Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: We consider a new class of batch arrival retrial queues. By contrast to standard batch arrival retrial
Received 19 November 2008 queues we assume if a batch of primary customers arrives into the system and the server is free then
Accepted 24 March 2009 one of the customers starts to be served and the others join the queue and then are served according
Available online 29 March 2009
to some discipline. With the help of Lyapunov functions we have obtained a necessary and sufficient con-
dition for ergodicity of embedded Markov chain and the joint distribution of the number of customers in
Keywords: the queue and the number of customers in the orbit in steady state. We also have suggested an approx-
Queueing
imate method of analysis based on the corresponding model with losses.
Markov processes
Retrials
Ó 2009 Elsevier B.V. All rights reserved.
Batch arrivals
Steady state
0377-2217/$ - see front matter Ó 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2009.03.033
G.I. Falin / European Journal of Operational Research 201 (2010) 786–790 787
The pool of sources of repeated calls may be viewed as a sort of arrived into the system during the service period. As we have
queue and often is said to be the orbit. Every such source produces noted, Ezm ¼ kðzÞ. Thus,
a Poisson process of repeated calls with intensity l. If an incoming
E zRdþ1 jRd ¼ n; E0 ¼ zn1 kðzÞ: ð1Þ
repeated call finds the server free, it is served and leaves the sys-
tem after service, while the source which produced this repeated 2. with probability kci
a group of i new customer arrives before
kþln
call disappears. Otherwise, the system state does not change. arrival of a repeated call from the orbit. Immediately after
Clearly a customer from the orbit can occupy the server only if this arrival the server begins the service of the customers
there is no primary customer in the queue in front of the server. from this group. The server is busy during i independent ser-
Thus, the returning customers may be viewed as low priority vice times and when all customers from the group are served
customers. the queue becomes empty, so that the departure epoch of the
The service time distribution function is BðxÞ for both primary last customer from the group is the time gdþ1 . The number of
calls and repeated calls. customers in the orbit at this epoch, Rdþ1 , equals to
We assume that the input flow of primary calls (i.e. arrival n þ m1 þ þ mi , where mj , 1 6 j 6 i, is the number of new pri-
epochs and sizes of batches), intervals between repeated trials, mary customers which arrived into the system during the
and service times are mutually independent. service period of the jth customer from the group. Random
R1
Let bðsÞ ¼ 0 esx dBðxÞ be the Laplace–Stieltjes transform of the variables m1 ; . . . ; mi are independent and have the same distri-
service time distribution function BðxÞ, Mb be the mean service bution with generating function kðzÞ. Thus,
time, r2b be the variance of the service time, q ¼ kMc M b the system
load due to primary customers (offered traffic), bðxÞ ¼ B0 ðxÞ= i
E zRdþ1 jRd ¼ n; Ei ¼ zn ðkðzÞÞ : ð2Þ
ð1 BðxÞÞ be the instantaneous service intensity given that the
elapsed service time is equal to x, kðzÞ ¼ bðk kcðzÞÞ. It is easy to
P
see that kðzÞ ¼ 1 n
n¼0 kn z , where kn is the probability that during
Eqs. (1) and (2) yield that
service time of a customer exactly n new primary customers arrive nl n1 k
into the system, and q ¼ k ð1Þ.
0
E zRdþ1 jRd ¼ n ¼ z kðzÞ þ zn cðkðzÞÞ: ð3Þ
k þ nl k þ nl
At time t let Q ðtÞ be the number of customers in the queue
(including the customer in service if any) and RðtÞ be the number Since evolution of the system after time gd depends on the his-
of sources of repeated calls (which may be viewed as a sort of tory of the system before time gd only through the value
low priority queue). The process ðQ ðtÞ; RðtÞÞ which describes the Rd ¼ n, the sequence of random variables Rd forms a Markov
number of customers in the system is the simplest and simulta- chain, which is the embedded chain for our queueing system.
neously the most important process associated with the above It should be noted Rd is not the only embedded Markov chain for
queueing system. our queueing system. The sequence of pairs ðQ 0d ; R0d Þ where Q 0d is
If the service time distribution is not exponential then the pro- the number of customers in the queue and R0d is the number of
cess ðQ ðtÞ; RðtÞÞ is not Markov. In this case we introduce a supple- customers in the orbit at the epoch of dth departure forms an-
mentary variable: if Q ðtÞ P 1, we define nðtÞ as the elapsed service other embedded Markov chain.
time of the call being served.
3.1. Ergodicity of the embedded Markov chain
Thus: nl k
¼ ð1 þ qÞ þ M c q:
k þ nl k þ nl
nl
1. with probability kþn l a customer from the orbit arrives before
arrival of new customers. Immediately after this arrival the As n ! þ1 there exists x ¼ lim xn ¼ 1 þ q. Applying Foster’s crite-
server begins the service of the customer and when the ser- rion (see, for example, Pakes (1969)) we can guarantee that the
vice ends the number of customers in the queue equals 0, so embedded Markov chain is ergodic if q < 1. Besides, since
that the departure epoch is the time gdþ1 . The number of cus- Rdþ1 Rd P 1, for chain Rd the mean down drift is bounded below.
tomers in the orbit at this epoch, Rdþ1 , equals to n 1 þ m, Thus, applying the criterion from Sennott et al. (1983) we can guar-
where m is the number of new primary customers which antee that inequality q < 1 is necessary for ergodicity. h
788 G.I. Falin / European Journal of Operational Research 201 (2010) 786–790
n¼0
k þ nl k þ Rd l ðk þ nlÞp0n ¼ p1n ðxÞbðxÞdx;
0
X
n
Since p0in ðxÞ ¼ ðk þ bðxÞÞpin ðxÞ þ k ck pi;nk ðxÞ; if i P 1;
k¼1
uðzÞ ¼ kwðzÞ þ lzw0ðzÞ; Z 1
pin ð0Þ ¼ piþ1;n ðxÞbðxÞdx þ kci p0;n þ ðn þ 1Þlp0;nþ1 di;1 ; if i P 1:
0
we get the following equation for the generating function wðzÞ:
l½kðzÞ z w0 ðzÞ ¼ k½1 cðkðzÞÞ wðzÞ: ð5Þ where dm;n is Kronecker’s delta, so that the multiplier di;1 indicates
that the corresponding term is present only in the case i ¼ 1.
Consider the coefficient f ðzÞ ¼ kðzÞ z. Note that: For generating functions
1. f ð1Þ ¼ kð1Þ 1 ¼ 1 1 ¼ 0; X
þ1 X
þ1
k 1 cðkðzÞÞ so that
w0 ðzÞ ¼ wðzÞ;
l kðzÞ z þ1 X
X þ1 X
þ1
Pþ1
dp0 ðzÞ Since y 1cðyÞ is the generating function of the tails ci ¼ k¼i ck :
pi ðz; 0Þ ¼ piþ1 ðz; 0ÞkðzÞ þ kci p0 ðzÞ þ l di;1 ; if i P 1 ð13Þ 1y
dz X
þ1
1 cðyÞ
correspondingly. y ¼ yi ci ;
1y i¼1
For generating function pðy; z; xÞ Eq. (13) becomes:
kðzÞ dp ðzÞ we may find the distribution of the queue length:
1 pðy; z; 0Þ ¼ p1 ðz; 0ÞkðzÞ þ kcðyÞp0 ðzÞ þ ly 0 :
y dz
ð14Þ PðQ ðtÞ ¼ 0Þ ¼ 1 q;
Eliminating p1 ðz; 0Þ from (12) and (14) we get: qð1 qÞ
PðQ ðtÞ ¼ 1Þ ¼ q2 þ ; ð18Þ
Mc
kðzÞ dp ðzÞ qð1 qÞ
1 pðy; z; 0Þ ¼ kð1 cðyÞÞp0 ðzÞ þ lðy zÞ 0 : ð15Þ PðQ ðtÞ ¼ iÞ ¼ ci ; if i P 2:
y dz Mc
Replacing y ¼ kðzÞ yield
It is interesting to note that this distribution does not depend on the
dp0 ðzÞ
l½kðzÞ z ¼ k½1 cðkðzÞÞp0 ðzÞ: ð16Þ parameter l which describes the behaviour of the blocked
dz
customers.
This equation is identical to Eq. (5) and thus On the other hand if we assume that l ¼ 0, i.e. blocked custom-
Z z ers are lost forever (and do not influence the queue), then it is easy
k 1 cðkðuÞÞ to see that the corresponding process ðQ 0 ðtÞ; n0 ðtÞÞ is Markovian.
p0 ðzÞ ¼ p0 ð1Þ exp du :
l 1 kðuÞ u Clearly this process is ergodic for any arrival rate and it can be
shown that in the case l ¼ 0 the stationary distribution of the
Now from (16) and (15)
queue length is given by
1 cðyÞ z y 1 cðkðzÞÞ
pðy; z; 0Þ ¼ ky þ p0 ðzÞ; 1 q
kðzÞ y kðzÞ y kðzÞ z PðQ 0 ðtÞ ¼ 0Þ ¼ ; PðQ 0 ðtÞ ¼ iÞ ¼ ci ; if i P 1:
1þq ð1 þ qÞMc
and so from (11) ð19Þ
1 kðzÞ 1 cðyÞ z y 1 cðkðzÞÞ
pðy; zÞ ¼ y þ p0 ðzÞ; Thus stationary performance characteristics of the queue are not
1 cðzÞ kðzÞ y kðzÞ y kðzÞ z
continuous at the point l ¼ 0.
However, there is an interesting relation between distributions
(18) and (19). With this goal consider the functioning of the origi-
Pðy; zÞ
nal system (with l > 0) from the point of view of an observer who
1 cðyÞ 1 kðzÞ z y 1 kðzÞ 1 cðkðzÞÞ has only information about the state of the queue (including the
¼ 1þy þy p0 ðzÞ:
kðzÞ y 1 cðzÞ kðzÞ y 1 cðzÞ kðzÞ z server). This observer sees that from time to time groups of cus-
tomers arrive into the system. If the server is free at the time of ar-
Now the unknown constant p0 ð1Þ can be found from the normaliz-
rival of a group, one customer from the group starts to be served
ing condition Pð1; 1Þ ¼ 1: p0 ð1Þ ¼ 1 q, which completes the
immediately and the rest join the queue. If the server is occupied
proof. h
at the time of arrival then (from the point of view of our observer)
With the help of the generating function Pðy; zÞ we can find var- the group disappears (although actually it joins the orbit).
ious performance characteristics of the system. For example, the Thus, from the point of view of an observer, which has only
distribution of the number of customers in the orbit has generating information about the state of the queue, the original retrial queue
function functions as a corresponding loss system. However, it is important
to realize that the input flow to this loss system is composed of all
1 z 1 cðkðzÞÞ
EzRðtÞ ¼ Pð1; zÞ ¼ ð1 qÞ 1 þ calls (both primary and repeated), i.e. is a composite of the initial
kðzÞ z 1 cðzÞ
Z z flow of primary calls, which reflects real needs in service, and the
k 1 cðkðuÞÞ flow of repeated calls, which is connected with delays in service.
exp du :
l 1 kðuÞ u Thus, this joint flow has a more complex structure than the flow
In particular, the mean number of customers in the orbit is of primary calls (as a matter of fact, we cannot even describe this
structure). We can, however, make some simplified assumptions
k qM c q2 qð2 qÞðMc 1Þ q2 about the structure of the joint flow, which allow us to transform
ERðtÞ ¼ þ þ þ
l 1 q 2ð1 qÞ 2ð1 qÞ 2ð1 qÞ the loss model into a tractable model. Assume (although formally
r2b qð2 qÞ r2c it is not correct) that the flow of repeated calls is Poisson with
þ : ð17Þ the rate (see (17))
M2b 2ð1 qÞ M c
and does not depend on the (Poisson) flow of groups of primary cus-
1 cðyÞ q
EyQ ðtÞ ¼ Pðy; 1Þ ¼ ð1 qÞ þ yq2 þ y ð1 qÞ: tomers. Then the joint flow is Poisson with rate
1 y Mc
1 q þ qM c
In particular, the mean queue length is k ¼ k þ r ¼ k :
1q
qð1 qÞðMc 1Þ r2c
EQðtÞ ¼ q þ þ qð1 qÞ : At every arrival epoch with probability kk a group of primary
2 2M c
customers arrives and with probability kr single customer from
790 G.I. Falin / European Journal of Operational Research 201 (2010) 786–790
the orbit arrives. Thus, the distribution of the group size for this References
joint flow is given by:
Artalejo, J.R., Gomez-Corral, A., 2008. Retrial Queueing Systems. A Computational
Approach. Springer.
k r qMc þ c1 ð1 qÞ
c1 ¼ c1 þ ¼ ; Breuer, L., Dudin, A.N., Klimenok, V.I., 2002. A retrial BMAP/PH/N system. Queueing
k k 1 q þ qM c Systems 40, 433–457.
k ci ð1 qÞ Breuer, L., Klimenok, V., Birukov, A., Dudin, A., Krieger, U.R., 2005. Modeling the
ci ¼ ci ¼ : access to a wireless network at hot spots. European Transactions on
k 1 q þ qM c Telecommunications 16, 309–316.
Dudin, A.N., Klimenok, V.I., 2000. A retrial BMAP/SM/1 system with linear repeated
requests. Queueing Systems 34, 47–66.
The mean size of the group for the joint flow is Dudin, A.N., Krishnamoorthy, A., Joshua, V.C., Tsarenkov, G.V., 2004. Analysis of the
BMAP/G/1 retrial system with search of customers from the orbit. European
X
þ1
Mc Journal of Operational Research 157 (1), 69–179.
M c ¼ ci ¼ Falin, G.I., 1976. Aggregate arrival of customers in one-line system with repeated
i¼1
1 q þ qM c calls. Ukrainian Mathematical Journal 28 (4), 437–440.
Falin, G.I., 1981. Functioning under nonsteady conditions of a single-channel system
with group arrival of requests and repeated calls. Ukrainian Mathematical
and thus the offered traffic from the point of view of the observer Journal 33 (4), 429–432.
which has information only about the queue is Falin, G.I., 1988. On a multiclass batch arrival retrial queue. Advances in Applied
Probability 20 (2), 483–487.
q Falin, G.I., Templeton, J.G.C., 1997. Retrial Queues. Chapman and Hall.
q ¼ k Mc Mb ¼ : Grishechkin, S.A., 1992. Multiclass batch arrival retrial queues analyzed as
1q branching processes with immigration. Queueing Systems 11 (4), 395–418.
He, Q.-M., Li, H., Zhao, Y.Q., 2000. Ergodicity of the BMAP=PH=s=s þ K retrial queue
Now using (19) calculate the distribution of the queue length in this with PH-retrial times. Queueing Systems 35, 323–347.
model: Ke, J.-C., Chang, F.-M., 2009. M ½x =ðG1 ; G2 Þ=1 retrial queue under Bernoulli vacation
schedules with general repeated attempts and starting failures. Applied
Mathematical Modelling 33, 3186–3196.
1 Kim, C.S., Klimenok, V.I., Lee, S.C., Dudin, A.N., 2007. The BMAP/PH/1 retrial
PðQ ðtÞ ¼ 0Þ ¼ ¼ 1 q;
1 þ q queueing system operating in random environment. Journal of Statistical
Planning and Inference 137, 3904–3916.
q qð1 qÞ
PðQ ðtÞ ¼ 1Þ ¼ c 1 ¼ q2 þ ; Kim, C., Klimenok, V.I., Orlovsky, D.S., 2008. The BMAP/PH/N retrial queue with
ð1 þ q ÞMc Mc Markovian flow of breakdowns. European Journal of Operational Research 189
(3), 1057–1072.
q qð1 qÞ
PðQ ðtÞ ¼ iÞ ¼ c i ¼ ci ; if i P 2: Kulkarni, V.G., 1986. Expected waiting times in a multiclass batch arrival retrial
ð1 þ q ÞM c Mc queue. Journal of Applied Probability 23 (1), 144–159.
Langaris, C., Moutzoukis, E., 1995. A retrial queue with structured batch arrivals,
As one can see from (18) we obtained the exact distribution of the priorities and server vacations. Queueing Systems 20, 341–368.
queue length in the original model. Li, Q.L., Ying, Y., Zhao, Y.Q., 2006. A BMAP/G/1 retrial queue with Markovian a server
subject to breakdowns and repairs. Annals of Operations Research 141, 233–
The above analysis shows that from the point of view of the sta- 270.
tionary distribution of the process Q ðtÞ the retrial model under Pakes, A.G., 1969. Some conditions for ergodicity and recurrence of Markov chains.
consideration can be thought of as the corresponding loss model Operations Research 17, 1058–1061.
Sennott, L.I., Humblet, P.A., Tweedie, R.L., 1983. Mean drifts and the non-ergodicity
with increased arrival rate k ¼ k þ r and adjusted batch size distri- of Markov chains. Operations Research 31, 783–788.
bution. The additional load r can be thought of as a load formed by Yang, T., Templeton, J.G.C., 1987. A survey on retrial queues. Queueing Systems 2
sources of repeated calls. (3), 201–233.