Analysis of Aircraft Sortie Generation With The Use of A Fork-Join Queueing Network Mode

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Analysis of Aircraft Sortie Generation with the Use of a

Fork-Join Queueing Network Model


Dennis C. Dietz, 1 Richard C. Jenkins 2
1
Department of Operational Sciences, Air Force Institute of Technology,
Wright-Patterson Air Force Base, Ohio
2
Armstrong Laboratory, Air Force Materiel Command, Brooks Air Force Base, Texas

Received August 1995; revised March 1996; accepted 19 September 1996

Abstract: This article presents an approximate analytical method for evaluating an aircraft
sortie generation process. The process is modeled as a closed network of multiserver queues
and fork-join nodes that allow concurrent service activities. The model uses a variation of
mean value analysis (MVA) to capture the effect of mean service times, resource levels,
and network topology on performance measures including resource utilizations and the
overall sortie generation rate. The quality of the analytical approximation is demonstrated
through comparison with simulation results. q 1997 John Wiley & Sons, Inc.* Naval Research
Logistics 44: 153–164, 1997

1. INTRODUCTION
Central to many theater-level analyses involving military aircraft is a determination of
the frequency with which the aircraft can be employed in combat. Often referred to as
operational readiness, this factor can be measured by numerical values such as aircraft
availability (portion of time aircraft are mission capable) or sortie generation rate (sustain-
able number of aircraft launches per time period). These values can be difficult to determine
because they depend on both the inherent reliability/maintainability characteristics of the
aircraft and the resource constraints within the supporting logistics system [4]. In order to
gain insight into operational readiness issues, analysts have historically employed simulation
tools such as the logistics composite model (LCOM) [6, 8], the sortie generation model
(SGM) [1], theater simulation of airbase resources (TSAR) [7], and Dyna-Sim [12].
Unfortunately, comprehensive evaluation of operational concepts by using simulation can
be very tedious. Because of random variation in simulation output, multiple replications
and careful experimental design are required to place acceptable confidence bounds on
results. Analytical queueing network approaches have been considered as alternatives to
simulation, but no method has yet been been offered to address the problem of concurrent
(fork-join) repair of different aircraft subsystems. This article presents an analytical ap-

Correspondence to: D.C. Dietz

q 1997 John Wiley & Sons, Inc. *This article is a US Government work and, as such, is in the
public domain in the United States of America. CCC 0894-069X/97/020153-12

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


154 Naval Research Logistics, Vol. 44 (1997)

proach that produces excellent approximate results for a sortie generation system modeled
as a closed network of multiserver queues and fork-join nodes. The approach is based on
recent work by Rao and Suri on analysis of Kanban-type manufacturing systems [13]. The
quality of the analytical approximation is demonstrated through comparison with simulation
results.

2. MODEL DESCRIPTION

Although a sortie generation system can be quite complex, a simple model can capture
the fundamental process behavior. To demonstrate this concept, a fixed complement of
aircraft is assumed to operate from a single base in a closed-loop system. The aircraft flow
through a network of activities which include taxi, sortie, (flight), turnaround (refueling
and scheduled maintenance), munitions upload, and repair (unscheduled maintenance).
Because activity capacities may be constrained, the entire system can be represented by a
set of N statistically identical aircraft that circulate between M indexed stations in a closed
queueing network. Upon completion of the service activity at any station i, each aircraft
proceeds to station j with time-invariant probability pij . Figure 1 illustrates the movement
of aircraft between stations in a typical sortie generation cycle.
The employment of each aircraft begins with ground operations (taxi), during which
numerous preflight checks are accomplished. With probability p13 , at least one malfunction
is detected and the aircraft experiences a ground abort. A troubleshooting activity is per-
formed and the aircraft receives appropriate repairs. Because each type of repair requires
a particular set of scarce resources, queueing delay may be encountered. Any aircraft not
experiencing a ground abort flies a sortie and returns to the base. With probability p23 , the
aircraft lands with an indication of at least one type of malfunction requiring repair. If no
malfunction is indicated, the aircraft immediately awaits turnaround operations. The turn-
around queue contains all aircraft which have completed repairs or landed without malfunc-
tion indications. After turnaround, each aircraft awaits munitions upload. When the upload
is complete, the aircraft is ready to taxi, and the cycle is repeated.
During the troubleshooting activity, an aircraft may be found to require up to five types
of repair. Multiple types of repair can be performed concurrently if the required resources
are available. Thus, the aircraft can be viewed as generating temporary clones that are
rejoined into a single entity when all repairs are complete. This process occurs at the fork-
join node represented by Station 4 in Figure 1. The different types of repair are required
with independent probabilities q4k , k Å 1, . . . , 5. A false malfunction indication thus occurs
with probability ∏ 5kÅ1 (1 0 q4k ), in which case the affected aircraft bypasses repair and
proceeds directly to the turnaround station.
Table 1 displays repair-probability values that are representative of current fighter aircraft.
The table also displays mean durations (in hours) and resource levels for each activity in the
network. The resource levels reflect the maximum number of aircraft that can simultaneously
receive each type of service. This article is concerned with aggregating these known service
parameters and the network structure to produce long-run (steady-state) mean performance
measures such as the sortie generation rate and resource utilizations. For purposes of analyti-
cal tractability, all service times are approximated by exponentially distributed random
variables. Fortunately, performance measures for closed queueing networks are often quite
insensitive to variances and higher-order moments of service times [14]. This property is
particularly evident when modeling aircraft maintenance systems [12].

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


Dietz and Jenkins: Aircraft Sortie Generation 155

Figure 1. Aircraft sortie generation process.

3. MEAN VALUE ANALYSIS

If the concurrent repair activities were removed from the example network, performance
measures could be computed with the use of mean value analysis (MVA). In general, MVA
can be applied to a closed or capacitated network with N customers (aircraft) circulating
through M stations having symmetric service disciplines (FCFS exponential, processor
sharing, infinite server, or LCFS preemptive) [2]. Each station may consist of a single
server, multiple identical servers, or an infinite number of servers (e.g., the sortie activity).
The service rate when n customers are at station i is given by mi (n) Å min(n/si , ri /si ).
Performance measures obtainable through MVA include mean response time Ri (waiting
and service time), throughput li , queue length Qi (number of customers waiting or in
service), and server utilization Ui (expected number of busy servers). Excellent expositions
of the MVA algorithm are given in Bruell and Balbo [3], or Conway and Georganas [5].
A critical foundation for MVA is the arrival theorem, proven in Lavenberg and Reiser
[11]. This theorem states that, for a closed network with N customers, an arriving customer
observes the same distribution of customers at a station as the stationary (random observer’s)

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


156 Naval Research Logistics, Vol. 44 (1997)

Table 1. Sortie generation model parameters.


Repair Resource
Activity Service time probability level
Taxi s1 Å 0.25 r1 Å `
Sortie s2 Å 2.00 r2 Å `
Troubleshoot s3 Å 0.50 r3 Å `
Rpr 1 (airframe) s41 Å 2.20 q41 Å 0.17 r41 Å 1
Rpr 2 (electrical/hydraulic) s42 Å 2.27 q42 Å 0.39 r42 Å 3
Rpr 3 (engine) s43 Å 2.37 q43 Å 0.21 r43 Å 2
Rpr 4 (avionics) s44 Å 1.50 q44 Å 0.27 r44 Å 1
Rpr 5 (radar/weapons control) s45 Å 1.19 q45 Å 0.46 r45 Å 2
Turnaround s5 Å 0.75 r5 Å 6
Munitions upload s6 Å 0.50 r6 Å 4

distribution for the same network with N 0 1 customers. The arrival theorem leads to the
marginal local balance theorem, which states that

mi (n)Pi (nÉN) Å li (N)Pi (n 0 1ÉN 0 1), (1)

where Pi (nÉN) is the probability that n customers are at station i given N customers are
in the network, and li (N) is the customer throughput for station i when N customers are in
the network. The marginal local balance theorem can be applied to recursively compute
performance measures. First, note that the mean queue length at any station i can be
written as

N N
nli (N)
Qi (N) Å ∑ nPi (nÉN) Å ∑ Pi (n 0 1ÉN 0 1). (2)
nÅ1 nÅ1 mi (n)

The throughput li (N) in Eq. (2) is unknown, but application of Little’s law yields

N
n
Ri (N) Å ∑ Pi (n 0 1ÉN 0 1). (3)
nÅ1 mi (n)

Note that if station i has only one server, then


N
Ri (N) Å si ∑ nPi (n 0 1ÉN 0 1) Å si {1 / Qi (N 0 1)}, (4)
nÅ1

and if it has an infinite number of servers, then Ri (N) Å si for all N.


Equations (3) and (4) relate the response time of a station when N customers are present
in the network to the distribution of customers at the station when N 0 1 customers are
present. Thus, by starting at N Å 1 [where Pi (0ÉN 0 1) Å 1 and Qi (N 0 1) Å 0 for all
i], performance measures can be obtained recursively. To determine station throughputs at
each iteration, it is necessary to determine the average cycle time for a customer at an
arbitrary reference station (e.g., station 1). The average time between two successive depar-
tures by the same customer from station 1 is given by

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


Dietz and Jenkins: Aircraft Sortie Generation 157

M
£i Ri (N)
CT 1 (N) Å ∑ , (5)
iÅ1
£1

where each ratio £i / £1 is the mean number of visits a customer makes to station i for every
visit to Station 1. These ratios can be obtained by setting £1 equal to an arbitrary value
(e.g., £1 Å 1) and solving the system of equations P Å vt , where P is the matrix of routing
probabilities for the network stations. For the example model with historical ground-abort
probability p13 Å 0.05 and postsortie troubleshooting probability p23 Å 0.30, the routing
matrix is

0 0.95 0.05 0 0 0
0 0 0.30 0 0.70 0
0 0 0 1 0 0
PÅ . (6)
0 0 0 0 1 0
0 0 0 0 0 1
1 0 0 0 0 0

If we use the cycle time from Eq. (5), throughput at each station can be determined as

N£ i
li (N) Å . (7)
CT 1 (N) £1

By applying Little’s law, queue length and utilization for each station can then be computed
as

Qi (N) Å Ri (N) li (N), (8)

Ui (N) Å si li (N). (9)

If station i has a single server, then the result from Eq. (8) can be used in Eq. (4) to obtain
response times with one more customer in the network. If station i has multiple servers,
then the marginal local balance theorem can be applied to determine the new distribution
of customers as

li (n)Pi (n 0 1ÉN 0 1)
Pi (nÉN) Å , n ú 0, (10)
mi (n)

N
Pi (0ÉN) Å 1 0 ∑ Pi (nÉN). (11)
nÅ1

These probabilities can then be used in Eq. 3 to obtain the multiple server response times
for the next iteration. The process is repeated until N is equal to the desired number of
customers in the network.

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


158 Naval Research Logistics, Vol. 44 (1997)

4. FORK-JOIN NODES
With the addition of concurrent repair activities, the product-form nature of the network
is destroyed and the MVA algorithm can no longer be directly applied [5]. However, a
heuristic based on MVA was recently developed by Rao and Suri to analyze a single fork-
join system of single-server queues [13]. This article extends the heuristic concepts to
accomodate multiple fork-join nodes in a larger network, multiple server activities, and
probabilistic service requirements.
Consider the case of a fork-join node i containing Ki substations where qik Å 1 for all k
Å 1, . . . , Ki . A customer arriving at station i immediately generates Ki clones, which enter
separate queues for the substations and are rejoined to the parent customer after their
respective service completions. The parent customer can be viewed as holding position at
fork-join node i until all of its clones have completed service, when it then moves to a new
network station according to known routing probabilities. Mean response time, queue length,
and utilization can be estimated for clone traffic at each substation ik and are denoted
Rik (N), Qik (N), and Uik (N), respectively. Following Rao and Suri, two key approximations
lead to the evaluation of network performance measures:

Á Approximation 1. For a network with N customers, a clone arriving at a


substation sees the stationary (random observer’s) distribution of clones at the
substation for the same network with N 0 1 customers.
Á Approximation 2. The response time experienced by a clone at a substation
can be represented as an exponentially distributed random variable and is inde-
pendent of the response time for clones at other substations.

Based on Approximation 2, the response time for any substation ik can be denoted by the
exponential random variable Tik (N) with rate parameter uik (N) Å 1/Rik (N). This response
time includes substation service time and any waiting time between clone creation and
service initiation. The mean time that a parent customer holds at fork-join node i, which
must be determined to obtain network cycle time, is therefore E[maxkÅ1,. . . ,Ki {Tik (N)}].
The assumption that qik Å 1 for all k Å 1, . . . , Ki can be relaxed by simply conditioning
on the subset of service activities S that may be required by a customer. Let Vi be the
union of all possible subsets for a particular fork-join node i, and let pi (S) be the probability
that subset S is required. The number of subsets in Vi is given by ( KkÅ0 i ( Kk i ) Å 2 Ki . For the
5
example sortie generation network, the 2 Å 32 possible subsets in V4 include M, {1},
{2}, {3}, {4}, {5}, {1, 2}, {1, 3}, . . . , {1, 2, 3, 4, 5}. If suitable system data were
available, exhaustive probabilities pi (S) could be explicitly assigned to each subset. Under
the assumption that all subsystem malfunctions occur independently, the probability that a
customer requires a particular subset S at fork-join node i is given by

pi (S) Å ∏ qik ∏ (1 0 qik ). (12)


k√S k√
/S

The mean of the associated conditional holding time at the fork-join node is E[maxk √ S
{Tik (N)}] (defined to be zero if S Å M ).
The computation of mean holding time mirrors a familiar problem from reliability theory
(that is, determining the mean time to failure for a parallel system of independent compo-
nents with exponentially distributed failure times [10]). Because the substation response

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


Dietz and Jenkins: Aircraft Sortie Generation 159

times are assumed to be independent by Approximation 2, the cumulative distribution


function (CDF) for conditional holding time is

F(t) Å ∏ P{Tik (N) ° t},


k√S

Å ∏ (1 0 exp{ 0 uik (N)t}). (13)


k√S

`
It is well known that E[X ] Å *0 {1 0 F(t)} dt for any nonnegative continuous random
variable X with CDF F(t), so

E[max{Tik (N)}] Å
k√S
*H
0
`
10 ∏
k√S
(1 0 exp{ 0 uik (N)t}) J dt

*H
`
Å 1 0 1 / ∑ exp{ 0 uik (N)t}
0 k√S

0 ∑ ∑ exp{ 0 ( uik (N) / uil (N))t}


k√S l√S
l xk

/ ∑ ∑ ∑ exp{ 0 ( uik (N) / uil (N) / uim (N))t}


k √ S l √ S m√ S

H S∑ D JJ
l xk mxk,l

0 rrr / ( 01) K ( S ) /1 exp 0 uik (N) t dt, (14)


k√S

where K(S) is the number of substations in S. Evaluation of the integral yields

1 1
E[max{Tik (N)}] Å ∑ 0 ∑ ∑
k√S k√S uik (N) k√S l√S uik (N) / uil (N)
l xk

1 1
/ ∑ ∑ ∑ 0 rrr / ( 01) K ( S ) /1 . (15)
k √ S l √ S m√ S
uik (N) / uil (N) / uim (N) (k √ S uik (N)
l xk mxk,l

The MVA algorithm can now be modified to accomodate a general network composed
of a set I of simple service stations and a set J of fork-join nodes. At each iteration, response
times for simple service stations are computed with the use of Eq. (3) or (4). For fork-
join substations with multiple servers, Approximation 1 permits analogous application of
the marginal local balance theorem and Little’s law to yield the formula
N
n
Rik (N) Å ∑ Pik (n 0 1ÉN 0 1). (16)
nÅ1 mik (n)

For single-server fork-join substations, Approximation 1 directly yields

Rik (N) Å sik {1 / Qik (N 0 1)}. (17)

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


160 Naval Research Logistics, Vol. 44 (1997)

The substation response rates uik (N) Å 1/Rik (N) are then used in Eq. (15) to determine
conditional mean holding times for the fork-join nodes. Cycle time for Station 1 can subse-
quently be calculated as

£i Ri (N) £i
CT 1 (N) Å ∑ / ∑ ∑ pi (S)E[max{Tik (N)}]. (18)
i√I
£1 i√J
£1 S⊆Vi k√S

With the cycle time determined, Eqs. (7) – (9) can still be used to compute throughputs,
queue lengths, and utilizations for simple service stations. Because a fork-join substation
ik is visited by a clone an average of qik times for each customer visit to node i, substation
throughput is given by

N£i qik
lik (N) Å . (19)
CT 1 (N) £1

The remaining substation performance measures are obtained with the use of Little’s law as

Qik (N) Å Rik (N) lik (N), (20)

Uik (N) Å siklik (N). (21)

A successive iteration of the algorithm can be initiated with the use of Eq. (3) or (4)
for simple service stations or Eq. (17) for fork-join substations with single servers. For
multiple server substations, Approximation 1 yields probability distributions for clone occu-
pancy as

lik (n)Pik (n 0 1ÉN 0 1)


Pik (nÉN) Å , n ú 0. (22)
mik (n)
N
Pik (0ÉN) Å 1 0 ∑ Pik (nÉN). (23)
nÅ1

These probabilities can then be employed in Eq. (16) to obtain new response times. Iterations
are continued until N is equal to the desired number of customers in the network.
It should be noted that the modified MVA algorithm produces exact results for any system
where N Å 1. When a single customer is in the network, no customer or clone ever waits
for service. Response times at fork-join substations therefore consist only of exponentially
distributed service times, so Approximations 1 and 2 are satisfied exactly and the product-
form nature of the network is preserved. For model applications where network servers
have relatively low utilizations (due to small customer population or abundant resources),
the probability that a customer or clone must wait for service is small, so the approximate
MVA algorithm produces highly accurate performance measures.

5. IMPLEMENTATION AND RESULTS


The algorithm described in the previous section has been implemented on a 80486/33-
MHz personal computer with a PASCAL software development environment. Recursive

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


Dietz and Jenkins: Aircraft Sortie Generation 161

procedures are used to form activity subsets and calculate the corresponding mean holding
times. Run time for the example network (including input and output) is less than 1 s for
30 aircraft and less than 2.5 s for 70 aircraft. Table 2 displays a complete set of performance
results for each network station at four different levels of network congestion (N √ {10,
30, 50, 70}). For a realistic complement of 30 aircraft, the estimated sortie generation rate
(throughput at the sortie station) is 5.507 launches per hour. The analytical sortie generation
rate as a function of N is shown by the line in Figure 2. The points shown in the figure
represent statistical estimates of the actual throughput values, produced through batched
means analysis of 10 7 hours of simulated operating time at eight congestion levels (each
with a 10 6-hour warmup period). The small bar inside each point has a vertical thickness
equal to the 95% confidence interval width for the simulation estimator. When executed
on a DEC AXP (Alpha) Model 500MP computing system, the simulation analysis required
9.23 hours of core processing time for N Å 30 and 11.97 hours for N Å 70. It is noteworthy
that the analytical and simulation results for sortie throughput agree within 0.3% over a
full range of network congestion levels.
In addition to analytical results, Table 2 displays simulation results for queue length and
utilization. Estimators for mean performance measures are shown, along with 95% confi-
dence interval half-widths. For each case where an analytical performance measure falls
outside the simulation confidence interval, percent error is also shown. The most significant
errors occur in the estimation of mean queue length at stations experiencing high utilization
factors. However, the largest observed error in Q (type 4 repair at N Å 50) is less than
2.4%, with most errors being much smaller. Generally, the analytical results agree very
closely with those produced by the lengthy simulations.
The example model confirms the exactness of the analytical results when N Å 1, and
verifies high accuracy at low network loading. System behavior at very high network loading
is also interesting. As explained by Kant [9], a bottleneck station eventually emerges in
any closed queueing network as N is increased. A very long queue forms at this station,
causing mean value analysis to eventually fail computationally as the network cycle time
overflows. For the example network, the bottleneck occurs at the type 4 (avionics) repair
activity. As N r ` , this substation operates at its maximum throughput of 1/s44 aircraft
per hour. The asymptotic sortie generation rate for the overall system is therefore l2 (` ) Å
£2 /( £4q44s44 ) Å 6.998 sorties per hour. At N Å 70, the bottleneck resource operates at a
utilization factor in excess of 98%. The corresponding analytical sortie generation rate is
6.909, which falls within 1.3% of the asymptotic value. Figure 2 and Table 2 both suggest
that the approximate algorithm produces accurate performance measure estimates for the
example network, even when the network is nearly saturated with a large customer pop-
ulation.

6. CONCLUSIONS
The analytical method presented in this article offers highly accurate estimates of mean
performance measures for an aircraft sortie generation process. In just a few seconds, results
can be obtained that would require hours of run time in a simulation model. The method
is therefore ideally suited for iterative analytical problems, such as the derivation of an
optimal resource structure under specified budget constraints. Results are particularly accu-
rate when resource utilization factors are relatively low.
The analytical model obviously cannot capture all of the variables that might appear in
a detailed simulation analysis. However, the MVA approximation could provide valuable

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


162 Naval Research Logistics, Vol. 44 (1997)

Table 2. Performance results for the example model.


Queue length (Q) Utilization (U)
% %
Activity R l MVA Simulation Error MVA Simulation Error
N Å 10
Taxi 0.250 2.303 0.575 0.576 { 0.000 00.13 0.575 0.576 { 0.000 00.13
Sortie 2.000 2.186 4.371 4.383 { 0.000 00.26 4.371 4.383 { 0.000 00.26
Troubleshoot 0.500 0.771 0.385 0.386 { 0.000 0.385 0.386 { 0.000
Repair 1 2.938 0.131 0.385 0.386 { 0.001 00.30 0.288 0.289 { 0.001
Repair 2 2.288 0.301 0.688 0.688 { 0.001 0.682 0.683 { 0.001
Repair 3 2.434 0.162 0.394 0.393 { 0.001 0.384 0.383 { 0.001
Repair 4 2.058 0.208 0.428 0.428 { 0.003 0.312 0.312 { 0.001
Repair 5 1.229 0.355 0.436 0.436 { 0.001 0.422 0.422 { 0.001
Turnaround 0.750 2.301 1.726 1.724 { 0.000 /0.09 1.725 1.724 { 0.001 /0.09
Mun. Upload 0.502 2.301 1.155 1.163 { 0.000 00.63 1.150 1.158 { 0.000 00.63
N Å 30
Taxi 0.250 5.796 1.449 1.445 { 0.001 /0.25 1.449 1.455 { 0.001 /0.25
Sortie 2.000 5.507 11.013 11.020 { 0.006 00.06 11.013 11.020 { 0.006 00.06
Troubleshoot 0.500 1.942 0.971 0.968 { 0.000 /0.33 0.971 0.968 { 0.000 /0.33
Repair 1 6.900 0.330 2.278 2.298 { 0.013 00.90 0.726 0.728 { 0.001 00.23
Repair 2 2.744 0.757 2.078 2.087 { 0.003 00.42 1.719 1.723 { 0.002 00.23
Repair 3 3.018 0.408 1.231 1.235 { 0.003 00.37 0.966 0.969 { 0.001 00.22
Repair 4 5.553 0.524 2.912 2.962 { 0.009 01.69 0.786 0.789 { 0.001 00.28
Repair 5 1.604 0.893 1.433 1.443 { 0.002 00.72 1.063 1.069 { 0.001 00.55
Turnaround 0.860 5.796 4.987 4.987 { 0.003 4.347 4.345 { 0.001 {0.04
Mun. Upload 0.650 5.796 3.768 3.744 { 0.004 /0.66 2.898 2.869 { 0.002 /1.02
N Å 50
Taxi 0.250 6.951 1.738 1.727 { 0.001 /0.65 1.738 1.727 { 0.001 /0.65
Sortie 2.000 6.603 13.206 13.225 { 0.008 00.14 13.206 13.225 { 0.008 00.14
Troubleshoot 0.500 2.329 1.164 1.172 { 0.001 00.63 1.164 1.172 { 0.001 00.63
Repair 1 14.147 0.396 7.188 7.184 { 0.015 0.871 0.871 { 0.001
Repair 2 3.369 0.908 3.039 3.061 { 0.009 00.73 2.061 2.070 { 0.003 00.41
Repair 3 3.531 0.489 1.727 1.734 { 0.005 00.45 1.159 1.161 { 0.002 00.16
Repair 4 15.002 0.629 9.432 9.663 { 0.023 02.39 0.944 0.943 { 0.000 00.10
Repair 5 1.972 1.071 2.112 2.120 { 0.004 00.38 1.275 1.276 { 0.001 00.14
Turnaround 1.228 6.951 8.536 8.521 { 0.018 /0.18 5.213 5.157 { 0.002 /1.08
Mun. Upload 1.034 6.951 7.188 7.184 { 0.015 3.475 3.442 { 0.002 /0.97
N Å 70
Taxi 0.250 7.241 1.810 1.794 { 0.002 /0.89 1.810 1.794 { 0.002 /0.89
Sortie 2.000 6.879 13.759 13.739 { 0.012 /0.14 13.759 13.739 { 0.012 /0.14
Troubleshoot 0.500 2.426 1.213 1.216 { 0.001 00.25 1.213 1.216 { 0.001 00.25
Repair 1 20.773 0.412 8.567 8.590 { 0.092 0.907 0.907 { 0.001
Repair 2 3.615 0.946 3.421 3.435 { 0.015 2.148 2.152 { 0.002 00.20
Repair 3 3.716 0.509 1.893 1.889 { 0.005 1.207 1.207 { 0.002
Repair 4 30.920 0.655 20.252 20.704 { 0.087 02.18 0.983 0.982 { 0.001 /0.07
Repair 5 2.116 1.116 2.361 2.371 { 0.009 00.41 1.328 1.328 { 0.002
Turnaround 1.584 7.241 11.473 11.356 { 0.050 /1.03 5.431 5.326 { 0.004 /1.98
Mun. Upload 1.395 7.241 10.098 10.016 { 0.056 /0.82 3.621 3.562 { 0.003 /1.65

insight when used as an adjunct to simulation. For example, the model could be employed
to produce external control variates or identify a starting point for a simulation search. It
therefore should be a useful tool for anyone concerned with the analysis of aircraft sortie

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


Dietz and Jenkins: Aircraft Sortie Generation 163

Figure 2. Results for the sortie generation rate.

generation or comparable problems involving multiserver queues and concurrent service


activities. Other potential applications include industrial machine maintenance, concurrent
tool changes in flexible manufacturing, and vehicle fleet maintenance.

ACKNOWLEDGMENT
The authors are grateful to the anonymous reviewers, whose thoughtful suggestions have
greatly improved the article.

REFERENCES
[1] Abell, John B., The Sortie Generation Model System, Logistics Management Institute, Washing-
ton, DC, Sept. 1981.

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914


164 Naval Research Logistics, Vol. 44 (1997)

[2] Baskett, Forest et al., ‘‘Open, Closed, and Mixed Networks of Queues with Different Classes
of Customers,’’ Journal of the A.C.M., 22, 248–260 (1975).
[3] Bruell, Steven C., and Balbo, Giofranco, Computational Algorithms for Closed Queueing Net-
works, Elsevier North Holland, New York, 1980.
[4] Conlon, John C. et al., Test and Evaluation of System Reliability, Availability, and Maintainabil-
ity, DOD Handbook, Vol. 3235.1H, U.S. Government Printing Office, Washington, DC, March
1982.
[5] Conway, Adrian E., and Georganas, Nicolas D., Queueing Networks—Exact Computational
Algorithms: A Unified Theory Based on Decomposition and Aggregation, The MIT Press,
Cambridge, MA, 1989.
[6] Cronk, Richard, and Wallace, Alan J., LCOM User Manual, Version 94.B, Air Force Aeronauti-
cal Systems Center, Wright-Patterson AFB, OH, March 1994.
[7] Emerson, Donald E., ‘‘An Introduction to the TSAR Simulation Program,’’ Technical Report
No. R-2584-AF, The Rand Corporation, Santa Monica, CA, Feb. 1982.
[8] Fisher, R.R. et al., ‘‘The Logistics Composite Model: An Overall View,’’ Technical Report
No. RM-5544-PR, The Rand Corporation, Santa Monica, CA, May 1968.
[9] Kant, Krishna, Introduction to Computer System Performance Evaluation, McGraw-Hill, Inc.,
New York, 1992.
[10] Kapur, K.C., and Lamberson, L.R., Reliability in Engineering Design, John Wiley and Sons,
Inc., New York, 1977.
[11] Lavenberg, S.S., and Reiser, M., ‘‘Stationary Probabilities at Arrival Instants for Closed
Queueing Networks with Multiple Types of Customers,’’ Journal of Applied Probability, 17,
1048–1061 (1980).
[12] Miller, L.W. et al., ‘‘Dyna-Sim: A Nonstationary Queuing Simulation with Application to the
Automated Test Equipment Problem,’’ Technical Report No. N-2087-AF, The Rand Corpora-
tion, Santa Monica, CA, July 1987.
[13] Rao, P.C., and Suri, Rajan, ‘‘Approximate Queueing Network Models for Closed Fabrication/
Assembly Systems. Part I: Single Level Systems,’’ Production and Operations Management,
4, 244–275 (1994).
[14] Suri, Rajan, ‘‘Robustness of Queueing Network Formulas,’’ Journal of the A.C.M., 30, 564–
594 (1983).

/ 8M0E$$0914 12-27-96 10:26:36 nra W: Nav Res 0914

You might also like