Computer Networks

Modeling arrivals and service with Poisson

Saad Mneimneh
Computer Science
Hunter College of CUNY
New York

Now remember this:

A poisson has a short memory span of few
months, but a Poisson process is memoryless.

Introduction

In computer networks, packet arrivals and service are modeled as a stochastic process
in which events occur at times t1 , t2 , . . . For instance, in the figure below, t1 , t2 , . . . can
be interpreted as the packet arrival times, or the service completion times. Accordingly,
Ti , defined as ti ti1 for i > 1, can be interpreted as the inter-arrival times of packets
(intervals between subsequent arrivals), or the delays experienced by the served packets
(assuming that the server is always busy). Similarly, A(t) denotes the number of
packets that arrive in [0, t], or the number of packets served in [0, t]. In the rest of this
document, we will refer to packets rather than service, but it should be clear that the
discussion applies to both.
T3
T2

A(t)

T1
t1

t2

t3

Figure 1: Arrival/service process

Poisson

The packet inter-arrival times are typically modeled as a Poisson process, i.e. Ti are
independent and identically distributed (IID, so we drop i from the term Ti in the
following expression), and obey an an exponential distribution:
FT (t) = P (T t) = 1 et
where is a parameter. We will give a name shortly.
The probability density function for T is therefore (the derivative of 1 et with
respect to t):
fT (t) = et
Therefore,

t2

P (t1 T t2 ) =
t1

et = et |tt21 = et2 + et1

The expected value of T can be obtained as:

tet dt =

E[T ] =
0

Therefore, the parameter is called the arrival rate, or simply rate 1 . Similarly,

Z
E[T 2 ] =

t2 et dt =
0

2
2

Therefore, the variance is:

2 (T ) = E[(T E[T ])2 ] = E[T 2 + E[T ]2 2E[T ]T ]
By the linearity of expectation,
2 (T ) = E[T 2 ] + E[T ]2 2E[T ]E[T ] = E[T 2 ] E[T ]2 =

1
2

Poisson process is memoryless

Now we prove a unique property of the exponential process, known as the memoryless
property. Consider the waiting time until some arrival occurs. The memoryless property states that given that no arrival has occurred by time , the distribution of the
remaining waiting time is the same as it was originally. Mathematically,
P (T > + t|T > ) = P (T > t)
The proof is simple as a direct consequence of the exponential distribution:
1A

mathematical
P result, known as the law of large numbers, says that if Xi are IID,

then limn
tn
limn A(t
n)

i=1

Xi

= E[X]. Therefore, the inverse of the rate can be expressed as

= limn

Pn

i=1

Ti

= E[T ] =

1
.

P (T > + t|T > ) =

P (T > + t and T > )

P (T > + t)
e( +t)
=
=
= et
P (T > )
P (T > )
e

As a direct application of the memoryless property, consider Z1 to be the waiting

time until the first arrival occurs after some time t. Assume also that A(t) = n, and
tn = t. Then
P (Z1 > z|A(t) = n, tn = ) = P (Tn+1 > z + t |Tn+1 > t , tn = )
Note that Tn+1 is independent of all earlier inter-arrival times, and thus of tn . Therefore,
P (Z1 > z|A(t) = n, tn = ) = P (Tn+1 > z + t |Tn+1 > t )
Using the memoryless property:
P (Z1 > z|A(t) = n, tn = ) = P (Tn+1 > z) = ez
Since the result does not depend on , we see that conditional on A(t), Z1 is
independent of tn .
P (Z1 > z|A(t) = n) = ez
Since the result also does not depend on n, Z1 is independent of A(t).
P (Z1 > z) = ez
The same argument applies if we conditioned not only on tn , but also on t1 , t2 , . . . tn1 .
Therefore,
P (Z1 > z|{A( ), 0 t}) = ez
If we now define Zi for i > 1 to be the inter-arrival time from the i 1st arrival to the ith arrival after time t, we see that conditional on A(t) = n, Zi = Ti+n
and, therefore, Z1 , Z2 , . . . are IID exponentially distributed conditional on A(t) = n.
Since the distribution is independent of n, Z1 , Z2 , . . . are IID exponentially distributed
and independent of A(t). It should be also clear that Z1 , Z2 , . . . are independent of
{A( ), 0 t}.
The memoryless property shows that the portion of a Poisson process starting at
some time t > 0, is a probabilistic replica of the process starting at t = 0. Therefore,
t + ) = A(t + ) A(t) has the same distribution as A(). This property is called
A(t,
t + ) is independent
the stationary increment property. We have shown also that A(t,
1 , t2 ), . . . , A(t
k1 , tk )
of all arrivals in [0, t]. Therefore, for t1 < t2 < . . . < tk , A(t1 ), A(t
are independent. This property is called the independent increment property.

A paradox

In this section we consider a hypothetic example of a Poisson process applied to

bus arrivals/departures. Suppose a bus arrives at (then departs) a station according to a Poisson process with average inter-arrival time of 20 minutes (i.e. = 0.05
bus/minute). When a customer arrives to the station at time t, what is the average
waiting time until the next bus? According to the previous section, the waiting time

is Z1 , which is exponentially distributed (with parameter ) and independent of everything prior to t. Therefore, E[Z1 ] = 1/ = 20 minutes. Now, when a customer
arrives to the station at time t, what is the average time since the last departure? Let
us call this time Y1 . We dont know how many arrivals/departures we had in [0, t], so
let us condition on A(t).
P (Y1 > y|A(t) = n) = P (Tn+1 > y) = ey
Since this is independent of n, Y1 is independent of A(t) and is exponentially
distributed (with parameter ). Therefore, E[Y1 ] = 1/ = 20 minutes.
By definition, the inter-arrival time is the time from the last departure to the next
arrival. Therefore, the average inter-arrival time is E[Y1 + Z1 ] = E[Y1 ] + E[Z1 ] =
20 + 20 = 40 minutes, by the linearity of expectation. But it should be 20 minutes
only! What happened? This is an example of what is known as random incidence.
The customer is more likely to fall in a large interval. For an intuition, consider the
following example of randomly throwing a ball into bins.

Figure 2: Random incidence

The average bin size is:
(1 ) +
1
=
2
2
The ball falling at random will observe a bin of size (1 ) with probability (1 ),
and a bin of size with probability . Therefore, the average bin size observed by the
ball is:
1
(1 )(1 ) + = (1 )2 + 2 1 2 >
2
For a mathematical interpretation of the paradox, let Z(t) be the time until the
first arrival after t.

Figure 3: Time until next arrival

Similarly, let Y (t) be the time since the last departure before t.
Now, Y (t) + Z(t) is shown below:
The time average of Y (t) + Z(t) can be computed as:

Figure 4: Time since last departure

Figure 5: Y(t)+Z(t)

lim

1
t

Pn

But limn

Pn
i=1

Ti2 /n = E[T 2 ] and limn

lim

Pn

Ti2
Ti2 /n
sum of areas of squares
= lim Pi=1
= lim Pi=1
n
n
n
n
length
T
T /n
i=1 i
i=1 i

(Y (t)+Z(t))dt =

1
t

Pn

(Y (t) + Z(t))dt =
0

i=1

Ti /n = E[T ]. Therefore,

2/2
2
=
1/

Poisson as a counting process

Another characterization of a Poisson process can be obtained by starting to observe

the relation between t and T :
tn =

n
X

Ti = T1 + T2 . . . Tn

i=1

From this relation, we can show that the probability density function for the arrival
times is:
ftn (t) =

n tn1 et
(n 1)!

To describe a Poisson process as a counting process (that counts the number of

arrivals), we are interested in obtaining the probability mass function for A(t), i.e.
P (A(t) = n) for some integer n. Note that
P (tn+1 (t, t + ])

=
=
=
=

P (A(t) = n and Z1 )
P (A(t) = n) P (Z1 |A(t) = n)
P (A(t) = n) P (Z1 ) (memoryless property)
P (A(t) = n) (1 e )

Therefore,
P (A(t) = n) =

P (tn+1 (t, t + ])
1 e

Now,

t+

P (tn+1 (t, t + ]) =
t

n+1 n e
n+1 tn et
d

n!
n!

as it can be seen from the figure below:

t t+

(n 1)/

Figure 6: Area under ftn (t) around [t, t + ]

Therefore,
P (A(t) = n) =

n+1 tn et
n!(1 e )/

Note that e = 1 + ()2 /2! ()3 /3! + . . .. Therefore, (1 e )/ =

+ o()/, where lim0 o()/ = 0. We can now take the limit as 0 (which
justifies the integral approximation done above) to obtain:
P (A(t) = n) =

(t)n et
n!

It is not hard to show that any counting process that satisfies the above probability
mass function, in addition to the stationary and independent increment properties, is
a Poisson process. For instance P (T1 > t) = P (A(t) = 0) = et . Also, P (Tn >
t + ) = 0|A(
, ) = 1, A(0,
) = n 2) =
t|tn1 = ) = lim0 P (A(t,
P (A(t) = 0) (stationary and independent increment).
From this probability mass function, we can obtain the following:
E[A(t)] = t
2

E[A (t)] = t + (t)2

2 (A(t)) = t

PASTA

Let N (t) be the number of customers in the system at time t. Define

pn (t) = P (N (t) = n)
to be the probability of having n customers in the system at time t. If the system has
a steady state distribution, then we can also define
pn = lim pn (t)
t

=
This of course means that N
npn = limt 1t
i=0
the same probability as seen by an arriving customer:

Rt
0

N ( )d . Let us now consider

an (t) = P (N (t) = n|a customer arrives just after time t)

and
an = lim an (t)
t

Is an = pn , n = 0, 1, . . .? We provide two examples to show that this is not necessary.

6.1

Example 1

Assume customer inter-arrival times are unifromly distributed between 2 and 4 seconds. Assume also that customer service times are all equal to 1. Then an arriving
customer always finds an empty system. Therefore, a0 = 1. On the other hand, the
average number of customers in the system is N = 1/3 (apply Littles theorem on
= 1/3 and T = 1). Therefore, p0 6= 1.

6.2

Example 2

Assume customer inter-arrival times are exponentially distributed (Poisson process).

Assume also that customer service times and future arrival times are correlated, e.g.
the service of the nth customer is half the inter-arrival time between customers n and
n + 1. Clearly, an arriving customer always finds an empty system. Therefore, a0 = 1.
However, the average number of customers in the system is easily seen to be 1/2.
Therefore, p0 6= 1.
It turns out that this cannot happen with a Poisson process where service times
and future arrivals are independent.
an (t)

=
=

t + )
lim0 P (N (t) = n|A(t,
t + ) = 1)
lim0 P (A(t) D(t) = n|A(t,

t + ) and A(t)
where D(t) is the number of customers that depart in [0, t]. Since A(t,
are independent, and service times are independent of future arrivals, A(t) D(t) and
t + ) are independent. Therefore, lim0 P (A(t) D(t) = n|A(t,
t + ) = 1) =
A(t,
P (A(t) D(t) = n) = P (N (t) = n) = pn (t). Taking the limit as t , we have
an = pn . This property of a Poisson process is called PASTA (Poisson Arrivals See
Time Averages).

Merging and splitting Poisson processes

Consider the number of arrivals of a Poisson process in the interval [t, t + ].

t + ) = 0) = e = 1 + o()
P (A(t,
t + ) = 1) = e = + o()
P (A(t,
t + ) 2) = 1 (1 + )e = o()
P (A(t,
where lim0 o()/ = 0.
It is not hard to show that any process that satisfies the above probabilities, in
addition to the stationary and independent increment properties, is a Poisson process.

 t + ) = 0) = P (A(t) = 0)P (A(t,

t+
For instance P (A(t + ) = 0) = P (A(t) = 0, A(t,
) = 0) = P (A(t) = 0)(1 + o()). Therefore,
lim

P (A(t + ) = 0) P (A(t) = 0)
= P (A(t) = 0)

d
P (A(t) = 0)) = P (A(t) = 0)
dt
So P (A(t) = 0) = et . In general, one can show that P (A(t) = n) = (t)n et /n!.
This alternative definition of a Poisson process allows us to think about merging and
splitting Poisson processes.

7.1

Merging

We can show that the sum of two Poisson processes with rates and is a Poisson
process with rate + .
t + ) = 0) = (1 + o())(1 + o()) = 1 ( + ) + o()
P (A(t,
t+) = 1) = (+o())(1+o())+(+o())(1+o()) = (+)+o()
P (A(t,
t + ) 2) = o()
P (A(t,
Since the two processes satisfy the stationary and independent increment properties, the resulting process does too. Therefore, the resulting process is a Poisson
process with rate + .

Figure 7: Merging two Poisson processes

7.2

Splitting

A Poisson process with rate can be split into two independent Poisson processes as
follows: each arrival is independently directed to process 1 with probability p and to
process 2 with probability 1 p.
P (A1 (t, t + ) = 0) = (1 ) + (1 p) + o() = 1 p + o()
P (A1 (t, t + ) = 1) = p( + o()) + o() = p + o()
P (A1 (t, t + ) 2) = o()
Similar calculation can be done for process 2 by exchanging p and 1 p. Since the
original process satisfies the stationary and independent increments properties, so do
process 1 and process 2. Therefore, process 1 and process 2 are both Poisson processes
with rates p and (1 p) respectively.

1p

(1 p)

Figure 8: Splitting a Poisson process

It remains to show that process 1 and process 2 are independent. Note that
conditional on the original process, process 1 and process 2 are not independent. In
fact, one determines the other.

P (A1 (t) = m, A2 (t) = n|A(t) = m + n) =

m+n
m

pm (1 p)n

m+n
. This is simply the binomial distribution, since, given
= (m+n)!
m!n!
m
m + n arrivals to the original process, each independently goes to process 1 with
probability p.
where

P (A1 (t) = m, A2 (t) = n|A(t) = m + n) =

P (A1 (t) = m, A2 (t) = n)

P (A(t) = m + n)

Combining the two equalities, we get:

(pt)m ept [(1 p)t]n e(1p)t
m!
n!
P (A1 (t) = m, A2 (t) = n) = P (A1 (t) = m)P (A2 (t) = n)

P (A1 (t) = m, A2 (t) = n) =

which proves that A1 (t) and A2 (t) are independent. To show that process 1 and process
2 are independent we must show that for any t1 < t2 < . . . < tk , {A1 (ti1 , ti ), 1
i k} and {A2 (tj1 , tj ), 1 j k} are independent. The argument above shows this
independence for i = j. For i 6= j, the independence follows from the independent
increment property of A(t).

References
Dimitri Bertsekas and Robert Gallager, Data Networks
Rober Gallager, Discrete Stochastic Processes