Bounds for point and steady state availability

Description

Niveau: Secondaire, Lycée, Terminale
Bounds for point and steady-state availability: an algorithmi approa h based on lumpability and sto hasti ordering A. Busi and J.M. Fourneau PRiSM, Universite de Versailles-Saint-Quentin, 45, Av. des Etats-Unis 78000 Versailles, Fran e Abstra t. Markov hains and rewards have been widely used to eval- uate performan e, dependability and performability hara teristi s of omputer systems and networks. Despite onsiderable works, the nu- meri al analysis of Markov hains to obtain transient or steady-state distribution is still a diÆ ult problem when the hain is large or the eigenvalues badly distributed. Thus bounding te hniques have been pro- posed for long to analyze steady-state distribution. Here, we show how to bound some dependability hara teristi s su h as steady-state and point availability using an algorithmi approa h. The bound is based on sto hasti omparison of Markov hains but it does not use sample-path arguments. The algorithm builds a lumped Markov hain whose steady-state or transient distributions are upper bounds in the strong sto hasti sense of the exa t distributions. In this paper, the implementation of algorithm is detailed and we show some numeri al results. We also show how we an avoid the generation of the state spa e and the transition matrix to model hains with more than 10 10 states.

• has been


• matrix whi


• between up


• sto hasti


• state distribution


• markov hains


• transition matrix


• theoreti al

Sujets

• Première
• United States
• Fourneau
• Has Been
• Stochastic matrix

Informations

• Publié par: profil-zyan-2012
• Nombre de lectures: 12
• Langue: English

Extrait

Bounds
for
p
oin
the
of
of
t
solution
and
hains
steady-state
mo
a

v
do
ailabilit
0
y:
CI
an
ailabilit

b

an
h
dis-
based
is
on
op
lumpabilit
the
y
8
and
w
sto
Sure-P

del
hastic
denition
ordering
is
A.
remaining
Bu
useful

v
si
y



=
and
to
J.M.
b
F
do
ourneau
t
PRiSM,
in
Univ
=
ersit
to

10
e
orted
de

V
tro
ersailles-Sain

t-Quen
y
tin,
ecoming
45,
large-scale
Av.
lev
des
and
Etats-Unis
memory
78000
that
V

ersailles,

F
tin
rance
whose

is
Mark
there
o
the
v
h

of
hains
v
and
separating
rew

ards
state
ha
the
v
state;
e
U
b
reliabilit
een
the
widely
a
used
b
to
R
ev
X
al-
;
uate
del
p
more

states.
dep
is
endabilit
y
y

and
pro
p
1
erformabilit
The
y
o

to

system
of
a

is
systems

and
generation
net
o
w
from
orks.
sp
Despite
ely

t
w
time
orks,
ts.
the
y
n
actually
u-
o

and
analysis

of
it.
Mark
nite
o
Mark
v
hain


hains
rate
to
en
obtain
.
transien
the
t

or
o
steady-state
X
distribution
the
is
the
still

a
An
Æ
y
problem
b
when
states
the
w

states
hain
states.
is
said
large
up
or
tem
the
for
eigen
it
v
.
alues
the
badly
states.
distributed.
at
Th
dened
us
y
b
has
ounding
b

up
hniques
w
ha
t
v
t
e
r
b
2
een
2
pro-
[)
p
mo
osed

for
with
long
than
to
10
analyze
This
steady-state
ork
distribution.
supp
Here,
b
w
A
e
S
sho
ecurit
w
e,
ho

w
aths
to
In
b

ound
use
some
Mark
dep
v
endabilit
hains
y
mo



reliabilit

and
h
v
as
y
steady-state
b
and

p
The
oin
and
t
of
a
Mark
v
v
ailabilit
dels
y
high
using
el
an


relativ

easy
h.

The
in
b
oth
ound
and
is
requiremen
based
The
on
Æ
sto
is

of
hastic
solving

Mark
of
v
Mark
hain
o
deriving
v
p


hains
from
but
Consider
it

do

es
uous-time
not
o
use

sample-path
X
argumen
sto
ts.
has-
The
transition
algorithm
matrix
builds
giv
a
b
lump
Q
ed
Then
Mark
exists
o
steady-state
v
tribution

of
hain
Mark
whose
v
steady-state
hain
or
whic
transien
is
t
unique
distributions
to
are
system
upp
equations
er
Q
b
0.
ounds
a
in
ailabilit
the
measure
strong
dened
sto
y

the
hastic
in
sense
t
of
o
the
up
exact
and
distributions.
wn
In
A
this
is
pap
to
er,
e
the
if
implemen
sys-
tation
is
of
erational
algorithm
that
is
otherwise
detailed
is
and
wn
w
Let
e
denote
sho
set
w
up
some
The
n
y
umerical
time
results.
is
W
as
e
probabilit
also
that
sho
system
w
alw
ho
ys
w
een
w
the
e
states

et
a
een
v
and
oid
:
the
(
generation
)
of
P
the
(
state
s
space
U;
and
s
the
[0
transition
t
matrix
:The
p
oin
(
Note
prop
t
these
a
ha
v
.
ailabilit
b
y
e
is
v
the

probabilit
P
y
uniformization:
that
Q
the
t
system
them
is
y
op
w
erational
atten
at
.
time
distributed.
t
een
:
b
P
)
AV
(
(
P
t
w
)
Th
=
matrix
P
is
r
or
(
pro
X
b
t
b
2

U
v
)
y
and
to
the
of
steady-state
is
a
steady-state
v

ailabilit
Rubino
y
:
is
function
the
P
limit,

if
t
it
n
exists,
n
of
:
this
the
probabilit
P
y
v
.
m
It
probabilit

for
also
h
b
p
e
e
dened
distribution.
as
metho
the
tee
exp
h
ectation
of
of
are
a
uous
rew
to
ard
discrete
on
Bounding
the
ys
steady
erfor-
state
the
distribution
solv
of
b
X
the
:
of
A
when
=
eigen
P
h
i
y
j
Mun
i
metho
is
b
up
[10].

U
(
1
i
the
).
set
The
e
usual
(
w
a
a
based
y
AV
to


)
these
e
quan
)
tities

is
1
based
P
on
v
the
steady-state
uniformization

metho

d.
the
Let
y
Æ
w
b

e
steady
an
distribution
arbitrary
.
p
y
ositiv
are
e
that
v
ev
alue
to
and
to

steady-
=
transien
max
us
i
ust
f
whic
Q
a
(
these
i;
and
i
not
)
the
g
state
+
w
Æ
terested
.

Let
CTMC,
us
er,
build
uniformization
the
e
uniformized
Mark
v
hain
ersion
ds
of
alw
Q
ed
b
in
y:
or
P
aluation.
=
w
(
e
I
are
d

Q=
analyzed
).
or
P
umerical
is
steady-state
a
o
discrete
is
time

Mark
or
o
alues
v
main

b
hain.
v
Let
b
us
b
denote
and
b
9].
y
has
P
v
U

the
Mahev
blo
0

n
k
1
of
Let
P
U
asso
e


to
of
transitions
U
b
W
et
get
w
AV
een
t
up
after
states
similar
and

let
on

P
0
(
b
)
e
N
the
t;
initial
X
distribution
=0
of
t
X
t
.
n
Using
!
uniformization,
0
w
n
e
U

As

and
R
ha
(
e
t
same
)
distribution,
b
e
y:
use
R
to
(
or
t
ound
)
a
=
ailabilit
1
.
X
us,
n
e
=0
ust
e
transien
t
and
(
state
t
y
)
for
n
P
n
But
!
man

problems
0
matrices
P
so
n
uge
U
this
1
not
and,
en
b
ossible
ecause
build
of
or
the
solv
prop
the
erties
state
of
the
the
t
exp
Th
onen
w
tial
m
function,
use
the
ds
summation
h

vide
b
guaran
e
on

reliabilit
W
measures
e
whic
rst
are

limited
N
y
(
size
t;
the

space.
)
that
whic
e
h
in
is
in
the
ounding
minimal
tin
v
time
alue
ho
of
ev
n
due

the
h
pro
that
w
P

n
time
n
o
=0

e
(DTMC).
t
meth