You are on page 1of 19

M |Ek |1 Model

We consider a single-server queueing system


where the arrival process is Poisson with rate
λ and service times are independent and identi-
cally distributed random variables with density
function
kµ (kµt)k−1 −kµt
f (t) = e ,
(k − 1)!
i.e. service times have the Erlang distribution
of order k (k-Erlang distribution) with mean
1
. So, the service time may be thought of as
µ
consisting of k independent exponentially dis-
1
tributed phases, each with mean . So, we

assume that after the arrival a customer has
k phases of service. Until all k phases of ser-
vice have been completed, the next customer
should wait in the queue. It is convenient to
number the phases of service in reverse order:
the first phase is phase k, the second phase
is phase k − 1, and so on. Hence, a customer
leaving phase 1 actually leaves the system.
1
Let pni(t) be the probability that at time t there
are n customers in the system and the cus-
tomer in service is in phase number i (recall
that the phases are numbered backward). In
other words, pni(t) is the probability that there
are n customers in the system and the cus-
tomer in service needs to complete i remaining
phases in order to leave the system. Let p0(t)
be the probability that at time t there are no
customers in the system.

For example, consider M |E2|1. The transition


diagram for this model is

λ λ
λ λ λ

0 1,2 2µ 1,1 2,2 2µ 2,1

2µ 2µ

2
In the general case of M |Ek |1 model, for suffi-
ciently small h,

p0(t + h) = (1 − λh)p0(t) + kµhp1,1(t) + o(h)


p1,i(t + h) = (1 − λh − kµh)p1,i(t)
+ kµhp1,i+1(t) + o(h), 1 ≤ i ≤ k − 1,
p1,k (t + h) = (1 − λh − kµh)p1,k (t) + kµhp2,1(t)
+ λhp0(t) + o(h)
pnk (t + h) = (1 − λh − kµh)pnk (t) + kµhpn+1,1(t)
+ λhpn−1,k (t) + o(h), n ≥ 2,
pni(t + h) = (1 − λh − kµh)pni(t) + kµhpn,i+1(t)
+ λhpn−1,i(t) + o(h), n ≥ 2 and
1 ≤ i ≤ k − 1.

3
Hence,
p0(t + h) − p0(t) o(h)
= −λp0(t) + kµp1,1(t) +
h h
p1,i(t + h) − p1,i(t)
= −(λ + kµ)p1,i(t)
h
o(h)
+ kµp1,i+1(t) + ,
h
1 ≤ i ≤ k − 1,
p1,k (t + h) − p1,k (t)
= −(λ + kµ)p1,k (t)
h
o(h)
+ kµp2,1(t) + λp0(t) +
h
pnk (t + h) − pnk (t)
= −(λ + kµ)pnk (t)
h
+ kµpn+1,1(t) + λpn−1,k (t)
o(h)
+ , n ≥ 2,
h
pni(t + h) − pni(t)
= −(λ + kµ)pni(t)
h
+ kµpn,i+1(t) + λpn−1,i(t)
o(h)
+ , n ≥ 2 and
h
1 ≤ i ≤ k − 1.
4
By letting h → 0,

p′0(t) = −λp0(t) + kµp1,1(t)


p′1,i(t) = −(λ + kµ)p1,i(t) + kµp1,i+1(t),
1 ≤ i ≤ k − 1,
p′1,k (t) = −(λ + kµ)p1,k (t) + kµp2,1(t) + λp0(t)
p′nk (t) = −(λ + kµ)pnk (t) + kµpn+1,1(t)
+ λpn−1,k (t), n ≥ 2,
p′ni(t) = −(λ + kµ)pni(t) + kµpn,i+1(t)
+ λpn−1,i(t), n ≥ 2 and 1 ≤ i ≤ k − 1.
For the steady state

p0 = lim p0(t)
t→∞
p1,i = lim p1,i(t), 1 ≤ i ≤ k − 1,
t→∞
p1,k = lim p1,k (t)
t→∞
pnk = lim pnk (t), n ≥ 2,
t→∞
pni = lim pni(t), n ≥ 2 and 1 ≤ i ≤ k − 1.
t→∞
The existence of these limits implies that

5
lim p′0(t) = 0
t→∞
lim p′1,i(t) = 0, 1 ≤ i ≤ k − 1,
t→∞
lim p′1,k (t) = 0
t→∞
lim p′nk (t) = 0, n ≥ 2,
t→∞
lim p′ni(t) = 0, n ≥ 2 and 1 ≤ i ≤ k − 1,
t→∞
and consequently
λp0 = kµp1,1 (1)
(λ + kµ)p1,i = kµp1,i+1, 1 ≤ i ≤ k − 1, (2)
(λ + kµ)p1,k = kµp2,1 + λp0 (3)
(λ + kµ)pnk = kµpn+1,1 + λpn−1,k , n ≥ 2, (4)
(λ + kµ)pni = kµpn,i+1 + λpn−1,i, (5)
n ≥ 2 and 1 ≤ i ≤ k − 1
Similar to the birth-and-death process, the bal-
ance equations (1)-(5) can be interpreted as
stating that the rate at which the system leaves
each state equals the rate at which the system
enters this state. In other words, for any state
of the system
entering rate = leaving rate
6
Therefore, the balance equations (1)-(5) can
be obtained directly from the transition dia-
gram. For example, consider M |E2|1 and the
corresponding transition diagram
λ λ
λ λ λ

0 1,2 2µ 1,1 2,2 2µ 2,1

2µ 2µ

Consider state (0). There is only one arc for
which node (0) is a starting node. The rate
associated with this arc is λ. So, the leaving
rate for node (0) is λp0. There is only one
arc for which node (0) is an end node. This
arc is (1,1)-(0), and the corresponding rate is
2µ. Therefore, the entering rate for node (0)
is 2µp1,1. This gives the balance (conservation
of flow) equation
λp0 = 2µp1,1
which is a particular case of (1) with k = 2.
7
Now consider node (2,1). There are two arcs
for which node (2,1) is a starting node. These
arcs are (2,1)-(3,1) with rate λ and (2,1)-(1,2)
with rate 2µ. Hence, the leaving rate for node
(2,1) is
(λ + 2µ)p2,1
Similarly, there are only two arcs for which
node (2,1) is an end node. These arcs are
(2,2)-(2,1) with rate 2µ and (1,1)-(2,1) with
rate λ. Hence, the entering rate for node (2,1)
is
λp1,1 + 2µp2,2
The balance equation associated with node
(2,1) is

(λ + 2µ)p2,1 = λp1,1 + 2µp2,2


which is a particular case of (5) with n = 2,
k = 2 and i = 1.

8
In general,
k [
∞ ∑

]
1 i
Wq = (n − 1) + pni + 0p0
n=1 i=1 µ kµ
k [
∞ ∑

]
k(n − 1) + i
= pni
n=1 i=1 kµ
Consider the following probability generating
function ∞ ∑ k

g(z) = z k(n−1)+ipni + p0
n=1 i=1
Then, g ′(1)
Wq = . (6)

In order to find g(z), we multiply (5) by z k(n−1)+i,
(4) by z kn, (2) by z i, (3) by z k , and sum up all
resultant equations and (1). We have

(λ + kµ)g(z) = kµp0 + [g(z) − p0] + λz k g(z),
z
λ
and by dividing by kµ and denoting r = ,

1
(1 + r)g(z) = p0 + [g(z) − p0] + rz k g(z),
z
9
which gives
( )
1 k (z − 1)p0
g(z) 1 + r − − rz =
z z
(1 − z)p0
g(z) = .
1 − (1 + r)z + rz k+1

The probability p0 can be found from the con-


dition
lim g(z) = 1.
z→1
Using L’Hopital’s rule, we have
−p0 p0
lim g(z) = lim = .
z→1 z→1 −(1 + r) + r(k + 1)z k 1 − rk
Hence, p0
= 1,
1 − rk
and therefore, λ
p0 = 1 − kr = 1 − .
µ
λ
By denoting ϱ = and using L’Hopital’s rule
twice, we have µ
(k + 1)ϱ
lim g ′(z) = ,
z→1 2(1 − ϱ)
and by virtue of (6),
10
g ′(1) k+1 ϱ k+1 λ
Wq = = =
kµ 2k µ(1 − ϱ) 2k µ(µ − λ)
Then 1
W = Wq + ,
µ
and using Little’s formulae,
k+1 λ2
Lq = λWq =
2k µ(µ − λ)

k+1 λ2 λ
L = λW = +
2k µ(µ − λ) µ

M [Y ]|M |1 Model

In this model we assume that customers arrive


in groups. The arrival stream of groups form a
Poisson process with rate λ. The size of each
group is a random variable Y , where P (Y =
k) = πk . The service times are independent
and identically distributed exponential random
1
variables with mean .
µ
11
A stochastic process {Y (t), t ≥ 0} is a com-
pound Poisson process if
X(t)

Y (t) = Yi ,
i=1
where {X(t), t ≥ 0} is a Poisson process, and all
Yi are independent and identically distributed
random variables. So, in the considered model,
customers arrive according to a compound Pois-
son process. The arrival stream of groups of
size k can be considered as a Poisson process
with rate λπk .
This model is still Markovian in the sense that
future behaviour depends only on the current
state. For example, if πk = 0 for all k ≥ 3,
then the transition diagram is

λπ2 λπ2 λπ2 λπ2 λπ2


λπ1 λπ1 λπ1 λπ1 λπ 1
0 1 2 3 4
µ µ µ µ µ

12
Using our usual technique, we can derive the
following set of differential equations

p′n(t) = −(λ + µ)pn(t) + µpn+1(t)


n

+ λπk pn−k (t), n ≥ 1,
k=1
p′0(t) = −λp0(t) + µp1(t)
For the steady state probabilities
n

(λ + µ)pn = µpn+1 + λ pn−k πk , n ≥ 1,(7)
k=1
λp0 = µp1 (8)
Multiplying each equation in (7) by the corre-
sponding z n and adding all the resultant equa-
tions and (8), we obtain

∑ ∞
∑ ∞

µ
λ pn z n + µ pn z n = pn z n
n=0 n=1 z n=1
∞ ∑
∑ n
+ λ pn−k πk z n (9)
n=1 k=1

13
Let ∞

g(z) = pn z n
n=0

∑ ∞

q(z) = πnz n = πnz n
n=0 n=1
Then,
∞ ∑
∑ n
pn−k πk z n = g(z)q(z),
n=1 k=1
and (9) can be rewritten as
µ
λg(z) + µ[g(z) − p0] = [g(z) − p0] + λg(z)q(z)
z
which gives
µp0(1 − z)
g(z) =
µ(1 − z) − λz[1 − q(z)]
Now we can obtain p0 from lim g(z) = 1. By
z→1
applying L’Hopital’s rule, and taking into ac-
count that lim q(z) = 1 and lim q ′(z) = E[Y ],
z→1 z→1
−µp0
lim g(z) = lim
z→1 z→1 −µ − λ + λq(z) + λzq ′ (z)

14
−µp0
=
−µ + λE[Y ]
Hence, λE[Y ]
p0 = 1 −
µ
λE[Y ]
Let ϱ = , then p0 = 1 − ϱ. So, we need
µ
ϱ < 1. Since p0 and q(z) are known, we can
find
L = lim g ′(z).
z→1

Output Process for M |M |m Model


Consider the steady state of the M |M |m model.
Let Ws be the expected time in service and Ls
be expected number of customers in service.
We have
W = Wq + Ws
L = Lq + Ls
L = λW
Lq = λWq
15
Hence, we have another Little’s formula
Ls = λWs,
which also can be obtained directly using the
method utilized for other Little’s formulae. Since
1 λ
Ws = , we have Ls = . If at time t there are
µ µ
i customers in the system and h is sufficiently
small, then the probability that service will be
completed in (t, t + h) equals
min[i, m]µh + o(h)
So, the probability of a customer leaving in
(t, t + h) is
m−1
∑ ∞

iµhpi + mµhpi + o(h)
 i=0 i=m
m−1
∑ ∞

= ipi + m pi µh + o(h) = Lsµh + o(h)
i=0 i=m
λ
= µh + o(h) = λh + o(h)
µ
This leads to the conclusion that the departure
process is a Poisson process with rate λ. The
same result holds for M |M |1 and M |M |∞.
16
Network of Queues

The above result permits to analyze networks


of queues, where a customer may require ser-
vice at more than one service stations (service
centers) and a service station may consist of
a single server, several servers, or even an un-
limited number of servers. The service stations
are considered as nodes of the corresponding
network of queues. Customers enter the sys-
tem from outside, follow a path through one
or several service stations (without visiting any
station more than once), and then leave the
system. Customers may arrive at a node from
outside the network or/and from other nodes.
Arrivals to node i from outside the network
form a Poisson arrival stream with rate λi (if
there are no arrivals from outside world to node
i, then λi = 0). After leaving node i a cus-
tomer may go to node j with probability πij or
leave the system.
17
For example,

outside world outside world


1 − π4,5
λ4
queue 4

π4,5
π1,4
queue 2
outside world
π1,2
λ1 queue 5
queue 1 queue 3
π1,3
π1,5

In this network queue 1 is the queue of a M |M |3


queueing system; queue 2 is the queue of a
M |M |2 queueing system; queue 3 is the queue
of a M |M |1 queueing system; queue 4 is the
queue of a M |M |3 queueing system; and queue
5 is the queue of a M |M |4 queueing system.

18
So, customers arrive at the first service station
according to the Poisson process with rate λ1.
Taking into account the theorems about split-
ting and merging Poisson processes, we con-
clude that
• customers arrive at the second service sta-
tion according to the Poisson process with
rate λ1π1,2;
• customers arrive at the third service station
according to the Poisson process with rate
λ1π1,3;
• customers arrive at the fourth service sta-
tion according to the Poisson process with
rate λ1π1,4 + λ1π1,2 + λ4;
• customers arrive at the fifth service station
according to the Poisson process with rate

λ1(π1,3 + π1,5) + [λ1(π1,2 + π1,4) + λ4]π4,5.

Therefore, all service station can be analyzed


separately.
19

You might also like