.
Morse, P. M. (1958). Queues, inventories and maintenance. John Wiley & Sons.
Neuts, M. F. (1979). A versatile Markovian point process. Journal of Applied Probabil-
ity, 17, 764799.
Neuts, M. F. (1981). Matrix-geometric solutions in stochastic models. Johns Hopkins
University Press.
OCinneide, C. A. (1990). Characterization of phase-type distributions. Communica-
tions in Statistics - Stochastic Models, 6(1), 157.
Sreenivasan, C., Chakravarthy, S. R., & Krishnamoorthy, A. (2013). MAP/PH/1 queue
with working vacations, vacation interruptions and N policy. Applied Mathemat-
ical Modelling, 37, 38793893.
Van Houdt, B., & Alfa, A. S. (2005). Response time in a tandem queue with block-
ing, Markovian arrivals and phase-type services. Operations Research Letters, 3,
373381.