Professional Documents
Culture Documents
Kappas 1991
Kappas 1991
To cite this article: George Kappas & Tunce M. Yegulalp (1991) An application of closed queueing networks theory
in truck-shovel systems, International Journal of Surface Mining, Reclamation and Environment, 5:1, 45-51, DOI:
10.1080/09208119108944286
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
I~ternationalJournal of Surface Mining and Reclamation 5 (1991):45-53
ABSTRACT: Primary applications of queueing networks theory have been in the area of computer systems and com-
puter communication networks analysis and performance. The application introduced here is a steady state anal-
ysis of a generalized truck-shovel system in a typical open pit mine. The service time distributions of each
activity are assumed to be of a general form. The approach used here minimizes the error of performance param-
eter estimates of the most saturated service center. The formulation allows one to obtain estimates of system
Downloaded by [Universitaets und Landesbibliothek] at 11:31 11 December 2013
parameters relatively fast and very close to those obtained by simulation. The relative error of the critical
performance parameter estimates as compared to those obtained by simulation do not exceed 5%. The output of
the model can be used within various optimization schemes for design and operational control of open pit min-
t
ing systems.
··(~)=l n
1979 ): 1 if nm=O
(4) n~
(Xm t: ifnm>O (13)
bm (r)
,-I
m-I
Bm (nm) (15)
a customer (trucK) departs from service center k
and arrives at service center m. The product pm Thus the estimation of steady state probabili-
bm(r) denotes the effective service rate of ser- ties, requires the evaluation of the normalizing
vice center m when there are r customers (trucks) constant G(N). A straight forward evaluation of
in this service center. It is noted that the ser- G(N) is avoided becAuse the number of feasible
vice rate in service center m is a function of the states, given by equation (3), increases rapidly
Downloaded by [Universitaets und Landesbibliothek] at 11:31 11 December 2013
number of customers (trucks) in this service cen- as a function of Nand M. Despite the large number
ter. The paramete~ pm can be viewed as the number of arithmetic operations required to determine
of customers (trucks) departing service center m G(N), there exist an efficient algorithm which com-
per unit of busy time of the service centet m. putes not only G(N) but also G(I), ••• ,G(N-l). The
It is clear from equation (5), that the vector procedure was developed by Buzen and has since be-
come known 4S the Buzen algorithm (BUZen, 1971;
a = (a1,....am,...,aM ) T (6) 1973).
The characteristics of Markovian networks men-
is an eigenvector of the transition probability ma- tioned above, will be used next in the description
trix and analysis of the extended product form tech-
nique. This technique will assume product form
r
1l
112 ••• 11M
1'21 1'22 ... '7IJ
J steady state probabilities for general networks.
R= (7)
[
rMI'M'Z .. ·rMM
EXTENDED PRODuCT FORM TECHNIQUE
The parameters am, V, m = 1, ••. ,M are viewed as AS it is pointed out by Reiser at al., (1976),
the relative throughputs since they are the the process X(t) of the closed network can be
throughputs for N~l. It is noted that Ql,.··,aM viewed as a B~~atified samoling process of an op~~
can be determined up to a mUltiplicative constant net~rk. This means that the state space of the
only. Thus any vector ka is also an eigenvector of closed network is generated from the state space
the transition probability matrix R. Equation (5) of an open network which has the same number of
in fact equates the flow of customers (trucks) in service centers, by choosing only the states in
and out of each service center. The matrix R is ir- which the total number of customers in the open
reducible; it is also stochastic, that is, the sum network equals N. Bence, the state probabilities
of the elements of each row is equal ~o 1: of the closed queueing network can be produced by
conditioning the state probabilities of the open
M
network on the event that the sum of the customers
L rlj= 1, V i= l, ...,M (8)
is N.
j=1
Given the above rationale, if Fm(nm), VI m=I,2, •
•• , M were the marginal probabilities of the open
The function Pm(.)is defined by: Markovian network, then, it can be shown that if
they were used in equations (4) and (13) instead
o ifr=O
of the functions Bm(nm), they would lead to the
bm(r)= min(r.Sm) if O<r:SN&FCFS (9) same steady state probability eetimates for the as-
{r if O<r:sN&IS sumed Markovian network. This observation shows
that the same state probabilities are obtained by
where Sm is the number of servers in service cen- different sets of input parameters. since the
ter m. The quantity G(N) in equation (4) is a nor- input parameter a, (which is an eigenvector of R),
malizing constant such that is the only one that is not determined uniquely,
it can be modified without affecting the state
L "(0) = 1 (10)
probabilities. Therefore, if a* is another eigen-
DeS(N,M)
vector of R, then
and is given by the following equation:
(16)
eters). There is only one set of input parameters 1,= max {1m) .If m=1....,M (23)
which minimizes the difference between actual m=l
otate probabilities and the ones obtained by prod-
uct form expressions. Let k be the estimate of the unknown value of k.
The method to be applied here for finding k is the
The sets of parameters a~ V m a 1,2, ••• ,M are golden section search method. The computer routine
viewed as the relative throughputs. They are used used for this purpose (ZXGSM, IMSL, version 8),
8S arrival rates in the marginal probability ex-
computes k that minimizes the unimodal function
pressions. Since they are functions of k, so are z(k) to a desired accuracy e. The function z(k) is
the marginal probabilities. For this reason, the defined for positive values of k and is assumed to
marginal probability expressions are written as be continuous. The method employed, requires a
P'm(~1k) • known finite interval K which contains the indepen-
The objective 1s to identify the set a~ , V, rna dent variable k. This interval i8 given as fol-
1, ••• ,K, such that it is the closest to the set of lows:
absolute throughputs (arrival rates) ~, V, m =
1,2, •.. ,M. The absolute throughputs (arrival (24)
rates) are functions of k since, by definition,
they are functions of Fm("m;k):
N
Am(k) =!'mL bm (nm)Fm (nm ;k) (18)
nm-t DERIVATION OF MARGINAL PROBABILITIES OF M/G/1/N
QUEUEING SYSTEMS
The arrival rates, as it was mentioned earlier,
satisfy the flow equations -they compose an eigen- The approach to be 'used in obtaining the marginal
vector of R- and since a is also an eigenvector of probabilities Fm(~;k), is operational analysis
R, there should be a constant w SUch that (BUzen, 1976; Denn~ng et al., 1977; Shum, 1980).
Listed below are some definitions. Let
(19) .Fm(~;k), "mcO,I, •.• , N, be the marginal
probability that "m custo~ers are in ser-
since the parameters ..l-(k), V, m CI 1,2, ••• ,M are vice center m which is an MIGI11N queueing
functions of k, w is ;1so a function of k. But system.
~m(k), V, m a 1,2, ••• ,H, defined by (18), are hypo-
,
• F m("m;k), !\n=0,1, ." , N-1 , be the mar-
thetical absolute arrival rates which satisfy the ginal probab~lity that a departing cus-
flow equations, therefore their exact values may tomer from service center m "sees" ~ cus-
not exist. Considering equation (19), let: tomers behind him.
Am(k) .~(Ilm,k), "maO, 1, 2, •.. , be the mar-
wm(k) 1::1 am t If m =1•...,M (20)
g~nal probability that "m customers are in
service center m when it is viewed as an
Adjusting k so that the arrival rates A (k) sat- MIGII queueing system.
isfy equation (19), is equivalent to adju~ting k
The equations that relate the above probabili-
so that all values of wm(k), V, m = 1,2, ••• ,M
ties are:
are equal. Such a value for k may not exist. Thus
an algorithm is necessary for minimizing the dif-
ferences between several values of w (k). For this Fm(nmJ<) = f';' (nm;k) If nm = 0.1..... N-1 (25)
purpose one may look for a value ofmk which mini- Fm(O;k) +Pm
mizes the following function:
Fm(N;k) = 1 _ 1 (26)
z(k) -
_ M
L 1m (~
(k) - 1
)2 (21) where
Fm(O;k) +pm
m-J Ws
When Cm=O, the probabilities Qm(nm;k) are given by are dependent upon the state probability distribu-
the following expressions (shum, 1980): tion given by equatione (14) and (15). since the
state probabilities are functions of the normaliz-
Qm(O;k) = 1-pm (30) ing constant G(N), we need a tool for computin9
G(N). For this purpose we will use the Buzan algo-
Qm(l;k) = (l-pm)(epm -1) (31) rithm (Buzen, 1971; 1973). To derive the algorithm
let
[~ e jPm(_1)"m-j(jPm)nm.- j + m
Qm(nm;k)=(l-pm)
j_1 (nm - J) I ~(n) = L
nE S(n,m) i=l
n Fi(n/;k)
(jpm)nm -j -1]
(nm _ j _ 1) I '
'f
nm = 2.3.... (32) Y m = 1....,m, n = O,I,...,N (36)
where
m
where the second factor of the r.h.s. in equation
(32) is ignored for j=nm. L ni=n.
i""1
When CmS1, the actual distribution will be ap-
proximated by means of erlang distribution. The By definition, from (15) and (36), G(n)aqM(n),
marginal probabilities are obtained by interpola-
tion of values of marginal probabilities of queue- ¥ , m=!, ••• ,M. The function gM(n) is viewed as the
ing systems that have different coefficients of normalizing constant for the closed queueing net-
variation. The three queueing systems are: a) work in which there are n customers. From (36),
M/D/l. The service times in this system are deter- and for m=l we have 91(n)=F 1(n;k),¥ , n-1, .•• ,N.
ministic (constant) and the coefficient of varia- Also, for n=O, 9m(D)c1,¥ m=l, .•• ,M. For m >1,
tion of the service time distribution is o. b) we have from (36):
M/E2/1 where the service times follow the 2-erlang m
distribution that has a coefficient of variation
of l/{2, and C) MIMI 1 where the service time dis-
~(n) =L Il Fi(n/;k)
n ES(n,m) i=1
tribution is exponential and the coefficient of
variation is 1. Procedures for obtaining the mar- n
ginal probabilities of MIDll and MIMll were dis- = L Fm(j;k)~-I(n-j) (37)
cussed above, thus we have to obtain only the j=O
probabilities Qm(Omik) for the M/E2/l queueing sys-
tem. These probabilities are given as follows: computation of various performance parameters re-
quires the estimation of the marginal probability
distribution of customers in each service center.
Ii nm = 0,1,2,... (33) It is useful to begin by considering the marginal
distribution of customers in the Hth service cen-
ter. The probability that there are j customers in
where zl' 22 are the roots of the characteristic service center Mis:
equation of the z-transform of 0m(Om;k) and A, B
are functions of zl' z2 • P(nM=j) = L J«nl.nz,....nM)
When Cm ~ 1, we use hyperexponential-2 distribu- n ES(n,m); nJr"j
tion for approximating the actual distribution.
The density function of an hyperexponential-2 dis- =F ("k) IPI-1(N-j) (38)
M J. G(N)
tribution is determined by three parameters p, Pml
and Pm2 and is given by the following expression: Marginal distributions of customers at any other
service center can be obtained by permuting the
order in which the service centers are numbered 80
that the facility of interest is designated as the
The queueing system of service center m, is Mth service center and then by applying the algo-
equivalent to one in which there are two servers, rithm to the permuted network. considering that
with service rates of ,um1 and Pm2 respectively,_ the service center of interest is in the Mth posi-
and any arriving customer chooses one of the two tion, we can express the probability that service
servers with probability p and 1-p respectively. center M is idle. For nM=O we have:
~since the distribution function hm(tm) is de-
fined by three parameters p,Pm1and ,um2 and only
the first two moments of the actual distribution (39)
are known, another relation between P,Pml and P m2
50
The above ie the probability that an arriving also the expected utilization of each server of
customer (truck) will be served immediately. We service center H.
consider that the service center M is busy if at The expected arrival rate at service center H
least ona server of the service center is bUSy. is.
Therefore P(service center M busy) a1-p(nMDO), N
where P(DM-O) is obtained from (38). AM = I'M}; bM IJ)P(nM =j)
)al
consider the event that at least i servers of
service center H are busy if nK ~ i. The probabil-
ity of this event lSI N (N 0)
=I'M}; bMIJ)FM(j;k)8M-1 -) (46)
P(nM ~ i ) C 1 - P(nM < i) )=1 G(N)
The expected utilization PH of the service cen-
I-I
ok) B.H(N - j) ter H can also be expressed by the following ex-
C 1-}; V
FM, G(N) pression: PM= ~M/sM"M)· The expected tims spent by
)aO
(40) a customer in service center H is obtained by the
Y i = 0, 1, ...,SM
Little formula (Little, 1961). E(ttM)= g(nM)/A M,
and the expected time spent by a customer in
All servers 8 M of the service center M are busy queue in service center M iSI E(tqM)Q E(qM)/AH "
(service center ful~y utilized) lac If 8 is the set of service centers corresponding
to dumping or crushing points, and D is the capac-
P(nM ~SM) = 1- P(nM < SM) ity of a truck (tons), then the production rate of
each truck is:
=1-};
'''-1 FM(j;k)KM-1
(N
-/
0)
(41)
)-0 G(N) D
PR = - } ; Am
Downloaded by [Universitaets und Landesbibliothek] at 11:31 11 December 2013
MODEL VERIFICATION
1 _.!. ~ ( 0) F (j k) /1M-I (N - j) (42)
S" )-0 S" - J M; G(N)
In order to test the relative accuracy and speed
of the analytical approach described above, a se-
The probability that i - i n number - particular ries of tests have been conducted. These tests
servers in service center H are busy, iSI were designed to observe the performance of the
model for critical values of the coefficient of
P ( I partkul., ) =1 _ ~ F (j k"'.-I eN - j) + variation of service time distributions and the
~
0
sctverl arc busy M, G(N) number of customers (trucks) in the system. The
)-0
EPF approach was compared to the markovian ap-
proach and monte carlo simulation. In these tests
the difference between the two results were mea-
sured by the absolute error, E, and the relative
error, e, defined as:
The expected number of customers in service cen-
ter M (queue and service) iss E
E =x -X and e=-
X
N
E(nM)c}; jP(nM=j) where x is the value of a system parameter ob-
j-I tained using the analytical approach, and X is the
corresponding value obtained by simulation. These
(N- 0) tests as reported by Kappas (KappaS 1999) led to
OF vok)8M-1 J (44)
J M. G(N) the following conclusions:
A customer will be in queue of service center M • In general, the Extended Product Form
if all servers 8M are occupied. This event will method for obtaining performance and pro-
occur when nM ~ 8 M• Therefore, the expected number duction parameters of a network is found
of customers in queue lSI to be superior to the one based on Marko-
vian assumptions" When the coefficient of
N variation is not equal to or close to
E (q.. ) =}; (j - s" )P(n" =j) zero, the relative error does not exceed
j-sM + 1 0.5
not les8 than 1% and the simulation clock trucks until N is near 10, and than it is leveled
reached 105 units of time. off near 5 tons/min. The same behavior also holds
for the total production (Figure 4.)~ On the other
• The main reasons for the relatively signif- hand, the production rate of the system (network)
icant time required for implementing the decreases as N increases (Figure 2.). If the objec-
EPF approach were: tive of this exercise WAS to maximize the trans-
ported tonnage with the minimum number of trucks,
- Numerical estimation of the optimal then 10 trucks would have been the optimum choice
value of k ,
for this system. Figure 5 shows the expected
- Estimation of marginal probability throughput for each operation (service center) in
expressions required three steps.
number of trucks per minute, as a function of the
Despite these facts, the authors believe that total number of trucks (N) in the system. For N >
significant improvements can be made on CPU time 10 no significant improvement in throughput is at-
tained. Figure 6 shows the utilization factor for
by improving the efficiency of the computer code.
each operation as a function of N, and Figure 7
shows the expected number of trucks in each opera-
EXAMPLE PROBLEM tion. It is clear from these figures that the re-
pair and maintenence operation becomes quickly sat-
In order to demonstrate the use of this approach urated for N>10, and act AS a bottleneck in the
system. Figure 8 shows the probabilities of find-
in mining, a simple truck-shovel system is consid- ing a service center (operation) bUsy~
ered here as an example problem. Five typical ac-
tivities are identified as the service centers of
the system. These areJ
ACKNOWLEDGEMENTS
1. Loading
2. Traveling loaded The authors wish to thank New york Mining and Min-
3. Dumping eral Resources Research Institute for ~he partial
4. Traveling empty support provided by the Department Interior'S Min-
5. Repair and/or maintenance. eral; Institutes program administered under allot-
ment grant number G1184136~
A schematic representation of this network is
shown in Pigure 1. corresponding transition proba-
bility matrix is assumed to be: REFERENCES