Computers & Operations Research 29 (2002) 1353}1374

Approximate solution for multi-server queueing systems with

Erlangian service times
Marcos Escobar , Amedeo R. Odoni
, Emily Roth *
McKinsey & Company, Inc., Arquimedes 130, 4th Floor, Polanco, Mexico D.F. 11560, Mexico

Operations Research Center, Room 33-217, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
Department of Mathematical Sciences, Bentley College, 175 Forest Street, Waltham, MA 02452-4705, USA
Received 1 January 2000; received in revised form 1 November 2000

Abstract

Multi-server queueing systems with Poisson arrivals and Erlangian service times are among the most
applicable of what are considered `easya systems in queueing theory. By selecting the proper order,
Erlangian service times can be used to approximate reasonably well many general types of service times
which have a unimodal distribution and a coe$cient of variation less than or equal to 1. In view of their
practical importance, it may be surprising that the existing literature on these systems is quite sparse. The
probable reason is that, while it is indeed possible to represent these systems through a Markov process,
serious di$culties arise because of (1) the very large number of system states that may be present with
increasing Erlang order and/or number of servers, and (2) the complex state transition probabilities that one
has to consider. Using a standard numerical approach, solutions of the balance equations describing systems
with even a modest Erlang order and number of servers require extensive computational e!ort and become
impractical for larger systems. In this paper we illustrate these di$culties and present the equally likely
combinations (ELC) heuristic which provides excellent approximations to typical equilibrium behavior
measures of interest for a wide range of stationary multiserver systems with Poisson arrivals and Erlangian
service. As system size grows, ELC computational times can be more than 1000 times faster than those for the
exact approach. We also illustrate this heuristic's ability to estimate accurately system response under
transient and/or dynamic conditions.  2002 Elsevier Science Ltd. All rights reserved.

1. Introduction

Multi-server queueing systems with Poisson arrivals and Erlangian service times (M/E /n and
I
variations) are among the most applicable of what are considered `easya systems in queueing

* Corresponding author. Tel.: #1-781-891-2990; fax: #1-780-891-2457.

E-mail address: eroth@bentley.edu (E. Roth).

0305-0548/02/$ - see front matter  2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 3 0 5 - 0 5 4 8 ( 0 1 ) 0 0 0 3 6 - 3
 1354 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
theory. The Erlangian family, E , of distributions encompasses service times ranging from the
2 I
3 negative exponential (k"1) to constant (for in"nite k), while the opportunity to set the number of
4 parallel and independent servers, n, to an appropriate integer value adds #exibility to the model.
5 Moreover, by selecting the proper value of k, Erlangian service times can be used to approximate
6 reasonably well many general types of service times which have a unimodal distribution and
7 a coe$cient of variation less than or equal to 1.
In view of their practical importance, it may be surprising that the existing literature on M/E /n
8 I
9 systems is quite sparse. The probable reason is that, while it is indeed possible to represent these
10 systems through a Markov process (hence the characterization as `easya), serious di$culties arise
11 because of (1) the very large number of system states that may be present with increasing Erlang
12 order, k, and increasing number of servers, n, and (2) the complex state transition probabilities that
13 one has to consider (Section 2). This makes it di$cult to compute numerical solutions to systems
14 with even modest values of k and n*while obtaining general closed-form expressions seems
15 intractable. The practitioner is thus left with little to go on.
16 In this paper we attempt to address this paucity of practical tools. Our principal result is
a heuristic approximation scheme for solving M(t)/E (t)/n/n#q systems, where q is the queue
17 I
18 capacity. We show that, under stationary conditions, our heuristic provides a computationally
e$cient and tractable way for approximating extremely well the exact solutions to M/E /n/n#q
19 I
20 systems and, by implication, to some of the M/G/n/n#q queueing systems which arise in practice.
21 The heuristic also solves e$ciently systems with a large value of q, thus making it possible to
22 compute steady-state solutions (when such solutions exist) for many systems with in"nite queue
23 capacity. Finally, a number of initial computational experiments has indicated that the heuristic
performs very well in the dynamic case (M(t)/E (t)/n/n#q).
24 I
As noted above, the existing literature on M/E /n systems and their variations ("nite queue
25 I
26 capacity, non-stationary parameters, etc.) is limited for n'1. The standard classical papers are
27 those of Shapiro [1] and Mayhugh and McCormick [2] which provide a state description for
M/E /n with unlimited queue size, list the steady-state balance equations, describe approaches to
28 I
29 solving these equations and prove that the equations have unique solutions. Neither paper
30 provides closed-form expressions. Murray and Kelton [3] use a complicated, specially tailored
approach to solve M/E /2 systems. It is di$cult to extend their technique to more than two servers.
31 I
32 Maal~e [4] proposes two approximate expressions for the mean waiting time in steady state for
M/E /n systems (unlimited queue size). These approximations work well for some combinations of
33 I
34 k, n and the system utilization ratio
, and not well for others. Smith [5] provides approximate
solutions for the steady-state distribution of the number of customers in M/E /n systems for
35 I
36 k"2,3 and R using the approach suggested by Hokstad [6]. Finally, some well-known, approx-
37 imate steady-state results for various performance measures of M/G/n systems with in"nite or "nite
capacities (see, e.g., [7}16]) can be applied with varying degrees of success to systems of the M/E /n
38 I
39 type. Escobar [17] provides a detailed review of these approximation methods and compares their
40 performance for several selected cases. None of these approaches provides (numerical) solutions to
the general M(t)/E (t)/n/n#q system as we attempt to do in this paper.
41 I
42 The paper is organized as follows: Sections 2.1}2.3 discuss the exact approach to the solution of
M/E /n/n#q systems. This is necessary in order, "rst, to explain the complications mentioned
43 I
44 above and, second, to introduce two alternative state descriptions of these systems. While the "rst,
`traditionala description is more natural, the second provides a good starting point for developing
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1355
1
2 our heuristic approach. This second description requires a `pattern generatora algorithm, detailed
3 in Appendix A, to generate all states of the system. Section 2.4 then presents the heuristic approach.
4 The heuristic is based on the idea of working with a simpli"ed and `condenseda version of the
5 second state description. Section 3 provides a partial summary of results from a large number of
numerical experiments we have performed with M/E /n/n#q systems, under a large variety of
6 I
7 utilization rates and settings of system parameters, in order to validate our heuristic under
8 stationary conditions. This section also brie#y describes a smaller number of numerical experi-
9 ments aimed at validating the heuristic under dynamic conditions. A concluding section summar-
10 izes our "ndings and identi"es areas for future investigation. The paper also includes an Appendix
11 B, which illustrates both the exact and heuristic solution approaches for a speci"c system, the
M(t)/E (t)/3/4 queue. Readers are encouraged to review this appendix because it greatly facilitates
12 
13 understanding of the paper.
14
15
2. The M(t)/E (t)/n and M(t)/E (t)/n/n#q queueing systems
16 I I
17
In this section, we describe in detail how to obtain exact solutions for the M(t)/E (t)/n and
18 I
M(t)/E (t)/n/n#q queueing systems, as well as a heuristic solution technique for such systems. We
19 I
start by reviewing the method of stages used to represent the Erlang distribution in the M(t)/E (t)/1
20 I
21 system to enable solution of such systems. Then, we extend this approach to the case of multiple
22 servers with limited or unlimited queue size. Note that all the results in this section apply under
23 both stationary and non-stationary conditions, unless otherwise speci"ed.
24
25 2.1. State description for single-server systems
26
27 Each customer that enters the system can be considered as a package of k consecutive,
28 exponentially distributed tasks to be performed by the service facility. The service rate for each
29 stage is k(t), with a corresponding expected time of 1/k(t) per stage. The service rate for
30 completing all stages is (t) with a corresponding expected service time of 1/(t).
31 The usefulness of this approach is that we can derive a state transition diagram with indepen-
32 dent, exponentially distributed transitions that completely describes the queue. Due to the mem-
33 oryless property of Poisson processes, in any time increment t, the state can change only as
indicated in the diagram. In the M(t)/E (t)/1 system, the states are de"ned fully by the total number
34 I
35 of stages (or tasks) remaining in the system to be completed for all customers. Fig. 1 shows the state
36 transition diagram for this queue. For clarity, the time dependence has been omitted.
37 Using Fig. 1, we can derive the Chapman}Kolmogorov equations describing the probability
38 distribution of the number of stages and, therefore, the number of customers in the system over
39 time. We can, thereby, compute such performance measures such as expected queue length and
40 expected waiting time.
41
42 2.2. State description for multi-server systems
43
44 If the system has multiple servers, the total number of stages is not su$cient to de"ne the state of
the system because, for a particular number of stages, the distribution of such stages among the
 1356 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
2
3
4
5
6
7
Fig. 1. M(t)/E (t)/1 queue.
8 I
9
10
11 servers may not be unique. Instead, the state must identify: the number of uncompleted stages in the
12 system; the number of customers in service and in the queue; and the distribution of uncompleted
13 stages among the busy servers.
14 We use two di!erent, but equivalent, state representations in this work. The "rst, proposed
15 recently by Lee [18] (and di!ering only slightly from those of Shapiro [1] and Mayhugh and
McCormick [2]) is a (k#1)-tuple state description of the form (a , a , , a , a ) where
16 I I\ 2  O
a indicates the number of servers with i stages remaining to be completed, for 1)i)k, and
17 G
a indicates the number of customers in the queue, waiting for service. We shall refer to this as
18 O
19 Description 1.
20 Using Description 1, the number m of customers in the system is given by


21
I a , a "0
22 m" G G O
23 n#a , a '0
O O
24
25 and the total number of stages left in the system, l, is
26
27 I
l" a i#a k.
28 G O
G
29
30 We now present a more compact state description, Description 2. De"ne the state by a three
31 element descriptor (l, m, r), where l and m are de"ned as above and r is the `pattern identi"era
32 needed when (l, m) is not unique. Note that the number of busy servers is given by the minimum of
33 m and n. To better explain the existence and generation of multiple patterns for a given combina-
34 tion of l and m, consider the following example.
Let the state of the system be (11,5, r) in an M(t)/E (t)/5/5#q queue (see Fig. 2). In this state, 11
35 
36 uncompleted stages from 5 customers are distributed among the 5 available servers. This can occur
37 in one of three ways, shown by the patterns (a), (b) and (c) of Fig. 2 in which the number of shaded
38 circles in each row denotes the number of uncompleted stages at a particular server. For example,
39 pattern (a) has one server with three stages remaining to be completed and four servers with two
40 stages remaining.
41 The two representations, Description 1 and Description 2, are equivalent. For example, state
42 (1,4,0,0) of Description 1 is equivalent to the state (11,5,a) of Description 2, as shown in Fig. 2.
43 Description 1 is more convenient for generating the Chapman}Kolmogorov equations. Descrip-
44 tion 2 motivates the heuristic presented in Section 2.4. It must, however, be used in conjunction
with an algorithm that generates all of the system states such as the one described in Appendix A.
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1357
1
2
3
4
5
6
7
8 Fig. 2. Alternative patterns for the state (11,5, r) in a M(t)/E (t)/5 queue.
9 
10
11 2.3. Exact solution technique
12
13 In this section, using Description 1 and extending Lee's [18] work, we write the equations
needed to obtain the state probabilities of the M(t)/E (t)/n and M(t)/E (t)/n/n#q systems. Let
14 I I
S be the array containing the state probabilities when m(n, and let S (a 2, a ) be the
15   I 
probability of state (a ,2, a ,0) for which I a (n. Let Q be the state probability array of
16 I  G G 
states (a ,2, a ,0) when I a "n, and m"n, with the state probabilities speci"ed by
17 I  G G
Q (a ,2, a ).
18  I 
Similarly, let Q be the state probability arrays for the case in which a '0, and let Q (a ,2, a )
19 Q O Q I 
be the state probabilities of state (a ,2, a , s) when there are s customers waiting for service in the
20 I 
21 queue. The total number of customers in the system in this case is n#s. If the system has in"nite
22 queue size, then 1)s(R. If the queue size is limited, 1)s)q, where q is the maximum number
of customers that can wait for service and the array Q has the state probabilities Q (a ,2, q )
23 O O I 
24 when there are n#q customers in the system.
25
26 2.3.1. Total number of states
27 The total number of states in the system that need to be considered is given by
28



n#k n#k!1
29 ¹ " #q (1)
30 1 n n
31
as proved by Lee [18]. Note that the "rst term indicates the number of states when the queue is
32
empty, and the second term indicates the number of states for the customers waiting in the queue.
33
Table 1 shows various combinations of k, n and q and, in column 4, their corresponding number of
34
states which increase with the order of max(n#q)!, q(n#k!1)!.
35
36
37 2.3.2. State-to-state transitions
38 A transition between states occurs due to a stage completion (rate k(t)) or to an arrival of a new
39 customer to the system (rate (t)). Table 2 shows the transitions for each type of state and their
40 corresponding state-to-state transition rates. These state transitions lead directly to the Chap-
man}Kolmogorov equations for the M(t)/E (t)/n and M(t)/E (t)/n/n#q systems.
41 I I
(A detailed example of the M(t)/E (t)/3/4 queueing system is presented in Appendix B including
42 
43 the state transition diagram and the associated state-to-state transition probabilities. We strongly
44 recommend that the reader consult this appendix before proceeding in order to appreciate the
complexity of the state transitions even for such a small system as the M(t)/E (t)/3/4 queue.)

 1358 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
Table 1
2
Number of states for various k, n and q
3
4 k n q Exact Heuristic
5
6 4 3 1 55 32
7 5 5 5 882 171
3 15 5 1496 411
8 4 15 5 7956 606
9 5 15 5 34,884 801
10 3 50 25 56,576 5126
11 10 10 5 646,646 961
12 4 50 25 901,901 7651
13
14
15
16 Table 2
17 State-to-state transitions for the exact solution technique
18
From state In To state In With rate
19
20 (a ,2, a ,0) S (a #1, a , , a ,0) S (t) if m(n!1
I   I I\ 2  
21 (a #1, a , , a ,0) Q (t) if m"n!1
I I\ 2  
22 (a !1, a #1, a ,2, a ,0) S a k(t)
I I\ I\   I
23 (a , a !1, a #1,2, a ,0) S a k(t)
I I\ I\   I\
(a ,2, a !1, a #1,2, a ,0) S a k(t) for i"k22
24 I G G\   G
(a , a ,2, a !1,0) S a k(t)
25 I I\   
(a ,2, a ,0) Q (a ,2, a ,1) Q (t)
26 I   I  
(a ,2, a !1, a #1,2, a ,0) Q a k(t) for i"k22
27 I G G\   G
(a , a ,2, a !1,0) S a k(t)
28 I I\   
(a ,2, a , s) Q (a ,2, a , s#1) Q (t)
29 I  Q I  Q>
(1)s(q) (a ,2, a !1, a #1,2, a , s) Q a k(t) for i"k22
30 I G G\  Q G
(a #1, a , , a !1, s!1) Q a k(t)
I I\ 2  Q\ 
31
(a ,2, a , q) Q (a ,2, a !1, a #1,2, a , q) Q a k(t) for i"k22
32 I  O I G G\  O G
(q(R) (a #1, a , , a !1, q!1) Q a k(t)
33 I I\ 2  O\ 
34
35
36
37 Generalizing the equations presented in Lee [18], the equations for solving the state probabilit-
ies are divided into "ve cases. When I a "m(n!1, the equations are
38 G G
39
SQ (a ,2, a )(t)"!((t)#mk(t))S (a ,2, a )(t)#(t)S (a !1,2, a )(t)
40  I   I   I 
41

42 # (a #1)k(t)S (a ,2, a #1, a !1,2, a )(t)
G  I G G\ 
43 GI
44
#(a #1)k(t)S (a , a , , a #1)(t). (2)
I  I I\ 2 
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1359
1
When I a "m"n!1, they are given by
2 G G
3 SQ (a ,2, a )(t)"!((t)#mk(t))S (a ,2, a )(t)#(t)S (a !1,2, a )(t)
4  I   I   I 
5 
# (a #1)k(t)S (a ,2, a #1, a !1,2, a )(t)
6 G  I G G\ 
7 GI
#(a #1)k(t)Q (a , a , , a #1)(t) (3)
8 I  I I\ 2 
9 and
10
11 I
SQ (a ,2, a )(t)"0 if a Om where 0)m)n!1. (4)
12  I  G
13 G
14 Once the number of customers equals the number of servers, m"n, the equations are
15
QQ (a ,2, a )(t)"!((t)#nk(t))Q (a ,2, a )(t)#(t)S (a !1,2, a )(t)
16  I   I   I 
17 
# (a #1)k(t)Q (a ,2, a #1, a !1,2, a )(t)
18 G  I G G\ 
19 GI
20 #(a #1)k(t)Q (a , a , , a #1)(t) (5)
I  I I\ 2 
21
and
22
23 I
24 QQ (a ,2, a )(t)"0 if a On. (6)
 I  G
25 G
26 The equations of the state probabilities when there are customers waiting for service, m'n, are
27 given by
28
QQ (a ,2, a )(t)"!((t)#nk(t))Q (a ,2, a )(t)#(t)Q (a ,2, a )(t)
29 Q I  Q I  Q\ I 
30 
31 # (a #1)k(t)Q (a ,2, a #1, a !1,2, a )(t)
G Q I G G\ 
32 GI
33 #(a #1)k(t)Q (a !1, a , , a #1)(t) (7)
I Q> I I\ 2 
34
35 and
36 I
37 QQ (a ,2, a )(t)"0 if a On, (8)
Q I  G
38 G
39 where 1)s(q, with q"R if the queue size is unlimited. Finally, if q(R, then the following
40 set of equations are needed when the number of customers in the system is m"n#q:
41
42 QQ (a ,2, a )(t)"!nk(t)Q (a ,2, a )(t)#(t)Q (a ,2, a )(t)
O I  O I  O\ I 
43

44 # (a #1)k(t)Q (a ,2, a #1, a !1,2, a )(t) (9)
G O I G G\ 
GI
 1360 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
2 and
3 I
4 QQ (a ,2, a )(t)"0 if a On. (10)
O I  G
5 G
6
7 2.4. Heuristic solution technique: equally likely combinations (ELC)
8
9 Motivated by the complexity of this system, we have devised a heuristic to reduce the number of
10 equations to be solved by combining each collection of states (l, m, r) into a single state (l, m). When
11 combining multiple states into one, we also modify the transitions among states. Two possible
12 transitions can occur when a customer completes a stage. If the stage completed was the last one
13 needed for a customer to exit the service facility, the customer leaves the system and the system
14 moves to a state with one less customer and one less stage to complete. The second type of
15 transition is when a customer completes one stage of service but the customer remains in service. In
16 this case, the system changes to a state with one stage less but with the same number of customers.
17 Let the rates at which these transitions occur be  k(t) and  k(t), respectively, where the
JK JK
18 subscript l, m indicates the state from which the transition originates.
19 The heuristic utilized to make the transformation from states (l, m, r) to state (l, m) is what
20 determines the transition probabilities  and  . In Escobar [17], two heuristics are studied:
JK JK
21 equally likely patterns in which all patterns of state (l, m, r) are assumed to be equally likely, and
22 equally likely combinations (ELC), in which all combinations of uncompleted stages in the servers
23 are assumed to be equally likely. As the performance of the latter heuristic has proven to be clearly
24 dominant (based on numerous computational experiments), we discuss only ELC here.
25 Let P be the array of the state probabilities, and let P be the state probability of state (l, m).
JK
26 Suppose the system is in state (l, m). Let D be the set of all pattern identi"ers that satisfy l and m, and
27 let d be the total number of patterns; let s be the number of servers with only one stage remaining
G
28 in pattern i; and let  be the transition probability to a state with l!1 stages and m!1
JK
29 customers. (In the example of Fig. 2, D"a, b, c, s "0, s "1 and s "2.)
? @ A
30 For each pattern in state (l, m), we can count the number of di!erent combinations of stages
31 remaining in the servers. The number of combinations in pattern i is given by
32
p!
33 C" , (11)
G p !p !2p !
34   V
35 where p"minm, n is the number of busy servers, x is the number of di!erent combinations of
36 stages in the servers, p denotes the number of servers with equal number of uncompleted stages,
37 H
and j"1, 2,2, x. The total number of combinations for a particular state is given by
38
39
C " C. (12)
40  G
GZD
41
(In the example of Fig. 2, in pattern (a), x"2 and C "5; in pattern (b), x"3 and C "30; and in
42 ? @
pattern (c), x"2 and C "10.)
43 A
44 The fundamental assumption in the ELC heuristic is that all the possible combinations of
uncompleted stages in the servers, C , are equally likely. Under this assumption, the transition

 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1361
1
2 probability to a state with one fewer customer is then given by
3 1
4  " s C, (13)
J K pC G G
5  GZD
6 and the transition probability from state (l, m) to state (l!1, m) is
7  "1! , (14)
8 JK JK
9 where both states have the same number of customers.
10 The procedure for deriving the transition probabilities  and  , is described in more detail
JK JK
11 through an example in Appendix B.
12
13 2.4.1. Total number of states
14 The reduced number of states associated with the ELC heuristic and their transitions can be
15 illustrated through a diagram. Fig. 3 shows the state-transition diagram for the general
16 M(t)/E (t)/n/n#q queue. In the top row of the "gure, we show the number of customers, m; the
I
17 number of uncompleted stages, l, is shown inside the ovals. In Appendix B, we provide state
18 transition diagrams using both the exact and heuristic solution techniques for the M(t)/E (t)/3/4

19 queue to show the great reduction in size and complexity in going from the former to the latter. For
20 ease of reading, the time dependence has been omitted in all cases.
21 Every customer that enters the system adds k stages to be completed before he/she leaves the
22 service facility (see Fig. 3). We need to keep track of the number of servers in use to determine the
23 service rate at which the system operates. We must separate the analysis into two sections: m)n,
24 the `growinga section; and m'n, the `queueinga section (Fig. 3). In the growing section the number
25 of states increases each time a new customer arrives at the system and enters one of the available
26 servers. When the (n#1)th customer arrives at the system, the queueing section starts. All servers (n
27 total) remain busy when n or more customers are present.
28 In the growing section, the number of states with exactly m customers in the system and m servers
29 occupied, is given by m(k!1)#1. The total number of states in the growing section (up to
30 n customers present) is
31
32 L n(n#1)
S " [m(k!1)#1]"(k!1) #n#1. (15)
33 % 2
K
34 Every space for a customer waiting in queue adds n(k!1)#1 states to the queueing section and to
35 the system.
36 If the system has unlimited queue space the system, of course, has an in"nite number of states.
37 With q(R, the number of states in the queueing section is
38
39 S "q[n(k!1)#1] (16)
/
40 and the total number of states in the system is represented by
41
42 S "S #S ,
1 % /
43
 
n(n#1)
44 S "(k!1) #qn #q#n#1. (17)
1 2
 1362 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25 Fig. 3. M(t)/E (t)/n queue.
I
26
27
28
29 Notice that the number of states is of the order maxkn, knq, generally much smaller than the
30 number of states in the exact solution. See Column 5 of Table 1 for speci"c examples.
31
32 2.4.2. State-to-state transitions
In the heuristic, we need to evaluate the state-to-state transition probabilities  and  for
33 JK JK
34 each state in the growing section. This section of states includes the column of states where the
35 number of customers is equal to the number of servers in the system. All columns of states but the
last in the queueing section have the same state-to-state transition probabilities  and 
36 JK JK
37 as the column that has exactly n customers and n servers occupied. This is because the
38 distribution of stages remaining in the queue (waiting to enter a server) does not a!ect the stage
39 distribution for customers in service. The only di!erence occurs in the (n#q)th column where
40 transitions to and from the right do not occur since no more than n#q customers are allowed in
41 the system. Table 3 summarizes the state-to-state transitions for the reduced number of states (ELC
42 heuristic).
43 Now, we can write the Chapman}Kolmogorov equations under the ELC model for both queues
M(t)/E (t)/n and M(t)/E (t)/n/n#q. Eqs. (18)}(24) represents the dynamics of the growing
44 I I
section. When the number of customers m is less than the number of servers in the system, m(n;
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1363
1
Table 3
2
State-to-state transitions for the heuristic solution techniques
3
4 From state To state With rate
5
6 (l, m) (l#k, m#1) (t)
7 (l!1, m!1)  mk(t)
JK
(l!1, m)  mk(t)
8 JK
9 (l!k, m!1) (l, m) (t)
(l#1, m#1)  mk(t)
10 JK
(l#1, m)  mk(t)
11 JK
12
13
14
15 we have
16 PQ (t)"!(t)P (t)#k(t)P (t), (18)
17   
PQ (t)"!((t)#mk(t))P (t)#(t)P (t)
18 KI K KI K KI\I K\
19 # (m#1)k(t)P (t) (19)
KI> K> KI> K>
20
21 for m"1, 2,2, (n!1),
22 PQ (t)"!((t)#mk(t))P (t)#(t)P (t)
KI\V K KI\V K KI\I\V K\
23
# mk(t)P (t)# (m#1)k(t)P (t)
24 KI\V> K KI\V> K KI\V> K> KI\V> K>
25 (20)
26 for m"1, 2,2,(n!1), x"1, 2,2,(m!1)(k!1),
27
28 PQ (t)"!((t)#mk(t))P (t)# mk(t)P (t)
KI\W K KI\W K KI\W> K KI\W> K
29 # (m#1)k(t)P (t) (21)
30 KI\W> K> KI\W> K>
31 for
32 m"1, 2,2, (n!1),
33
34 y"(m!1)(k!1)#1,(m!1)(k!1)#2,2, m(k!1),
35 and m is the number of customers in the system.
36 When the number of customers equals the number of servers in the system, m"n, the equations
37 are
38
PQ (t)"!((t)#nk(t))P (t)#(t)P (t)# nk(t)P (t), (22)
39 LI L LI L L\I L\ L\I> L LI> L>
40 PQ (t)"!((t)#nk(t))P (t)#(t)P (t)
41 LI\V L LI\V L L\I L\
# nk(t)P (t)# nk(t)P (t), (23)
42 LI\V> L LI\V> L L\I\V> L LI\V> L>
43 where x"1,2,2,(n!1)(k!1),
44
PQ (t)"!((t)#nk(t))P (t)# nk(t)P (t), (24)
LI\W L LI\W L LI\W> L LI\W> L
 1364 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
2 where
3 y"(n!1)(k!1)#1, (n!1)(k!1)#2,2, n(k!1).
4
5 Eqs. (25)}(27) capture the dynamics of the queueing section when the queue size is unlimited,
6 q"R. The equations are given by
7
PQ (t)"!((t)#nk(t))P (t)#(t)P (t)
8 KI K KI K KI\I K\
9 # nk(t)P (t), (25)
L\I> L KI> K>
10
11 for m"(n#1), (n#2),2, (n#q!1),
12 PQ (t)"!((t)#nk(t))P (t)#(t)P (t)
13 KI\V K KI\V K KI\I\V K\
14 # nk(t)P (t)# nk(t)P (t) (26)
LI\V> L KI\V> K L\I\V> L KI\V> K>
15 for m"(n#1), (n#2),2, (n#q!1), x"1,2,2, (n!1)(k!1),
16
17 PQ (t)"!((t)#nk(t))P (t)#(t)P (t)
KI\W K KI\W K KI\I\W K\
18
# nk(t)P (t) (27)
19 LI\W> L KI\W> K
20 for
21
22 m"(n#1), (n#2),2, (n#q!1),
23 y"(n!1)(k!1)#1, (n!1)(k!1)#2,2, n(k!1).
24
25 In the case with limited size queue, q(R, Eqs. (28), (29) are needed. The equations for the states
26 when there are n#q customers in the system are
27
PQ (t)"!nk(t)P (t)#(t)P (t), (28)
28 L>OI L>O L>OI L>O L>O\I L>O\
29 PQ (t)"!nk(t)P (t)#(t)P (t)
L>OI\V L>O L>OI\V L>O L>O\I\V L>O\
30
# nk(t)P (t), (29)
31 LI\V> L L>OI\V> L>O
32 where x"1, 2,2, n(k!1).
33 And "nally,
34
35 PQ (t)"0 (30)
JK
36
for values (l, m) not included in Eqs. (18)}(29).
37
Appendix B illustrates graphically details of applying both the exact and heuristic solution
38
techniques to the M(t)/E (t)/3/4 queueing system. Included are "gures and explanations of all
39 
possible transitions and computations of associates 's and 's.
40
41
42 2.5. Performance measures of interest
43
44 After solving the Chapman}Kolmogorov equations and obtaining the state transition probabil-
ities for all states (l, m), the results can be processed to calculate the probability distribution for the
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1365
1
2 number of customers m in the system as follows:
3


IK p (t), 0)m)n,
4 P (t)" JK J K (31)
5 K IK p (t), n(m)q.
JL>K\LI J K
6
7 These occupancy probabilities may be used to compute many other measures of interest. Note that
8 for "nite capacity systems the queue size is limited and, therefore, the e!ective arrival rate to the
9 system when using Little's formulae is given by
10
 "(1!P ).
11 L>O
12
13
14 3. Validation of the heuristic
15
16 The validation e!ort concentrated primarily on comparing steady-state numerical results ob-
17 tained through the exact solution technique with those obtained through the equally likely
18 combinations (ELC) heuristic, under stationary conditions. It will be seen that the ELC heuristic
19 provides excellent approximations, for most practical purposes, to the exact results. As ELC
20 typically requires the solution of far fewer equations than the exact approach, results are obtained
21 much faster, making it possible to solve (approximately) much larger models than those currently
22 solvable through the exact solution technique. The performance of ELC was also tested, to a lesser
23 extent, under transient conditions with stationary arrival and service rates, as well as under
24 dynamic demand conditions. Results suggest good performance in these cases, as well. In Section
25 3.1 we summarize validation results under stationary conditions and in Section 3.2 under transient
26 and dynamic conditions.
27
28
29 3.1. Stationary conditions
30
31 In order to obtain a numerical solution we must, of course, limit the number of customers
that may join the queue. This means that in all cases we solve the M/E /n/n#q queue,
32 I
33 q(R, where q is the maximum number of customers waiting for service. In practice, if q is
su$ciently large and
(1, the M/E /n/n#q queue will provide an acceptable approximation to
34 I
the M/E /n queue. Included in our experiments are many examples with e!ectively in"nite
35 I
36 capacity.
37 A total of 170 di!erent cases were examined, selected to cover a wide range of system parameters.
38 We included large and small systems in terms of both the number of servers, n, and the
39 length of queue, q; large and small Erlang orders, k, to provide multiple distribution shapes
40 for the service times; and a range of utilization ratios,
, including under- and over-saturated
41 systems.
42 The tables in this section present results from applying the ELC heuristic in a representative
43 sample of the 170 cases. These results are for steady-state behavior. We assume a system has
44 e!ectively achieved steady state when the state probabilities are constant, up to 6 decimal places,
for at least 5 continuous output intervals. All systems start empty and idle.
 1366 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
2 Table 4 illustrates the accuracy of the heuristic solution technique in evaluating =, the expected
time in the system, and ¸ , the expected queue length, for 39 cases where
3 O
4  mP
5 =" K K ,

6
7 
¸ " (m!n)P .
8 O K
KL
9
See Escobar [17] for results with other performance measures including the expected delay,
10
expected number of customers in the system, and occupancy probabilities.
11
Using ELC, the expected queue length, ¸ , was within 3% of the exact solution in all 170 cases
12 O
and within 1% in 95% of the cases. The expected system time, =, obtained from the heuristic was
13
always within 1% of the exact solution.
14
For both aggregate measures, the results obtained using the heuristic are sometimes larger and
15
sometimes smaller than the exact solutions. Therefore, ELC does not provide either an upper or
16
lower bound for the exact results.
17
To illustrate the advantage of using the heuristic technique, we compare in Table 5 the number of
18
Chapman}Kolmogorov equations and the associated CPU times required to solve these equations
19
for each of the 39 cases in Table 4. The "rst three columns specify all combinations of n, q and
for
20
each value of k. Columns 4 and 5 specify the number of states in each model for the exact and
21
heuristic techniques, respectively. The CPU times required to obtain steady-state solutions with
22

"0.9 for the exact solution technique is in column 6, and the CPU time required using ELC is in
23
column 7. The "nal column shows the ratios of columns 6 and 7. We selected the utilization ratio of
24
0.9 as representative of a `typicala application.
25
The CPU times for running the experiments using the exact solution technique range from 0.42
26
up to 192,475 s (53.46 h) on a SUN SPARCstation 10 Model 41; note that the time for solving the
27
largest system studied was not available because we were unable to solve the extremely large
28
number of equations with the available computer hardware and software. Using the heuristic
29
solution technique, solution times ranged from 0.25 up to 1183.73 s (19.73 min). As expected, the
30
solution time depends on both the number of Chapman}Kolmogorov equations to solve and on
31
how long the model takes to reach steady state. The range of the CPU time ratios is from 1.68 times
32
faster in the smallest system up to 2646.07 times for the second largest system considered in the
33
analysis.
34
These CPU time ratios illustrate a clear advantage of using the heuristic solution techniques over
35
the exact solution technique. Speed in solving the desired systems is critical when the technique is
36
used for solving numerous queueing systems or a network of queues.
37
38
39 3.2. Transient and nonstationary results
40
41 In many applications, steady-state results are of limited use due to the presence of strongly
42 dynamic conditions and/or the importance of the transient period. Our numerical experiments
43 suggest that the ELC heuristic works well in both of these cases (see Escobar [17] for details).
44 We examined the transient expected number of customers in the system in twelve cases in which
systems started empty and idle. The cases are identi"ed in Table 6 which also indicates the worst
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1367
1
Table 4
2
Expected waiting time (=) and expected queue length (¸ )*exact vs. heuristic
3 
4 k n q
System time = Queue length ¸ % di!erence

5
6 Exact ELC Exact ELC = ¸

7
3 3 10 0.5 11.1125 11.1082 0.1669 0.1662 0.04 0.38
8 3 3 10 0.9 20.3851 20.3835 2.7444 2.7441 0.01 0.01
9 3 3 10 0.99 24.8242 24.8302 4.1709 4.1728 !0.02 !0.05
10 3 3 10 1.2 33.3901 33.4109 6.9587 6.9651 !0.06 !0.09
11 3 15 30 0.5 10.0120 10.0119 0.0090 0.0089 0.00 1.00
12 3 15 30 0.9 12.6157 12.6142 3.5290 3.5269 0.01 0.06
13 3 15 30 0.99 18.2410 18.2458 12.0455 12.0526 !0.03 !0.06
14 3 15 30 1.2 27.4561 27.4704 26.1836 26.2050 !0.05 !0.08
15 4 3 10 0.5 11.0545 11.0474 0.1582 0.1571 0.06 0.68
16 4 3 10 0.9 20.1745 20.1716 2.6948 2.6942 0.01 0.02
17 4 3 10 0.99 24.7866 24.7971 4.1728 4.1762 !0.04 !0.08
4 3 10 1.2 33.6587 33.6955 7.0486 7.0599 !0.11 !0.16
18
19 4 50 25 0.9 N/A 10.4410 N/A 1.9829 N/A N/A
20 5 3 10 0.5 11.0195 11.0105 0.1529 0.1516 0.08 0.88
21 5 3 10 0.9 20.0368 20.0329 2.6619 2.6611 0.02 0.03
22 5 3 10 0.99 24.7619 24.7757 4.1735 4.1779 !0.06 !0.11
5 3 10 1.2 33.8318 33.8802 7.1059 7.1207 !0.14 !0.21
23
24 5 3 100 0.5 11.0198 11.0108 0.1530 0.1516 0.08 0.88
25 5 3 100 0.9 26.5074 26.4846 4.4570 4.4508 0.09 0.14
5 3 100 0.99 129.4921 129.0503 35.4102 35.2803 0.34 0.37
26 5 3 100 1.2 332.4972 332.5524 96.7492 96.7657 !0.02 !0.02
27
8 3 10 0.5 10.9666 10.9548 0.1450 0.1432 0.11 1.21
28
8 3 10 0.9 19.8114 19.8061 2.6073 2.6062 0.03 0.04
29 8 3 10 1.2 34.1096 34.1784 7.1971 7.2180 !0.20 !0.29
30
10 3 10 0.5 10.9487 10.9362 0.1423 0.1404 0.11 1.33
31 10 3 10 0.9 19.7305 19.7250 2.5875 2.5864 0.03 0.04
32 10 3 10 1.2 34.2075 34.2839 7.2289 7.2521 !0.22 !0.32
33
10 10 5 0.5 10.0480 10.0467 0.0240 0.0233 0.01 2.71
34 10 10 5 0.9 11.1715 11.1886 1.0101 1.0259 !0.15 !1.56
35 10 10 5 1.2 12.6743 12.7213 2.6097 2.6577 !0.37 !1.84
36 15 3 10 0.5 10.9246 10.9111 0.1387 0.1367 0.12 1.46
37 15 3 10 0.9 19.6174 19.6121 2.5597 2.5586 0.03 0.04
38 15 3 10 1.2 34.3425 34.4295 7.2726 7.2990 !0.25 !0.36
39 15 5 10 0.5 10.3227 10.3162 0.0807 0.0791 0.06 2.01
40 15 5 10 0.9 15.3311 15.3362 2.3662 2.3688 !0.03 !0.11
41 15 5 10 1.2 24.3858 24.4666 7.1646 7.2053 !0.33 !0.57
42 20 3 10 0.5 10.9123 10.8986 0.1369 0.1348 0.13 1.51
43 20 3 10 0.9 19.5583 19.5535 2.5451 2.5441 0.02 0.04
44 20 3 10 1.2 34.4122 34.5045 7.2951 7.3231 !0.27 !0.38
 1368 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
Table 5
2
Summary of number of states and CPU times
3
4 Servers Queue Utilization Number of states CPU times (s),
"0.9
5
6 n q
Exact Heuristic Exact Heuristic Ratio
7 (max)
8 Erlang order (k): 3
9 3 10 0.5,0.9,0.99,1.2 120 86 10.16 1.71 5.94
10 15 30 0.5,0.9,0.99,1.2 4896 1186 1155.5 122.76 9.41
11 Erlang order (k): 4
12 3 10 0.5,0.9,0.99,1.2 235 122 15.49 5.12 3.03
13 50 25 0.9 901901 7651 N/A 1183.73 N/A
14 Erlang order (k): 5
15 3 10 0.5,0.9,0.99,1.2 406 158 38.54 4.68 8.24
16 3 100 0.5,0.9,0.99,1.2 3556 1328 2153.3 207.43 10.38
17 Erlang order (k): 8
18 3 10 0.5, 0.9, 1.2 1365 266 266.62 14.64 18.21
19 Erlang order (k): 10
20 3 10 0.5, 0.9, 1.2 2486 338 692.66 22.76 30.43
21 10 5 0.5, 0.9, 1.2 646646 961 192475.05 72.74 2646.07
22 Erlang order (k): 15
23 3 10 0.5, 0.9, 1.2 7616 518 4121.3 49.73 82.87
24 5 10 0.5, 0.9, 1.2 131784 926 78712.5 78.33 1004.88
25 Erlang order (k): 20
26 3 10 0.5, 0.9, 1.2 17171 698 25712.4 88.33 291.09
27
28
29
30 case performance observed in each case. Both under- and over-saturated models are included. To
31 better appreciate the quality of the approximation provided by ELC, we plot in Fig. 4 the expected
32 number of customers as a function of time for one of these systems.
To observe performance under dynamic conditions, we solved instances of M(t)/E (t)/5/5#15
33 
and M(t)/E (t)/18/18#3 models with a sinusoidal arrival rate, (t), along the experimental lines
34 
35 suggested by Green et al. [19].
For example, for the tests with the M(t)/E (t)/18/18#3 system, we used a service rate "0.1
36 
37 customers per minute per server (combined rate of 1.8 customers per minute when all servers are
38 busy) and an arrival rate of (t)"sin(2 t/30)#1.16 customers per minute, giving maximum and
39 minimum values of
of 1.2 and 0.888, respectively, during each 30-min long cycle of the sinusoid.
40 The results for this system are shown in Fig. 5 whose horizontal axis is in minutes. The ELC
41 estimates track the exact solution very well. Even in the magni"ed view of the initial period of
42 operation (see Fig. 6), there are only slight di!erences, for practical purposes, between the two sets
43 of results. The maximum di!erence for the expected number of customers in the system was 23.8%
44 at minute 61, when the absolute di!erence was negligible, less than 0.001 minutes. Most of the time
the two estimates were within 2% of each other.
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1369
1
Table 6
2
Transient examples and maximum percentage di!erence
3
4 System Utilization ratio % Di!erence E[Cust.] Actual di!erence
5
E[Cust.]
6
7 M/E /5/5#5 0.5 0.36 0.0065

0.9 0.48 0.0147
8 1.2 0.54 0.0220
9
M/E /18/18#3 0.5 0.54 0.0332
10 
0.9 0.58 0.0575
11 1.2 0.57 0.0760
12
M/E /15/15#30 0.5 1.07 0.0561
13 
0.9 1.16 0.1101
14
M/E /3/3#100 0.9 1.08 0.0241
15 
16 M/E /3/3#10 0.9 2.36 0.0559

17 1.2 2.58 0.0818
18 M/E /3/3#10

0.9 3.81 0.0939
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
Fig. 4. Transient results for M/E /18/18#3 queueing system.

 1370 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
2
3
4
5
6
7
8
9
10
11
12
Fig. 5. M(t)/E (t)/18/18#3 system response with sinusoidal .
13 
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28 Fig. 6. M(t)/E (t)/18/18#3 system response (zoom) with sinusoidal .
29 
30
31
32 The results reported here are typical of the observations made in all the other numerical
33 experiments of this type (see Escobar [17]). While this part of the model validation can be
34 characterized as only preliminary, it strongly suggests that ELC provides good approximations for
35 steady-state statistics as well as under dynamic and/or during transient periods.
36
37
38 4. Conclusions and future work
39
40 The main contribution of this paper is a heuristic approximation scheme, the equally likely
combinations (ELC) model, for solving the M(t)/E (t)/n/n#q queueing system. Experimental
41 I
42 results indicate that ELC: provides excellent, for most practical purposes, approximations to the
43 typical quantities of interest; increases considerably the size of systems that can be solved with
44 reasonable computational e!ort, as a result of the greatly reduced number of states required by the
approximate model; and speeds up the solution of some large systems by a factor of 1000 or more.
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1371
1
2
3
4
5
6
7
8
9
10
11
12 Fig. 7. The n;k matrix of circles of a pattern when m(n.
13
14 These experimental results cover a wide range of system parameters and utilization and appear
15 robust to parameter changes.
The practical signi"cance of this new heuristic extends beyond just M(t)/E (t)/n/n#q systems. In
16 I
the stationary case, M/E /n/n#q systems are often used to obtain estimates of the performance of
17 I
18 some M/G/n/n#q systems. Typically, this works as follows: assume that the service times, S, at the
19 original M/G/n/n#q system have expected value E[S] and variance <ar[S]. The original system is
then approximated through a M/E /n/n#q model by (1) setting the expected service times in the
20 I
latter system equal to E[S] and (2) selecting k so that the variance of the Erlang random variable E
21 I
22 (whose expected value has been set to E[S]) is as close as possible (remember that k must be
a positive integer) to <ar[S]. By extending signi"cantly the range of M/E /n/n#q systems that can
23 I
24 be solved, ELC similarly extends the range of M/G/n/n#q systems that can be approximated and
25 the ease with which this can be done.
26 In future work, additional research is necessary on several fronts, including: testing further the
27 performance of ELC under steady state and, especially, dynamic conditions; testing the quality of
28 the approximations to various types of M/G/n/n#q systems obtained through ELC; and obtain-
29 ing better insights, through ELC-computed solutions, into the qualitative behavior of
M(t)/E (t)/n/n#q systems under stationary and dynamic conditions for various parameter settings.
30 I
31
32
33 Appendix A. Pattern generator algorithm
34
35 This algorithm generates all patterns associated with speci"c values of l and m in Description 2.
36 It works by assigning uncompleted stages to servers in a left justixed, top justixed manner, in the
37 sense shown in Fig. 7, such that each busy server has at least one uncompleted stage.
38 To understand how the algorithm works, refer to Fig. 7. Each pattern is a matrix of circles, with
39 n rows and k columns, where the uncompleted stages are denoted by shaded circles as in Fig. 2 and
40 7. If there are fewer than n customers in the systems, m(n (thus, m busy servers), then the bottom
41 n!m rows contain only empty circles (the servers are idle). The rest of the servers will be assigned
42 stages (the circles will be "lled) according to this algorithm:
䢇 Input data: l, m, n, k
43
䢇 IF m(n
44
䢇 THEN last n!m rows are not used (idle servers)
 1372 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
2 䢇 IF m'n
3 䢇 THEN draw pattern for l"l!k(m!n) uncompleted stages, FLAG ON
4 䢇 stagesleft "l, column"k
5 䢇 kth COLUMN: minimum and maximum stages this pass
6 䢇 smin[column] " minimum number of stages in column k
7 䢇 smax[column] " maximum number of stages in column k
8 䢇 stagesleft " stagesleft ! smin[column]
9 䢇 FOR stages[column] " smin[column] TO smax[column]
10
column " column ! 1
11
(k!1)st COLUMN: minimum and maximum stages this pass
12
smin[column]" minimum number of stages in column k!1
13
smax[column] " maximum number of stages in column k!1
14
stagesleft " stagesleft ! smin[column]
15
FOR stages[column]"smin[column] TO smax[column]
16
column " column ! 1
17
(k!2)th TO 2nd COLUMNS: Nested loops as (k!1)th COLUMN loop
夹 1st COLUMN: No loop here since all stages left should be the same as the number of
18
19 busy servers
夹 stages[column]" stagesleft
20
夹 PRINT PATTERN LOOP
21
夹 column " column#1
22
23
stagesleft " stagesleft # smax[column]#1
24
column "column#1
25
stagesleft " stagesleft #smax[column]#1
26
column "column#1
27 䢇 stagesleft " stagesleft #smax[column]#1
28 䢇 IF m'n
29 䢇 THEN identify pattern obtained with original stage (l, m), FLAG OFF
30 䢇 END
31
32 The PRINT PATTERN LOOP has the following steps:
33 䢇 FOR row " 1 TO minm, n
34
FOR col " 1 TO k
35
IF stages[col]* row
36
THEN use location (row,col)
37
ELSE do not use location (row,col)
38 䢇 END
39
In the "rst loop of the algorithm, when column"k, the minimum and maximum number of
40
uncompleted stages is given by
41
42 smin[column]"max0, stagesleft!(column!1)p
43
 
stagesleft!p
44 smax[column]"min p, ,
column!1
 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374 1373
1
2 where p"minm, n indicates the number of busy servers. For the remaining k!2 loops in the
3 algorithm, for 2)column)k!1, the minimum and maximum values are
4 smin[column]"maxstages[column#1], stagesleft!(column!1)p
5

 
6 stagesleft!p
7 smax[column]"min p, .
column!1
8
9
10 Appendix B. States and state transitions in the M(t)/E (t)/3/4 queueing system
11 
12 Please see http://web.mit.edu/orc/www/faculty/odoni.html for this appendix.
13
14
15
16 References
17
[1] Shapiro S. The M-server queue with Poisson input and gamma-distributed service of order two. Journal of the
18 Operations Research Society of America 1966;14:685}94.
19 [2] Mayhugh J, McCormick R. Steady state solution of the queue M/E /r. Management Science 1968;14:692}712.
20 I
[3] Murray J, Kelton W. The transient behavior of the M/E /2 queue in steady state simulation. Computers and
I
21 Operations Research 1988;15(4):357}67.
22 [4] Maal~e E. Approximation formulae for estimation of waiting-time in multiple-channel queueing systems. Manage-
ment Science 1973;19:703}10.
23
[5] Smith V. Distribution of customers in M/E /m queues using Hokstad's approximation. Applied Statistics (Royal
24 2
Statistics Society) 1987;36:394}401.
25 [6] Hokstad P. Approximations for the M/G/m queue. Operations Research 1978;26(3):510}23.
26 [7] Cosmetatos G. Some approximate equilibrium results for the multi-server queue (M/G/r).. Operational Research
27 Quarterly 1976;27:615}20.
28 [8] Takahashi Y. An approximation formula for the mean waiting time of an M/G/c queue. Journal of the Operations
Research Society of Japan 1977;20(3):150}63.
29 [9] Nozaki S, Ross S. Approximations in "nite-capacity multi-server queues with Poisson arrivals. Journal of Applied
30 Probability 1978;15:826}34.
31 [10] Boxma O, Cohen J, Hu!els N. Approximations of the mean waiting time in an M/G/s queueing system. Operations
32 Research 1979;27:1115}27.
33 [11] Tijms H, van Hoorn M, Federgruen A. Approximations for the steady-state probabilities in the M/G/c queue.
Advance Applied Probability 1981;13:186}206.
34
[12] Yao D. Re"ning the di!usion approximation for the M/G/m queue. Operations Research 1985;33:1266}77.
35 [13] Miyazawa M. Approximation of queue-length distribution of an M/GI/s queue by the basic equations. Journal of
36 Applied Probability 1986;23:443}58.
37 [14] Ma B, Mark J. Approximation of the mean queue length of an M/G/c queueing system. Operations Research
38 1995;43(1):158}65.
39 [15] Kimura T. Di!usion approximation for an M/G/s queue. Operations Research 1983;31(2):304}19.
[16] Kimura T. A transform-free approximation for the "nite capacity M/G/s queue. Operations Research
40 1996;44(6):984}8.
41 [17] Escobar M. Approximate solutions for multi-server queueing systems with Erlangian service times and applications
42 to air tra$c management. PhD thesis, 1998 Massachusetts Institute of Technology.
43 [18] Lee D. M/E /N/N#Q and M/E /N/N#Q. LMI Internal Paper, January 1997.
 )
44 [19] Green L, Kolesar P, Svoronos A. Some e!ects of nonstationarity on multi-server Markovian systems. Operations
Research 1991;39(3):502}11.
 1374 M. Escobar et al. / Computers & Operations Research 29 (2002) 1353}1374
1
Marcos Escobar is the Director of Strategic Planning at Grupo Posadas, hospitality group in Latin America, Mexico.
2
He obtained his Ph.D. in Electrical Engineering at the Massachusetts Institute of Technology. He occasionally teaches
3 courses and seminars in operations research at various universities and institutes in Mexico.
4 Amedeo R. Odoni is the T. Wilson Professor of Aeronautics and Astronautics at the Massachusetts Institute of
5 Technology and the former Co-Director of the Operations Research Center there. He is the author or co-author of books
on Urban Operations Research and on Airport Planning and Design and of more than 50 journal publications. His
current research interests are in applied stochastic processes and in the applications of operations research to airports
and air tra$c control.
Emily Roth is an associate professor in the Mathematical Sciences Department at Bentley College. She received her
S.B. in Mathematics and Ph.D. in Operations Research from the Massachusetts Institute of Technology. She is the
author or co-author of several articles on heuristics for evaluating the behavior of queueing models as well as application
of stochastic processes to business problems. Currently, she is interested in the application of operations research
techniques to problems in air tra$c control.