### Lecture 3 Constructing and Solving Markov Processes

```Stochastic Processes
Markov Processes
Derivation of Performance Measures
Performance Modelling — Lecture 3
Constructing and Solving Markov Processes
Jane Hillston
School of Informatics
The University of Edinburgh
Scotland
23rd September 2014
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Stochastic Process
Formally, a stochastic model is one represented as a stochastic
process;
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Stochastic Process
Formally, a stochastic model is one represented as a stochastic
process;
A stochastic process is a set of random
variables{X (t), t ∈ T }.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Stochastic Process
Formally, a stochastic model is one represented as a stochastic
process;
A stochastic process is a set of random
variables{X (t), t ∈ T }.
T is called the index set usually taken to represent time.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Stochastic Process
Formally, a stochastic model is one represented as a stochastic
process;
A stochastic process is a set of random
variables{X (t), t ∈ T }.
T is called the index set usually taken to represent time.
Since we consider continuous time models T = R≥0 , the set
of non-negative real numbers.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
State Space
The state space of a stochastic process is the set of all possible
values that the random variables X (t) can assume.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
State Space
The state space of a stochastic process is the set of all possible
values that the random variables X (t) can assume.
Each of these values is called a state of the process.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
State Space
The state space of a stochastic process is the set of all possible
values that the random variables X (t) can assume.
Each of these values is called a state of the process.
Any set of instances of {X (t), t ∈ T } can be regarded as a path of
a particle moving randomly in the state space, S, its position at
time t being X (t).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
State Space
The state space of a stochastic process is the set of all possible
values that the random variables X (t) can assume.
Each of these values is called a state of the process.
Any set of instances of {X (t), t ∈ T } can be regarded as a path of
a particle moving randomly in the state space, S, its position at
time t being X (t).
These paths are called sample paths or realisations of the
stochastic process.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Properties of Stochastic Processes
In this course we will focus on stochastic processes with the
following properties:
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Properties of Stochastic Processes
In this course we will focus on stochastic processes with the
following properties:
{X (t)} is a Markov process.
This implies that {X (t)} has the Markov or memoryless property:
given the value of X (t) at some time t ∈ T , the future path X (s)
for s > t does not depend on knowledge of the past history X (u)
for u < t, i.e. for t1 < · · · < tn < tn+1 ,
Pr(X (tn+1 ) = xn+1 | X (tn ) = xn , . . . , X (t1 ) = x1 ) =
Pr(X (tn+1 ) = xn+1 | X (tn ) = xn )
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Properties of Stochastic Processes
In this course we will focus on stochastic processes with the
following properties:
{X (t)} is irreducible.
This implies that all states in S can be reached from all other
states, by following the transitions of the process. If we draw a
directed graph of the state space with a node for each state and an
arc for each event, or transition, then for any pair of nodes there is
a path connecting them, i.e. the graph is strongly connected.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Properties of Stochastic Processes
In this course we will focus on stochastic processes with the
following properties:
{X (t)} is stationary:
for any t1 , . . . tn ∈ T and t1 + τ, . . . , tn + τ ∈ T (n ≥ 1), then the
process’s joint distributions are unaffected by the change in the
time axis and so,
FX (t1 +τ )...X (tn +τ ) = FX (t1 )...X (tn )
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Properties of Stochastic Processes
In this course we will focus on stochastic processes with the
following properties:
{X (t)} is time homogeneous:
the behaviour of the system does not depend on when it is
observed. In particular, the transition rates between states are
independent of the time at which the transitions occur. Thus, for
all t and s, it follows that
Pr(X (t + τ ) = xk | X (t) = xj ) = Pr(X (s + τ ) = xk | X (s) = xj ).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Exit rate and sojourn time
In any stochastic process the time spent in a state is called the
sojourn time.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Exit rate and sojourn time
In any stochastic process the time spent in a state is called the
sojourn time.
In a Markov process the rate of leaving a state xi , qi the exit rate,
is exponentially distributed with the rate which is the sum of all
N
X
qi,j .
the individual transitions that leave the state, i.e. qi =
j=1,j6=i
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Exit rate and sojourn time
In any stochastic process the time spent in a state is called the
sojourn time.
In a Markov process the rate of leaving a state xi , qi the exit rate,
is exponentially distributed with the rate which is the sum of all
N
X
qi,j .
the individual transitions that leave the state, i.e. qi =
j=1,j6=i
This follows from the superposition principle of exponential
distributions.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Exit rate and sojourn time
In any stochastic process the time spent in a state is called the
sojourn time.
In a Markov process the rate of leaving a state xi , qi the exit rate,
is exponentially distributed with the rate which is the sum of all
N
X
qi,j .
the individual transitions that leave the state, i.e. qi =
j=1,j6=i
This follows from the superposition principle of exponential
distributions.
It follows that the sojourn time will be 1/qi .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Exit rate and sojourn time
In any stochastic process the time spent in a state is called the
sojourn time.
In a Markov process the rate of leaving a state xi , qi the exit rate,
is exponentially distributed with the rate which is the sum of all
N
X
qi,j .
the individual transitions that leave the state, i.e. qi =
j=1,j6=i
This follows from the superposition principle of exponential
distributions.
It follows that the sojourn time will be 1/qi .
Note: by the Markov property, the sojourn times are memoryless.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Transition rates and transition probabilities
At time τ , the probability that there is a state transition in the
interval (τ, τ + dt) is qi dt + o(dt).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Transition rates and transition probabilities
At time τ , the probability that there is a state transition in the
interval (τ, τ + dt) is qi dt + o(dt).
When a transition out of state xi occurs, the new state is xj with
probability pij , which must depend only on i and j (Markov).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Transition rates and transition probabilities
At time τ , the probability that there is a state transition in the
interval (τ, τ + dt) is qi dt + o(dt).
When a transition out of state xi occurs, the new state is xj with
probability pij , which must depend only on i and j (Markov).
Thus, for i 6= j, i, j ∈ S,
Pr(X (τ + dt) = j | X (τ ) = i) = qij dt + o(dt)
where the qij = qi pij , by the decomposition property.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Transition rates and transition probabilities
At time τ , the probability that there is a state transition in the
interval (τ, τ + dt) is qi dt + o(dt).
When a transition out of state xi occurs, the new state is xj with
probability pij , which must depend only on i and j (Markov).
Thus, for i 6= j, i, j ∈ S,
Pr(X (τ + dt) = j | X (τ ) = i) = qij dt + o(dt)
where the qij = qi pij , by the decomposition property.
The qij are called the instantaneous transition rates.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Transition rates and transition probabilities
At time τ , the probability that there is a state transition in the
interval (τ, τ + dt) is qi dt + o(dt).
When a transition out of state xi occurs, the new state is xj with
probability pij , which must depend only on i and j (Markov).
Thus, for i 6= j, i, j ∈ S,
Pr(X (τ + dt) = j | X (τ ) = i) = qij dt + o(dt)
where the qij = qi pij , by the decomposition property.
The qij are called the instantaneous transition rates.
The transition probability pij is the probability, given that a
transition out of state i occurs, that it is the transition to state j.
By the definition of conditional probability, this is pij = qij /qi .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Infinitesimal Generator Matrix
The state transition diagram of a Markov process captures all the
information about the states of the system and the transitions
which can occur between then.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Infinitesimal Generator Matrix
The state transition diagram of a Markov process captures all the
information about the states of the system and the transitions
which can occur between then.
We can capture this information in a matrix, Q , termed the
infinitesimal generator matrix.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Infinitesimal Generator Matrix
The state transition diagram of a Markov process captures all the
information about the states of the system and the transitions
which can occur between then.
We can capture this information in a matrix, Q , termed the
infinitesimal generator matrix.
For a state space of size N, this is a N × N matrix, where entry
q(i, j) or qi,j , records the transition rate of moving from state xi to
state xj .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Infinitesimal Generator Matrix
The state transition diagram of a Markov process captures all the
information about the states of the system and the transitions
which can occur between then.
We can capture this information in a matrix, Q , termed the
infinitesimal generator matrix.
For a state space of size N, this is a N × N matrix, where entry
q(i, j) or qi,j , records the transition rate of moving from state xi to
state xj .
By convention, the diagonal entries qi,i are the negative row sum
for row i, i.e.
N
X
qi,i = −
qi,j
j=1,j6=i
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
In performance modelling we are often interested in the probability
distribution of the random variable X (t) over the state space S, as
the system settles into a regular pattern of behaviour.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
In performance modelling we are often interested in the probability
distribution of the random variable X (t) over the state space S, as
the system settles into a regular pattern of behaviour.
This is termed the steady state probability distribution.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
In performance modelling we are often interested in the probability
distribution of the random variable X (t) over the state space S, as
the system settles into a regular pattern of behaviour.
This is termed the steady state probability distribution.
From this probability distribution we will derive performance
measures based on subsets of states where some condition holds.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Existence of a steady state probability distribution
For every time-homogeneous, finite, irreducible Markov process
with state space S, there exists a steady state probability
distribution
{πk , xk ∈ S}
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Existence of a steady state probability distribution
For every time-homogeneous, finite, irreducible Markov process
with state space S, there exists a steady state probability
distribution
{πk , xk ∈ S}
This distribution is the same as the limiting or long term
probability distribution:
πk = lim Pr(X (t) = xk | X (0) = x0 )
t→∞
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Existence of a steady state probability distribution
For every time-homogeneous, finite, irreducible Markov process
with state space S, there exists a steady state probability
distribution
{πk , xk ∈ S}
This distribution is the same as the limiting or long term
probability distribution:
πk = lim Pr(X (t) = xk | X (0) = x0 )
t→∞
This distribution is reached when the initial state no longer has any
influence.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Probability flux
In steady state, πi is the proportion of time that the process
spends in state xi .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Probability flux
In steady state, πi is the proportion of time that the process
spends in state xi .
Recall qij is the instantaneous probability that the model makes a
transition from state xi to state xj .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Probability flux
In steady state, πi is the proportion of time that the process
spends in state xi .
Recall qij is the instantaneous probability that the model makes a
transition from state xi to state xj .
Thus, in an instant of time, the probability that a transition will
occur from state xi to state xj is the probability that the model
was in state xi , πi , multiplied by the transition rate qij .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Probability flux
In steady state, πi is the proportion of time that the process
spends in state xi .
Recall qij is the instantaneous probability that the model makes a
transition from state xi to state xj .
Thus, in an instant of time, the probability that a transition will
occur from state xi to state xj is the probability that the model
was in state xi , πi , multiplied by the transition rate qij .
This is called the probability flux from state xi to state xj .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Global balance equations
In steady state, equilibrium is maintained so for any state the total
probability flux out is equal to the total probability flux into the
state.
πi ×
X
qij =
xj ∈S,j6=i
|
{z
flux out of xi
X
(πj × qji )
xj ∈S,j6=i
}
|
{z
flux into xi
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
}
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Global balance equations
In steady state, equilibrium is maintained so for any state the total
probability flux out is equal to the total probability flux into the
state.
πi ×
X
qij =
xj ∈S,j6=i
|
{z
flux out of xi
X
(πj × qji )
xj ∈S,j6=i
}
|
{z
flux into xi
}
(If this were not true the distribution over states would change. )
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Global balance equations
Recall that the diagonal elements of the infinitesimal generator
matrix Q arePthe negative sum of the other elements in the row,
i.e. qii = − xj ∈S,j6=i qij .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Global balance equations
Recall that the diagonal elements of the infinitesimal generator
matrix Q arePthe negative sum of the other elements in the row,
i.e. qii = − xj ∈S,j6=i qij .
We can use this to rearrange the flux balance equation to be:
X
πj qji = 0.
xj ∈S
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Global balance equations
Recall that the diagonal elements of the infinitesimal generator
matrix Q arePthe negative sum of the other elements in the row,
i.e. qii = − xj ∈S,j6=i qij .
We can use this to rearrange the flux balance equation to be:
X
πj qji = 0.
xj ∈S
Expressing the unknown values πi as a row vector π, we can write
this as a matrix equation:
πQ=0
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Normalising constant
The πi are unknown — they are the values we wish to find.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Normalising constant
The πi are unknown — they are the values we wish to find.
If there are N states in the state space, the global balance
equations give us N equations in N unknowns.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Normalising constant
The πi are unknown — they are the values we wish to find.
If there are N states in the state space, the global balance
equations give us N equations in N unknowns.
However this collection of equations is irreducible.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Normalising constant
The πi are unknown — they are the values we wish to find.
If there are N states in the state space, the global balance
equations give us N equations in N unknowns.
However this collection of equations is irreducible.
Fortunately, since {πi } is a probability distribution we also know
that the normalisation condition holds:
X
πi = 1
xi ∈S
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Normalising constant
The πi are unknown — they are the values we wish to find.
If there are N states in the state space, the global balance
equations give us N equations in N unknowns.
However this collection of equations is irreducible.
Fortunately, since {πi } is a probability distribution we also know
that the normalisation condition holds:
X
πi = 1
xi ∈S
With these n + 1 equations we can use standard linear algebra
techniques to solve the equations and find the n unknowns, {πi }.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Example
Consider a system with multiple CPUs, each with its own
private memory, and one common memory which can be
accessed only by one processor at a time.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example
Consider a system with multiple CPUs, each with its own
private memory, and one common memory which can be
accessed only by one processor at a time.
The CPUs execute in private memory for a random time
before issuing a common memory access request. Assume that
this random time is exponentially distributed with parameter λ
(the average time a CPU spends executing in private memory
between two common memory access requests is 1/λ).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example
Consider a system with multiple CPUs, each with its own
private memory, and one common memory which can be
accessed only by one processor at a time.
The CPUs execute in private memory for a random time
before issuing a common memory access request. Assume that
this random time is exponentially distributed with parameter λ
(the average time a CPU spends executing in private memory
between two common memory access requests is 1/λ).
The common memory access duration is also assumed to be
exponentially distributed, with parameter µ (the average
duration of a common memory access is 1/µ).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Example
If the system has only one processor, it has only two states:
1
2
The processor is executing in its private memory;
The processor is accessing common memory.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example
If the system has only one processor, it has only two states:
1
2
The processor is executing in its private memory;
The processor is accessing common memory.
The system behaviour can be modelled by a 2-state Markov
process whose state transition diagram and generator matrix are as
shown below:
λ
1
µ
6
?
2
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Q=
−λ λ
µ −µ
Stochastic Processes
Example
Markov Processes
λ
1
µ
6
Derivation of Performance Measures
?
2
Q=
−λ λ
µ −µ
Assumptions
If we consider the probability flux in and out of state 1 we obtain:
π1 λ = π2 µ. Similarly, for state 2: π2 µ = π1 λ.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Example
Markov Processes
λ
1
µ
6
Derivation of Performance Measures
?
2
Q=
−λ λ
µ −µ
Assumptions
If we consider the probability flux in and out of state 1 we obtain:
π1 λ = π2 µ. Similarly, for state 2: π2 µ = π1 λ.
We know from the normalisation condition that: π1 + π2 = 1.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Example
Markov Processes
λ
1
µ
6
Derivation of Performance Measures
?
2
Q=
−λ λ
µ −µ
Assumptions
If we consider the probability flux in and out of state 1 we obtain:
π1 λ = π2 µ. Similarly, for state 2: π2 µ = π1 λ.
We know from the normalisation condition that: π1 + π2 = 1.
µ
λ
π=
,
.
µ+λ µ+λ
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Example
Markov Processes
λ
1
µ
6
Derivation of Performance Measures
?
2
Q=
−λ λ
µ −µ
Assumptions
If we consider the probability flux in and out of state 1 we obtain:
π1 λ = π2 µ. Similarly, for state 2: π2 µ = π1 λ.
We know from the normalisation condition that: π1 + π2 = 1.
µ
λ
π=
,
.
µ+λ µ+λ
From this we can deduce, for example, that the probability that the
processor is executing in private memory is µ/(µ + λ).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Solving the global balance equations
In general our systems of equations will be too large to contemplate
solving them by hand, so we want to be able to take advantage of
linear algebra packages which can solve matrix equations of the
form Ax = b, where A is an N × N matrix, x is a column vector of
N unknowns, and b is a column vector of N values.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Solving the global balance equations
First we must resolve two problems:
1 Our global balance equation is expressed in terms of a row
vector of unknowns π, π Q = 0: the unknowns.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Solving the global balance equations
First we must resolve two problems:
1 Our global balance equation is expressed in terms of a row
vector of unknowns π, π Q = 0: the unknowns.
This problem is resolved by transposing the equation, i.e.
QT π = 0, where the right hand side is now a column vector
of zeros, rather than a row vector.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Solving the global balance equations
2 We must eliminate the redundancy in the global balance
equations and add in the normalisation condition.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Solving the global balance equations
2 We must eliminate the redundancy in the global balance
equations and add in the normalisation condition.
We replace one of the global balance equations by the
normalisation condition. In QT this corresponds to replacing
one row by a row of 1’s. We usually choose the last row and
denote the modified matrix QT
N.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Solving the global balance equations
2 We must eliminate the redundancy in the global balance
equations and add in the normalisation condition.
We replace one of the global balance equations by the
normalisation condition. In QT this corresponds to replacing
one row by a row of 1’s. We usually choose the last row and
denote the modified matrix QT
N.
We must also make the corresponding change to the
“solution” vector 0, to be a column vector with 1 in the last
row, and zeros everywhere else. We denote this vector, eN .
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Solving the global balance equations
2 We must eliminate the redundancy in the global balance
equations and add in the normalisation condition.
We replace one of the global balance equations by the
normalisation condition. In QT this corresponds to replacing
one row by a row of 1’s. We usually choose the last row and
denote the modified matrix QT
N.
We must also make the corresponding change to the
“solution” vector 0, to be a column vector with 1 in the last
row, and zeros everywhere else. We denote this vector, eN .
Now we can use any linear algebra solution package, such as
MatLab to solve the resulting equation:
QT
N π = eN
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Example
Consider the two-processor version of the multiprocessor with
processors A and B.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example
Consider the two-processor version of the multiprocessor with
processors A and B.
We assume that the processors have different timing
characteristics, the private memory access of A being governed by
an exponential distribution with parameter λA , the common
memory access of B being governed by an exponential distribution
with parameter µB , etc.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example: state space
Now the state space becomes:
1
A and B both executing in their private memories;
2
B executing in private memory, and A accessing common
memory;
3
A executing in private memory, and B accessing common
memory;
4
A accessing common memory, B waiting for common memory;
5
B accessing common memory, A waiting for common memory;
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Example: state space
µ
λB
2 1 3
λ
i
PP
1
µB
A
PP
PP
µB P P
µA
PP
P
λB
λA
PP
P
P
-
A
4
5
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example: generator matrix



Q=


−(λA + λB )
λA
λB
0
0
µA
−(µA + λB )
0
λB
0
µB
0
−(µB + λA )
0
λA
0
0
µA
−µA
0
0
µB
0
0
−µB
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes






Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example: modified generator matrix

QT
N


=


−(λA + λB )
µA
µB
0
0
λA
−(µA + λB )
0
0
µB
λB
0
−(µB + λA ) µA
0
0
λB
0
−µA 0
1
1
1
1
1
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes






Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
If we choose the following values for the parameters:
λA = 0.05
λB := 0.1
µA = 0.02
µB = 0.05
solving the matrix equation, and rounding figures to 4 significant
figures, we obtain:
π = (0.0693, 0.0990, 0.1683, 0.4951, 0.1683)
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Deriving Performance Measures
SYSTEM
STATE
TRANSITION
DIAGRAM
Q=
..... .....
.....
=
EQUILIBRIUM PROBABILITY
p , p , p , ..... DISTRIBUTION ..... , pN
1
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
.....
.....
.....
.....
.....
MARKOV
PROCESS
2
3
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Deriving Performance Measures
SYSTEM
STATE
TRANSITION
DIAGRAM
.....
.....
.....
.....
.....
MARKOV
PROCESS
Q=
..... .....
.....
=
EQUILIBRIUM PROBABILITY
p , p , p , ..... DISTRIBUTION ..... , pN
1
2
3
PERFORMANCE MEASURES
e.g. throughput, response time, utilisation
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Deriving Performance Measures
Broadly speaking, there are three ways in which performance
measures can be derived from the steady state distribution of a
Markov process.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Deriving Performance Measures
Broadly speaking, there are three ways in which performance
measures can be derived from the steady state distribution of a
Markov process.
These different methods can be thought of as corresponding to
different types of measure:
state-based measures, e.g. utilisation;
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Deriving Performance Measures
Broadly speaking, there are three ways in which performance
measures can be derived from the steady state distribution of a
Markov process.
These different methods can be thought of as corresponding to
different types of measure:
state-based measures, e.g. utilisation;
rate-based measures, e.g. throughput;
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Deriving Performance Measures
Broadly speaking, there are three ways in which performance
measures can be derived from the steady state distribution of a
Markov process.
These different methods can be thought of as corresponding to
different types of measure:
state-based measures, e.g. utilisation;
rate-based measures, e.g. throughput;
other measures which fall outside the above categories, e.g.
response time.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
State-based measures
State-based measures correspond to the probability that the model
is in a state, or a subset of states, which satisfy some condition.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
State-based measures
State-based measures correspond to the probability that the model
is in a state, or a subset of states, which satisfy some condition.
For example, utilisation will correspond to those states where a
resource is in use.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
State-based measures
State-based measures correspond to the probability that the model
is in a state, or a subset of states, which satisfy some condition.
For example, utilisation will correspond to those states where a
resource is in use.
If we consider the multiprocessor example, the utilisation of the
common memory, Umem , is the total probability that the model is
in one of the states in which the common memory is in use:
Umem = π2 + π3 + π4 + π5 = 93.07%
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
State-based measures
Other examples of state-based measures are idle time, or the
number of jobs in a system.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
State-based measures
Other examples of state-based measures are idle time, or the
number of jobs in a system.
Some measures such as the number of jobs will involve a weighted
sum of steady state probabilities, weighted by the appropriate value
(expectation).
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
State-based measures
Other examples of state-based measures are idle time, or the
number of jobs in a system.
Some measures such as the number of jobs will involve a weighted
sum of steady state probabilities, weighted by the appropriate value
(expectation).
For example, if we consider jobs waiting for the common memory
to be queued in that subsystem, then the average number of jobs
in the common memory, Nmem , is:
Nmem = (1 × π2 ) + (1 × π3 ) + (2 × π4 ) + (2 × π5 ) = 1.594
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Rate-based measures
Rate-based measures are those which correspond to the predicted
rate at which some event occurs.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Rate-based measures
Rate-based measures are those which correspond to the predicted
rate at which some event occurs.
This will be the product of the rate of the event, and the
probability that the event is enabled, i.e. the probability of being in
one of the states from which the event can occur.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example: rate-based measures
In order to calculate the throughput of the common memory, we
need the average number of accesses from either processor which it
satisfies in unit time.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Example: rate-based measures
In order to calculate the throughput of the common memory, we
need the average number of accesses from either processor which it
satisfies in unit time.
Xmem is thus calculated as:
Xmem = (µA × (π2 + π4 )) + (µB × (π3 + π5 )) = 0.0287
or, approximately one access every 35 milliseconds.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Other measures
The other measures are those which are neither rate-based or
state-based.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Other measures
The other measures are those which are neither rate-based or
state-based.
In these cases, we usually use one of the operational laws to derive
the information we need, based on values that we have obtained
from solution of the model.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Other measures
The other measures are those which are neither rate-based or
state-based.
In these cases, we usually use one of the operational laws to derive
the information we need, based on values that we have obtained
from solution of the model.
For example, applying Little’s Law to the common memory we see
that
Wmem = Nmem /Xmem = 1.594/0.0287 = 55.54 milliseconds
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Stochastic Hypothesis
“The behaviour of a real system during a given period of
time is characterised by the probability distributions of a
stochastic process.”
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Assumptions
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Assumptions
All delays and inter-event times are exponentially distributed.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Assumptions
All delays and inter-event times are exponentially distributed.
(This will often not fit with observations of real systems.)
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Assumptions
All delays and inter-event times are exponentially distributed.
(This will often not fit with observations of real systems.)
We make the assumption because of the nice mathematical
properties of the exponential distribution, and because it is
the only distribution giving us a Markov process.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Assumptions
All delays and inter-event times are exponentially distributed.
(This will often not fit with observations of real systems.)
We make the assumption because of the nice mathematical
properties of the exponential distribution, and because it is
the only distribution giving us a Markov process.
Plus only a single parameter to be fitted (the rate), which can
be easily derived from observations of the average duration.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Assumptions
All delays and inter-event times are exponentially distributed.
(This will often not fit with observations of real systems.)
We make the assumption because of the nice mathematical
properties of the exponential distribution, and because it is
the only distribution giving us a Markov process.
Plus only a single parameter to be fitted (the rate), which can
be easily derived from observations of the average duration.
The Markov/memoryless assumption — future behaviour is
only dependent on the current state, not on the past history
— is a reasonable assumption for computer and
communication systems, if we choose our states carefully.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Assumptions
All delays and inter-event times are exponentially distributed.
(This will often not fit with observations of real systems.)
We make the assumption because of the nice mathematical
properties of the exponential distribution, and because it is
the only distribution giving us a Markov process.
Plus only a single parameter to be fitted (the rate), which can
be easily derived from observations of the average duration.
The Markov/memoryless assumption — future behaviour is
only dependent on the current state, not on the past history
— is a reasonable assumption for computer and
communication systems, if we choose our states carefully.
We generally assume that the Markov process is finite, time
homogeneous and irreducible.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
Stochastic Processes
Markov Processes
Derivation of Performance Measures
Assumptions
Exercise
Consider the multiprocessor example, but with three
processors, A, B and C sharing the common memory instead
of two.
List the states of the system, and draw the state transition
diagram for this case.
What is the difficulty in doing this and what further
information do you need?
Solution will be presented at the beginning of the next lecture.
Jane Hillston School of Informatics The University of Edinburgh Scotland
Performance Modelling — Lecture 3 Constructing and Solving Markov Processes
```