ANALYSIS OF HM-NETWORKS WITH STOCHASTIC INCOMES FROM TRANSITIONS BETWEEN STATES
Michal Matalycki1, Katarzyna Koluzaewa2
1 Institute of Mathematics, Czestochowa University of Technology, Poland
2 Faculty of Mathematics and Computer Science, Grodno State University koluzaeva@gmail.com
Abstract. In article method of finding of expected incomes in systems of HM-network of arbitrary topology when incomes from transitions between network states are stochastic variables with given mean values is proposed. For expected incomes the system of linear non-homogeneous ordinary differential equations was obtained, to solve it we can find in- comes in network systems.
Introduction
Markov queueing networks (QN) with incomes were examined in works [1-3]
at the first time. They are described with help of Markov chains with continuous time and incomes which were introduced by R. Howard [4]. So at the recent time they are called HM(Howard-Matalytski)-networks [5, 6]. Before investigation of closed networks with account number of states was carry out. Herewith the next cases were examined: a) incomes from transitions between network states depend on states and time or b) incomes are stochastic variables (SV) with known finite moments of the first and the second orders.
For expected incomes of network systems in case a) the system of difference- -differential equations which are reduced to the system of linear non-homogeneous ordinary differential equations (ODE) can be obtained. For their solution different methods - method of multidimensional z-transformations and known methods: me- thod of Laplace transformation, matrix method, numerical methods were proposed [7-9].
In [5, 10] approximate relations for expected incomes and income variations in systems of exponential HM-networks in case b) were obtained. Technique of re- ceiving of these relations is based on interval partition of the network functioning by big number m of small intervals of size ∆t, income estimation on every interval and summing of these incomes by means of passage to the limit m→∞, ∆t→0.
Herewith mean value of messages in network system in unsteady condition was
found with help of developed recurrence by time moments method. Notice that
different methods of analysis and optimization of Markov HM-networks and their
application were described in [11]. In the present work method of finding of expec- ted incomes of network systems in case b) which is based on solution of the linear non-homogeneous ODE, which were received for expected incomes and mean values of messages, is proposed.
1. Finding of expected incomes in systems
Let us examine open exponential QN of arbitrary topology with one type mes- sages which consists of n queueing systems (QS) S
1, S
2,…,S
nwith m
iservice chan- nels in system S ,
ii = 1 , n . Denote via k ( t ) = ( k
1( t ), k
2( t ), ..., k
n( t ) ) - vector of net- work states, where k
i(t ) - message number in system S
i(in queue and service) at the moment t. Poisson flow of messages of the rate λ enters the network. Service rate of messages at the moment t µ
i(k
i(t)) in system S
idepends on message number in this system, i = 1 , n . Message when transiting from one QS to another brings to the last system some stochastic income and income of the first system reduces by this value correspondingly.
Let us consider dynamics of income changes of some network system S
i. De- note it’s income at the moment t as V
i(t). Let at the initial moment income of sys- tem equals V
i(0) = v
i0. Income of this QS at the moment t + ∆t can be presented as
) , ( ) ( )
( t t V t V t t
V
i+ ∆ =
i+ ∆
i∆ (1)
where ∆V
i(t,∆t) - income change of system S
ion time interval [t, t+∆t). For finding of this value we write probabilities of events which can appear during time
∆tand changes of incomes of system S
iwhich are connected with these events.
1. Message from the outside with probability λ
p0i∆t +o(∆t)will enter to system S
iand will bring to it income of size r
0i, where r
0i- RV with mathematical expectation (m.e.) M { } r
0i= a
0i, p
0i- probability of message enter from outside to the system S
i, i = 1 , n .
2. Message from the system S
iwith probability µ
i( k
i( t )) u ( k
i( t )) p
i0∆ t + o ( ∆ t ) will pass to the outside and income of the system S
iwill decrease by value R
i0, where R
i0- RV with m.e. M { } R
i0= b
i0, p
i0- probability of message leaving from system S
ito the outside, i = 1 , n ,
≤
= >
, 0 , 0
, 0 , ) 1
( x
x x
u - Heavyside function.
3. Message from the system S
jwith probability µ
j(kj(t))u(kj(t))pji∆t+o(∆t)will
pass to the system S
iand income of the system S
iwill increase by value r
jiand
income of the system S
jwill decrease by this value, where r
ji- RV with m.e.
{ } rji a
ji
M = , p
ji- probability of message transition from the system S
jto the system S
i, i , j = 1 , n , i ≠ j .
4. Message from the system S
iwith probability µ
i( k
i( t )) u ( k
i( t )) p
ij∆ t + o ( ∆ t ) will pass to the system S
jand income of the QS S
iwill decrease by value R
ijand in- come of the system S
jincrease by this value, where R
ij- RV with m.e.
{ } Rij b
ij
M = , i , j = 1 , n , i ≠ . j
5. State changes of system S
ion the time interval [t,t+∆t) with probability
∆ +
∆
+
+
− ∑
≠
=
) ( ))
( ( )) ( ( ))
( ( )) ( ( 1
1
0
k t u k t k t u k t p t o t
p
n
i j j
ji j j
j i
i i
i
µ µ
λ will not appear,
n i = 1 , .
Besides that during each small time interval t ∆ the system S
iincreases its in- come by value r
i∆ t , where r
i- RV with m.e. M { r
i} = c
i, i = 1 , n . Let also sup- pose that RV r
ji, R
ij, r
0i, R
i0are independent with respect to RV r
i, i , j = 1 , n .
Evidently that r
ji= R
jiwith probability 1, i.e.
n j i b
a
ji=
ji, , = 1 , (2)
Then from said follows
∆ +
∆
+
+ +
∆ −
∆ +
∆
∆ +
−
∆ +
∆
∆ +
∆ +
∆
∆ +
−
∆ +
∆
∆ +
=
∆
∆
∑
≠
=
) ( ))
( ( )) ( (
)) ( ( )) ( ( 1
y probabilit with
) ( ))
( ( )) ( ( y
probabilit with
) ( ))
( ( )) ( ( y
probabilit with
) ( ))
( ( )) ( ( y
probabilit with
) ( y
probabilit with
) , (
1 0
0 0
0 0
t o t p t k u t k
t k u t k t p
r
t o t p t k u t k t
r R
t o t p t k u t k t
r r
t o t p t k u t k t
r R
t o t p t
r r
t t V
n
i j j
ji j j
j
i i i i i
ij i i i i
ij
ji j j
j i
ji
i i i i i
i
i i
i
i
µ
µ λ µ µ
µ λ
, , , ,
(3)
Under fixed realization of the process k (t ) and taking to account (3) it can be
written:
{ }
) ( ))
( ( )) ( (
)) ( ( )) ( ( )
( / ) , (
1
1 0 0 0
0
t o t a p t k u t k
b p b
p t k u t k c a p t
k t t V M
n
i j j
ji ji j j
j
n
i j j
ij ij i
i i i i i i i i
∆ +
∆
+
+
+
− +
=
∆
∆
∑
∑
≠
=
≠
=
µ
µ λ
Average the last relation by
k(t)with taking to account normalization condition
( ( ) = ) = 1
∑
k
k t k
P for income change of system S
iwe will obtain
{ } ( ) { }
( )
{ }
( )
( ( ) ) ( ( )) ( ( )) ( ) )) ( ( )) ( ( ) (
)) ( ..., ), ( ), ( ( ) ( / ) , (
)) ( ..., ), ( ), ( ( ) ( ...
) ( / ) , ( )
( )
, (
1
1 0 0 0
0
2 1
0 0 0
2 1
1 2
t o t t k u t k k t k P a p
t k u t k k t k P b p b
p c a p
t k t k t k t k t t V M
t k t k t k t k P
t k t t V M k t k P t
t V M
n
i j j
j j
j k
ji ji
i i i k
n
i j j
ij ij i
i i i i
n i
k k k
n k
i i
n
∆ +
∆
= +
+
=
+
− +
=
=
=
∆
∆
×
×
=
=
=
∆
∆
=
=
∆
∆
∑ ∑
∑
∑
∑ ∑ ∑
∑
≠
=
≠
=
∞
=
∞
=
∞
=
µ
µ λ
Let the system S
icontains m
iidentical service channels, time of message ser- vice in every channel is distributed under exponential law with parameter µ
i, i = 1 n , . In this case
>
= ≤
, ) ( ,
, ) ( ), )) (
( (
i i
i i
i i i i i
i
m k t m
m t k t t k
k µ
µ µ µ
i( k
i( t )) u ( k
i( t )) = µ
imin( k
i( t ), m
i) , i = 1 , n
Let us suppose that averaging of expression µ
i( k
i( t )) u ( k
i( t )) brings )
), (
min(
i ii
N t m
µ , i.e.
) ), ( min(
) ), (
min( k
it m
iN
it m
iM = (4)
where N
i(t ) - average number of messages (waiting and serving) in the system S
iat the moment , t i = 1 , n . Taking in account this assumption we obtain the next approximate relation
{ }
) ( )
), ( min(
) ), ( min(
) , (
1
1 0 0 0
0
t o t a p m t N
b p b
p m t N c
a p t
t V M
n
i j j
ji ji j j j
n
i j j
ij ij i
i i i i
i i i i
∆ +
∆
+
+
+
− +
=
∆
∆
∑
∑
≠
=
≠
=
µ
µ λ
(5)
Since Poisson flow of messages with rate λ enters the network, i.e. probability of entering of l messages in the system S
iduring time ∆ has the appearance t
( )
p tl i
l
e
il t t p
P ∆
− ∆=
∆
0) !
( λ
0 λ, l = 0 , 1 , 2 , ... , so average number of messages which en- tered to the system S
ifrom the outside during time t ∆ equals λ p
0i∆ t . Denote the average number of busy service channels in system S
iat the moment t as ρ
i(t ) ,
n
i = 1 , . Then µ
iρ
i( t ) ∆ t - average number of messages that left the system S
idur- ing time ∆ t and ∑
≠
=
∆
n
i j j
ji j
j
t p t
1
) ρ (
µ - average number of messages that entered in the
system S
ifrom another QS during time ∆ . So t
t t t
p t t
p t N t t
N
i in
i j j
ji j j i
i
i
+ ∆ − = ∆ + ∑ ∆ − ∆
≠
=
) ( )
( )
( ) (
1
0
µ ρ µ ρ
λ , i = 1 , n
whence with ∆t → 0 system of ODE for N
i(t ) follows:
i i
i n
i j j
ji j j
i
t p t p
dt t dN
0 1
) ( )
) (
( = ∑ µ ρ − µ ρ + λ
≠
=
, i = 1 , n (6)
It is impossible to find variable ρ
i(t ) exactly so as we did earlier we approximate it by expression
=
>
= ≤ min( ( ), )
, ) ( ,
, ) ( ), ) (
(
i ii i i
i i i
i
N t m
m t N m
m t N t t N
ρ
Then system of equations (6) will take on form
i i i i
n
i j j
j j ji
j
i
p N t m N t m p
dt t dN
0 1
) ), ( min(
) ), ( ) min(
( = ∑ µ − µ + λ
≠
=
, i = 1 , n (7)
This system is system of linear ODE with discontinuous right parts. It’s neces- sary to solve it by means of segmentation of the phase space by set of the areas and finding solution in each of them. System (7) can be solved for example by using means of system of computer mathematics Maple 8.
Let us introduce denotation v
i( t ) = M { V
i( t )} , i = 1 , n . From (1), (5) we obtain
) ( )
), ( min(
) ), ( min(
) (
)}
, ( { ) ( ) (
1
1 0 0 0
0
t o t a p m t N
b p b
p m t N c
a p t v
t t V M t v t t v
n
j
ji ji j j j
n
i j j
ij ij i
i i i i
i i i i
i i
i
∆ +
∆
+
+
+
− + +
=
=
∆
∆ +
=
∆ +
∑
∑
=
≠
=
µ
µ λ
Then pass to limit with ∆t → 0 we receive non-homogeneous linear ODE of the first order
n i c a p a p m t N
b p b
p m t dt N
t dv
i n
i j j
i i ji ji j j j
n
i j j
ij ij i
i i i i
i
, 1 , )
), ( min(
) ), ( ) min(
(
1
0 0
1 0 0
= + +
+
+
+
−
=
∑
∑
≠
=
≠
=
λ µ
µ
(8)
Specify initial conditions v
i( 0 ) = v
i0, i = 1 , n , it is possible to find expected incomes of network systems.
If the network is functioning so that queues are not observed in it upon the average, i.e. min( N
i( t ), m
i) = N
i( t ) , i = 1 , n , then systems (7), (8) will take the appearance:
i i
i n
i j j
j ji j
i
p N t N t p
dt t dN
0 1
) ( )
) (
( = ∑ µ − µ + λ
≠
=
, i = 1 , n (9)
=
=
+ +
+
+
−
= ∑ ∑
≠
=
=
n i v v
c a p t N a p t
N b p b
dt p t dv
i i
i n
i j j
i i j
ji ji j i
n
j ij ij i
i i i
, 1 , ) 0 (
, )
( )
) ( (
0
1
0 0 1
0
0
µ λ
µ
(10)
System (9) can be rewrote in matrix form f t dt QN
t
dN ( ) = ( ) +
(11) where N
T( t ) = ( N
1( t ), N
2( t ),..., N
n( t )) , Q - quadratic matrix which consists of ele- ments q
ij= µ
jp
ji, if suppose p
ii= − 1 , i , j = 1 , n , f - column vector with elements
p
0iλ , i = 1 n , . Solution of the system (11) is
∫
−+
=
t t Q
Qt
f e d
e N t N
0 )
)
(0 ( )
(
ττ
where N ( 0 ) - some given initial conditions, but finding of elements of matrix e
Qtis difficult problem even for rather small values of n.
Fig. 1. Structure of the network with central QS
Let us consider closed network with central QS that consists of n systems (Fig. 1).
Let suppose that queues are not observed in peripheral systems of the network under the average, i.e. min( N
i( t ), m
i) = N
i( t ) , i = n 1 , − 1 , and central QS functions in the condition of heavy traffic, i.e. min( N
n( t ), m
n) = m
n. System (7) in this case will re- write as
−
=
−
= +
−
=
∑
−=
n n n
i i i n
ni n n i i i
m t dt N
t dN
n i p m t dt N
t dN
µ µ
µ µ
1
1
) ) (
(
, 1 , 1 , )
) ( (
(12)
S1 S2 Sn
…
SnGeneral solution of the system (12) with initial conditions N
i( 0 ), i = 1 , n , equals
∑
−=
−
−
+
−
=
−
= +
=
1
1
) (
, 1 , 1 , )
(
n
i i
ni n t n i n
i ni n t n i i
p e m
g K
t N
n p i
e m g t N
i i
µ µ µ
µ
µ µ
where
i ni n n i
i
p N m
g µ
− µ
= ( 0 ) , ∑
=
=
n
i i
t N K
1
)
( - number of messages in the network.
For such N
i( t ), i = 1 , n , system (10) for expected incomes of the network systems will take the appearance
n j
nj n t n j n
j jn j
n
i i
ni n t n i n
j nj nj n n
i n
j j
nj n t n j ni
ni n
i ni n t n i in i i
p c e m
g a
p K e m
g b
dt p t dv
n i p c
e m g K
a p
p e m
g dt b
t dv
j i j
i
+
+
+
+
−
+
=
−
=
+
+
− +
+
+
−
=
− −
=
−
=
−
−
=
−
=
−
−
∑
∑
∑
∑
µ µ µ
µ µ µ
µ µ µ
µ µ µ
µ µ µ
µ
1
1
1
1 1
1 1
1
) (
1 , 1 , )
(
Integrate given ODE with initial conditions v
i( 0 ) = v
i0, i = 1 , n , we will obtain
1 , 1 , )
(
0 1
1
1
1 1
1
−
= +
−
−
+
+
−
− +
+ +
=
∑
∑
∑
−
=
−
=
− −
=
−
n i v g g b a p
t c b p m a m Ka
p
e g b g e
a p t v
i i in n
j j
j ni ni n
i in n n
j j
nj ni n n ni ni n
t i in n
j
t j j ni ni n
i i
j
µ µ
µ µ µ
µ µ
µ µ(13)
0 1
1 1
1 1
1
1
1 1
1 1
1 1
1
1
1 1
1 1
1
) (
n n
j j jn n
j j
j n
i ni ni n
n n
j
jn nj n n
j nj nj n
i i
ni n
j nj nj n n n
n t j
j jn t
n
i i
i n
j nj nj n n
v g g a
b p
t c a p m b p p K
b p m
e g a g e
b p t
v
i j+ +
+
+
+
+
− +
+
−
−
=
∑
∑
∑
∑
∑
∑
∑
∑
∑
∑
−
=
−
=
−
=
−
=
−
=
−
=
−
=
− −
=
− −
=
−
=
µ µ µ µ µ
µ µ
µ µ(14)
2. Numerical example
Let us examine the network which was described in the previous section with
= 30
n , K = 61 , where K - the number of messages in the network. Service rates of messages in channels of network systems equal: µ
1= µ
13= µ
19= µ
29= 4 , µ
6= 6 ,
,
16
5
10 9
3
= µ = µ = µ =
µ µ
4= µ
8= µ
12= µ
14= µ
21= µ
24= µ
26= µ
28= 2 , µ
25= 1 , ,
27
= 7
µ µ
2= µ
5= µ
7= µ
11= µ
15= µ
17= µ
18= µ
20= µ
22= µ
23= 3 , µ
30= 40 , channel number in the central system − m
30= 2 , probabilities of message transitions between network QS – p
30i= 1 29 , p
i30= 1 , i = 1 , 29 , define also p
ii= − 1 , i = 1 , 30 , the rest of probabilities p
ij= 0 , i , j = 1 , 30 . Let also N
i( 0 ) = 2 , i = 1 , 29 , N
30( 0 ) = 3 . Charts of average number of messages in the QS are shown in Figures 2-5.
Fig. 2. Average number of messages in network systems Si,i=1,10
Fig. 3. Average number of messages in network systems Si, i=11,20
Fig. 4. Average number of messages in network systemsSi, i=21,29
Fig. 5. Average number of messages in central QS
Specify values of m.e. of incomes from transitions between network states:
29 , 1 , exp 100 ,
sin 50
30 , 1 2 , cos
20
30
30
=
=
=
=
−
=
n i a i
n a i
n i c i
i i
i
π
π π
Then using relation (13), (14) with initial condition v
i( 0 ) = 100 , i = 1 , 29 , 150
) 0
30
( =
v , the expressions for expected incomes of network systems were obtained. For example expression for expected income of central system is
1 . 15870 2
. 28654 1
. 93 7 . 823
7 . 149 9
. 2354 1
. 720 5
. 2970 4
. 795 ) (
7
6 2
5 3
4 30
+
−
−
−
− +
+ +
+
=
−
−
−
−
−
−
−
t e
e
e e
e e
e t
v
t t
t t
t t
t
Charts of expected incomes of the network systems are shown in Figures 6-9.
Fig. 6. Expected incomes of the systemsSi, i=1,10
Fig. 7. Expected incomes of the systems Si,i=11,20
Fig. 8. Expected incomes of the systemsSi,i=20,29
Fig. 9. Expected incomes of the central QS
References
[1] Matalytski M.A., Pankov A.V. Application of operation calculus for the investigation of banking models, Proc. of 17 Int. Conf, Modern Mathematical Methods of Analysis and Optimization of Telecommunication Networks, BSU, Minsk 2003, 172-177.
[2] Matalytski M., Pankov A., Incomes probabilistic models of the banking network, Scientific Re- search of the Institute of Mathematics and Computer Science, Czestochowa, University of Tech- nology 2003, 1(2), 99-104.
[3] Matalytski M., Pankov A., Probability analysis of incomes in banking networks, Westnik BSU, 2004, 2, 86-91.
[4] Howard R., Dynamic programming and Markov processes, Moscow, 1964 (in Russian).
[5] Matalytski M., Koluzaeva E. Analysis and optimization of Markov HM-networks with stochastic incomes from transition between their states, Scientific Research of the Institute of Mathemat- ics and Computer Science, Czestochowa University of Technology 2008, 1(7), 51-62.
[6] Matalytski M.A., Koluzaeva E.V., About methods of analysis and applications of queueing HM-networks, Review of Applied and Industrial Mathematics 2008, 3(15), 564-566 (in Russian).
[7] Matalytski M., Pankov A., Analysis of the stochastic model of the changing of incomes in the open banking network, Computer Science 2003, 5, 19-29.
[8] Matalytski M., Pankov A., Numerical analysis of the probabilistic model of incomes changing in banking network, Computer Science 2004, 6, 7-11.
[9] Matalytski M., Koluzaeva K., Finding of expected incomes in open exponential networks of arbitrary architecture, Scientific Research of the Institute of Mathematics and Computer Science, Czestochowa University of Technology 2007, 6, 179-190.
[10] Koluzaeva E.V., Matalytski M.A., Analysis of incomes in open HM-networks of arbitrary archi- tecture, Westnik GrSU, 2008, 1, 22-29 (in Russian).
[11] Matalytski M.A., About some results of analysis and optimization of Markov networks with incomes and their application, Automatics and Remote Control 2009, 10 (in Russian).