J. Virtamo 38.

3143 Queueing Theory / Loss system 1

Loss system
Consider a loss system, where the following parameters are given


 n = number of trunks (elements which are reserved)
 a = the intensity of the offered traffic


There are no waiting places in the system. Calls which upon arrival find all trunks reserved
are blocked and lost.
Question: what is the probability with which an arriving call is blocked?

• Time blocking refers to the proportion of time the system spends in the blocking state
where all n elements are reserved.
• Call blocking is the proportion of the arriving calls which are blocked.
• Traffic blocking is the ratio of the traffic intensity of the blocked traffic to that of the
offered traffic.

All these quantities are equal if

• the arrival process is Poisson
• the service (holding) times of the calls are independent and identically distributed
J. Virtamo 38.3143 Queueing Theory / Loss system 2

Blocking in the M/M/n/n system: Erlang’s formula

Assume that the arrival process of the customers arrive according to a Poisson process with
intensity λ and that the service time obeys the distribution Exp(µ)


 λ = the arrival intensity (rate) of the customers

 µ = the service rate of the server (the mean service time is 1/µ)
Denote


 N = number of elements reserved (number of customers in system)
 π
j = P{N = j} the equilibrium probability of state j


The state variable Nt constitutes a Markov process of the birth-death type

- the state can change only stepwise
J. Virtamo 38.3143 Queueing Theory / Loss system 3

Erlang’s formula (continued)

0 l 1 l 2 j-1 l j n-1 l n
The state transition diagram is ... ...
m 2m jm nm

The balance across the cut leads to the recursion

a
λπj−1 = jµπj eli πj = πj−1 (a = λ/µ = offered traffic intensity)
j

By repeated application of the recursion one obtains

aj
πj = π0 , j = 0, 1, . . . , n
j!
a a2 an
From the normalization condition π0 +π1 . . .+πn = 1 one can solve π0 = 1/(1 + + + . . . + )
1! 2! n!

aj - the equilibrium probabilities in the M/M/n/n system

j! - truncated Poisson distribution
πj = 2
a a an aj −a
1+ + + ...+ - → e , as n → ∞
1! 2! n! j!
J. Virtamo 38.3143 Queueing Theory / Loss system 4

Erlang’s formula (continued)

The prob. πn of the state n gives the time blocking prob. (= call blocking prob. in this model).
We get the celebrated Erlang’s formula (also called Erlang’s B-formula):

an
A canonical traffic theoretical relation: it relates the size
E(n, a) = n!
a a2 an of the system n, the offered traffic a and the experienced
1+ + + ...+ quality of service (blocking).
1! 2! n!

Example.
A modem pool consists of 4 modems and the offered traffic intensity is 2 erl.
What is the probability that a connection attempt is fails due to blocking?
What is the blocking probability, if the number of modems is increased to 6?
Answer: the original blocking probability is 9.5 % and after the increase of the number of
modems it is 1.2 %.

Insensitivity

Erlang’s formula holds more generally independent of the form of the service time distribution.
The blocking depends only on mean holding time 1/µ through the traffic intensity a = λ/µ.
 J. Virtamo 38.3143 Queueing Theory / Loss system 5

Graphs for Erlang’s blocking function

100% 100%

1 10 20 30 40
10% 10%

2 50 60 70
1% 1%

3 4 5 6 7 8 9 10 80 90 100
.1% .1%

.01% .01%
0.1 1 10 0 20 40 60 80 100

• Horizontal axis: the offered traffic intensity a

• The parameter of the family of curves: the size of the system n
• Vertical axis: blocking probability E(n, a)
J. Virtamo 38.3143 Queueing Theory / Loss system 6

The required capacity as a function of the load

The following table gives the required number of trunks n as a function of the offered traffic
intensity a when the allowed blocking is 1 %. The last column gives the required relative
oversizing n/a, i.e. the ratio of the number of trunks to the load.

a (erl) n n/a
3 8 2.7
10 18 1.8
30 42 1.4
100 117 1.17
300 324 1.08
1000 1029 1.03

• When the traffic intensity a is large the Poisson fluctuations in the occupancy are small
in relative terms, and the required oversizing is small
√ √
– for Poisson distribution the standard deviation to mean ratio is a/a = 1/ a.
• From the point of view of dimensioning the system it is then (a large) more important
that the value of a on which the dimensioning is based has been correctly estimated and
the uncertainties in it have been properly accounted for.
J. Virtamo 38.3143 Queueing Theory / Loss system 7

Recursion formula

Erlang’s function is given by a simple expression which is easy to evaluate.

Some planning tools, however, make very frequent calls to this function, possibly in nested
iterative loops. Then it is important to pay attention to the fast computation of the function.
Often the following recursion is advantageous:
an
E(n, a) = n!
a + a2 +···+ an
1+ 1! 2! n!
2 n−1 a n
E(n, a) an a + a +···+ a
1+ 1! 2! (n−1)! a n!
= n!
an−1 2 n = (1 − 2 n)
E(n − 1, a) (n−1)!
1+ 1!
a + a +···+ a
2! n! n 1 + 1! + 2! + · · · + an!
a a
| {z }
E(n,a)
a
E(n, a) = E(n − 1, a)(1 − E(n, a)) ⇒ E(n, a)(n + aE(n − 1, a)) = aE(n − 1, a)
n

E(0, a) = 1 F (0, a) = 1
aE(n − 1, a) n
E(n, a) = F (n, a) = 1 + F (n − 1, a)
n + aE(n − 1, a) a

The latter form has been obtained by writing the recursion for the inverse F (n, a) = 1/E(n, a).
In this recursion, one first computes F (n, a) from which one obtains E(n, a) = 1/F (n, a).
J. Virtamo 38.3143 Queueing Theory / Loss system 8

The insensitivity of the equilibrium distribution

Above we have derived the result that the equilibrium probability distribution is a truncated
Poisson distribution:

aj j = 0, 1, . . . , n
j!

πj = 2



 a = λX̄
a a an 

λ = Poisson arrival intensity
1+ + + ...+ 
1! 2! n! 

 X̄ = mean holding time (1/µ)



The derivation was based on the assumption that the arrival process is Poissonian and that
the holding time obeys exponential distribution.
Remarkably, however, the result holds more generally: the insensitivity result.

The formula for the equilibrium probabilities (and in particular for the blocking prob-
ability πn) is valid for any holding time distribution and depends on the distribution
through the mean holding time X̄ only. (Poisson assumption, however, is necessary.)

• The proof of the insensitivity is a non-trivial task in the general case.

• In the following we make considerations by which verify the insensitivity in the cases
n = 1 and n = ∞.
J. Virtamo 38.3143 Queueing Theory / Loss system 9

Insensitivity in the case n = 1 (the M/M/1/1 system)

In the case n = 1 the truncated Poisson distribution reduces to the form

1 a
π0 = , π1 = , where a = λX̄
1+a 1+a

The insensitivity claim: The state probabilities are valid irresespective of the form of the
holding time distribution.
_
Proof: The state of the system alternates be- X 1/l

ì
í
î
ì
í
î
tween “server busy” and “server idle”. 1

Consider a full cycle which consists of one re- 0

t

ì
í
î
served period and one idle period. jakso

All cycles are stochastically identical. The probability of the busy/idle state equals the average
proportion of the busy/idle period of the total length the total period.






X̄ = expected duration of a busy period



 1/λ = expected duration of an idle period






(the interarrival times are distributed according to Exp(λ); memoryless!)

 X̄ + 1/λ = expected duration of the period


1/λ 1 X̄ a Knowledge about the distribution

π0 = = , π1 = =
X̄ + 1/λ 1 + a X̄ + 1/λ 1 + a of X was not needed!
J. Virtamo 38.3143 Queueing Theory / Loss system 10

Insensitivity in the case n = ∞ (the M/M/∞/∞ system)

Now the system is non-blocking. Consid-
er the number of calls in progress at time
0. This is equal to the number of starting Poisson-prosessi, l
events (before 0) of calls which are still in
progress at time 0. t

• The call durations X for all calls are independent: tail distribution G(t) = P{X > t}.
• Select all the calls that extend over 0. A call arriving at time t < 0 is selected with the
probability G(−t).
• The arrival process of the selected calls is an inhomogeneous Poisson process with intensity
λ(t) = λ · G(−t).
• The number of calls in progress at time 0 equals the number of arrivals from the inho-
mogeneous
Z 0
Poisson process in the interval (−∞, 0). The number is ∼ Poisson(a), where
a = −∞ λ(t)dt.
Z 0 Z ∞ 
∞
Z ∞
0

a = −∞
λG(−t)dt = λ 0
G(t)dt = λ /| 0 t{zG(t)} − 0 G (t)
| {z }
t dt
0 −f(t)
Z ∞
= λ 0
tf (t) dt = λX̄

The number is distributed as Poisson(λX̄) which depends on the holding time only
through the mean X̄.
J. Virtamo 38.3143 Queueing Theory / Loss system 11

The covariance of the occupancy states at two different instants of time in

the M/M/∞/∞ system
Consider the covariance between the values of the occupancy N at two different instants t1
ja t2 .
• Calls (customers) arrive with a Poissonian intensity λ.
• The tail distribution of the holding time X is G(x) = P{X > x} (general distribution).

t1 t2

ì
í
î
Dt
Denote

 N1 = number of calls at time t1


 N
2 = number of calls at time t2


We wish to calculate Cov[N1 , N2].

• If t1 and t2 are two close instants, one can assume that N1 ≈ N2 and
Cov[N1, N2] ≈ V[N] = λX̄ = a.
• If t1 and t2 are far apart, N1 and N2 are approximately independent and Cov[N1 , N2] ≈ 0.
J. Virtamo 38.3143 Queueing Theory / Loss system 12

Covariance of the occupations (continued)

We split both N1 and N2 into independent components as follows



 N1 = K1 + K1,2
 N
2 = K2 + K1,2


whence the covariance arises only through the common component K1,2,
Cov[N1 , N2] = Cov[K1 + K1,2, K2 + K1,2]
= Cov[K1, K2] + Cov[K1, K1,2] + Cov[K1,2, K2] + Cov[K1,2, K1,2]
= Cov[K1,2, K1,2] = V[K1,2]

The component K1: the number of calls which are in progress

X
at time t1 but not at time t2
- an arrival at t < t1
t t1 t2
- condition for the duration X: t1 − t < X < t2 − t
- random selection from a Poisson stream λ in −∞ < t < t1
with the probability G(t1 − t) − G(t2 − t)
J. Virtamo 38.3143 Queueing Theory / Loss system 13

Covariance of the occupations (continued)

The component K2: the number of calls which are in progress

X
at t2 but which start after t1
- arrival at t1 < t < t2
t1 t t2
- the condition for the duration X: X > t2 − t
- random selection from a Poisson stream λ in t1 < t < t2
with the probability G(t2 − t)

Component K1,2: the number of calls which are in progress

X
both at t1 and at t2
- arrival at t < t1
t t1 t2
- the condition for the duration X: X > t2 − t
- random selection from a Poisson stream λ in t < t1
with the probability G(t2 − t)
J. Virtamo 38.3143 Queueing Theory / Loss system 14

Covariance of the occupations (continued)

l

The quantities K1 , K2 ja K1,2 represent total num- l G(t1-t) l G(t2-t)

ber of arrivals from inhomogeneous Poisson pro-

cesses, which result from random selection or ran- t1 t2
dom split in the respective intervals. niiden kutsujen saapumis-
intensiteetti, jotka ovat
niiden kutsujen saapumis-
intensiteetti, jotka ovat
käynnissä vielä hetkellä t1 käynnissä vielä hetkellä t2
By the general results on random selection / ran-
dom split, K1 , K2 and K1,2 all obey Poisson dis-
l
tribution and are independent,
a1 a2

t t2
a12 1
 Z t1
K1 ∼ Poisson(a1 ), a1 = λ (G(t1 − t) − G(t2 − t))dt







 −∞

 Z t
2


K2 ∼ Poisson(a2 ), a2 = λ t1
G(t2 − t)dt





Z t1


 K1,2 ∼ Poisson(a1,2 ), a1,2 = λ −∞
G(t2 − t)dt

Since V[K1,2] = a1,2 we obtain (change of variable x = t2 − t)

0 −f(t)
Z ∞
Cov[N1 , N2] = λ G(t)dt
z }| {
z }| { Z ∞
t2−t1 /∞
0 t G(t) − 0
t G0 (t) dt
zZ }| {
∞
When t1 = t2 , we find that V[N1] = Cov[N1 , N1] = λ 0
G(t)dt = λX̄ = a, as it should.