Beruflich Dokumente
Kultur Dokumente
Queueing Systems
Professor Izhak Rubin
Electrical Engineering Department
UCLA
rubin@ee.ucla.edu
2014-2015 by Izhak Rubin
Server
i ,
i 1.
0,
lim P X t j
t
P j ,
a
i 0
a
i 0
a
i 0
0 1... j 1
We have a0 1, a j
j, j 0 .
12 ... j
j
a
i 0
a
i 0
1 .
i 0
1 .
1
0,
lim P X t j
j 0.
j
t
1 , 1
Prof. Izhak Rubin
, 1.
1
, 1
1
E W E D 1
, 1.
We set:
( j ) = P{an arriving message finds j messages ahead of itself in the queue}.
We obtain
( j ) P( j ) (1 ) j ,
j 0,
by using a result that states that Poisson Arrivals See Time Averages (PASTA).
To calculate the waiting time distribution, we assume a FCFS service policy.
An arriving message that finds upon arrival that there are j messages ahead of itself, j 1, will have to wait
until these messages are served before it is admitted into service. The services times of these j messages
are statistically independent and identically distributed; this is also true for the message that is
in the midst of its service when the message for which the wait time is computed arrives into the
system. Clearly, the distribution of the sum of j independent service times (actually, the sum of the residual
service time of the message in service plus the service times of the waiting j-1 messages) is equal to the j-th order
convolution of the service time distribution. When j=0, the arriving message enters service without having to wait.
W t (0) 1 ( j ) (1 e
j 1
*j
)u (t ) , t 0.
We obtain:
W t 1 e
1 t
u(t ),
t 0 , <1.
1 Mbps
1
mess
confirming that =
20 [
].
50m sec
sec
Hence:
Traffic intensity = =
0.8 1,
so that the system is stable and reaches as it evolves steady state behavior.
Example 1 (Cont.)
The steady-state mean queue size is:
0.8
EX
4 [messages].
1 1 0.8
The steady-state mean message delay and waiting times are computed as:
1
4
E D 1E X
W t 1 e 1 t u (t ), so that at steady state the probability that the waiting time is longer than t
is expressed by: P{W > t}= e 1 t , t 0. Hence, the probability that the message wait time is longer
than 1 sec is computed to be:
P{W > 1}= e 1 0.8e 4 0.0146.
Thus, 1.46% of the messages are experiencing waiting times that are longer than 1 sec., noting that
the average message waiting time in the queue is equal to 0.2 sec.
Prof. Izhak Rubin
i min i, m , i 1
Server 1
Server 2
[ Erlangs],
m
Server m
10
0,
lim P X t j
t
P j ,
a
i 0
a
i 0
We have a0
a
i 0
, j 0.
j
fj
,
0 jm
j
j!
j !
1, a j
j
fj
, j m 1
m !m j m j m !m j m
a
j 0
1.
j 0
1
0,
lim P X t j
t
P j , 1.
Prof. Izhak Rubin
11
j0
1
where P 0 a j
j 0
The steady-state mean queue size is:
j P j ,
X j 0
,
1
1.
1
,
Departing message rate [mess/sec] = D
m , 1
f , 1
D ,max
fD
f D ,max
, 1
1, 1.
Prof. Izhak Rubin
12
Under a FCFS policy, a message that finds upon arrival j messages in the system,
will not have to wait if j < m. In turn, if j m, the message will have to wait until the number of
messages ahead of itself in the system falls down to m-1 (so that a server becomes available
to admit it into service); in this case, its waiting time is equal to the sum of j-(m-1)=j+1-m time intervals between
message departures. But the departure process during this period is a Poisson
process with departure rate equal to m , so that each inter-departure time interval is exponentially distributed
with mean interdeparture time equal to 1/m . Hence, we write:
m 1
j 0
j m
W t ( j ) ( j ) (1 e m t )u (t )
* j 1 m
j 0
j m
W t P( j ) P( j ) e
m y
j m !
j m
m y
m dy.
Prof. Izhak Rubin
13
Server 1
Server 2
i min i, m , i 1.
Server m
14
fj
,
0 jm
P 0
j!
P j
fj
P 0
m !m j m , m j N
1
m 1 f j N
fj
where:
P 0
,
j m
j
!
m
!
m
j m
j 0
f / .
fN
PB N P N P 0
m !m N m
The message departure rate (message throughout) is:
D 1 PB 1 P N [mess/sec]
The throughput load (Erlangs) is:
f D D 1 1 PB 1 1 PB f
[Erlangs].
15
A j
j
1 PB
P j
, 0 j N 1
1 PB
m 1
N 1
W t A i A i 1 e m t u t
i 0
im
Since A i
P i
P i
,
1 PB 1 P N
P j 1,
j 0
0 i N 1
N 1
where
* i m 1
and
j 1 , we have:
j 0
N 1
* i m 1
m 1
1
m t
W t
P i P i 1 e
, t 0.
1 P N i 0
im
16
PB N P N .
The message departure rate (message throughout) is:
D 1 PB 1 P N [mess/sec]
The throughput Erlang load rate is:
f D D 1 1 PB 1 1 PB f
[Erlangs].
E(X)
.
17
0.8 1.
Since the system's storage capacity N is finite, the system is stable and reaches as it
evolves steady state behavior.
18
19
i
fi
ai
, i 0;
i
i !
i!
[Erlangs].
a
k 0
Server 2
j!
m
k 0
Server 1
, 0 j m.
k!
Server m
Prof. Izhak Rubin
20
k 0
fk
.
k!
[Erlangs].
Note: This blocking probability formula also holds for the M/G/m/m system model.
Messages incur no waiting time, W = 0.
Message mean delay time (as confirmed by using Little's
Little's formula) is thus equal to the message mean service time:
X
1
D
.
1 PB
Prof. Izhak Rubin
21
0.8 [ Erlangs].
22
Hence,we obtain the following results for the blocking probabilities for different values of
selected system storage levels:
For N=m= 1, we obtain: PB 44.44%, so that 44.44% of arriving calls are blocked;
For N=m= 2, we obtain: PB 15.09%, so that 15.09% of arriving calls are blocked;
For N=m= 3, we obtain: PB 3.86%, so that only 3.86% of arriving calls are blocked, while
the remainder 96.64% of arriving messages are admitted into the system and are
immediately served without incurring any waiting time.
Prof. Izhak Rubin
23