Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Performance Analysis of Computer Network
Queueing SystemLecture 12
1
ERF – 2016-2
2
Inside a Router
3
Inside a Router
4
Element of Queue
5
Element of Queue
6
Kendall Notation of Queueing System
7
Kendall Notation of Queueing System
8
Performance Measure if Queueing System
9
Little’s Law
10
Litlle’s Law
Queueing System EquationPerformance Measure of Interests:–Jobs in the Queue–Jobs in the Server–Utilization or Traffic Intensity
–The Response Time:
[ ] [ ]E N E Rl=[ ] [ ]sE N E Sl=
[ ] [ ]
1
, [ ]
sE N E S
and
if X thus X ES
r llr rµ
l r
= =
= <
= =
[ ] [ ] [ ][ ][ ]1
E R E W E SE SE R
r
= +
=-
We further have the performance measure of interests:
2
[ ][ ]1
[ ]1
[ ]1q
E SE W
E N
E N
rr
rrrr
=-
=-
=-
Poisson Process
Poisson Process
Exponential and Posisson
Properties of Poisson
M/M/1 ModelArrival Model: Poisson process
Poisson process is viable model for M/M/1 whenpackets are originated from large population ( such asfrom Internet) of independent users
The superposition of Poisson process yields thePoisson process as depicted in the figures
The M/M/1 Performance MeasuresPoisson arrival with rates λ. Hence the average time of packetarrivals equals to E[A] = 1/ λ
Markovian Service time with rate µ. The average service timeE[S] = 1/ µ
The utilization of M/M/1 queue is obtained from λ/ µ. From Little’s law we also obtain ρ = λ E[S]
The average number of jobs in the system equals to E[N] = λ.E[R]. Thus, E[N] also can be obtained from ρ / 1- ρ
From the Little’s law, the average response time can be obtained as E[R] = E[S] / 1 – ρ
The average waiting time of a job in the queue can be measured by: E[W] = ρ .E[S] / 1- ρ
The DTMC Model of M/M/1
With the generator Matrix Q
M/M/1 SolutionFrom the given performance measure
Find the steady state probability using GBE
From the obtained steady state probability wecan find another performance measure
M/M/1 SolutionThere are three ways to solve the steady-state probabilities
Global Balance Equations
Suppose ‘probability flow balance’ hence:p0λ0 = p1µ1 Þ pi(λi + µi) = pi-1 λi-1 + pi+1µi+1 for I = 1, 2, 3, …; thus we have
1
0 0 10 1
1 0 1
1;1
ik
i ik kk
i k k
p p pllµµ
-
-¥= +
= = +
= =+
ÕåÕ
ip
1ijjp =å
M/M/1 SolutionSince station has infinite buffer, no loss involves, thus:–Overall arrival rate:
–Overall service rate:
If r=λ/µ < 1, the queue is stable and the throughput X = λ.
Then
å¥
=
=0i
iip ll
å¥
=
=0i
iip µµ
0 0[ ] ; [ ] ( 1)i q i
i iE N ip E N i p
¥ ¥
= =
= = -å å
E[R]=E[N]/λ; E[W]=E[Nq]/λ=E[R]-E[S].
The probability of having k jobs in the queuing station:
å¥
=
=ki
ipkB )(
Example: Simple Single Server
M/M/1 Solution: Local Balance
Solving flow equation is easier than GBE: piλi = pi+1µi+1; iÎÀ
The computation is done repeatedly, hence we can find all state probability
M/M/1 with Constant RatesAssume the rate is constant
λ = λi and µ = µi for all states. Hence– pi = p0(λ/µ)i– pi = p0(r)i; i = 0,1,2,…
If r = λ/µ we have by normalisation that– p0 = 1 - r
The steady-state probabilities:– Known then
– pi = (1 - r)ri; i = 0,1,2,…
p0 = 1 - r Þ 1 – p0 = åipi è utilisation.
Throughput X = λ hence r = XE[S].
å¥
=
=0
0 1i
i prå
=i
pr
10
Other Performance Measure
E[N] = ρ / 1 – ρ
E[R] = 1 / µ - λ
In case we have E[S] , thus E[R]= E[S]/1 - ρ
A Performance of Router
Router A sends 8 pps, on the average, to Router B. The mean size of a packet is 400 bytes with exponentially distributed. They are connected with 64 kbps line speed. –What is the utilization of the router ?
1864 / / (400 8 / ) 20 /
0.4
skbps s x bit packet packtes s
lµ
lrµ
-== =
= =
A Performance of Router–What is the E[N] ?
–What is the probability of the number of packet is 10 or more ?
[ ] 0.671
E N rr
= =-
10 410Nr r -= =
M/M/1 AnalysisWe start with simple Queuing M/M/1
Model as a birth death process
Generate Markov Chain
Birth death is a Markov Process in which transition are only allowed between neighboring states
Steady state solution of a birth death process–Steady state probability of being in state n is expressed as:
and is probability being in state 0
np
0 1 10
0 1... 2
... ; 0,1,2,...,nnp p wherenl l l
µ µ µ-= = ¥
0p
For the proof of
where previously we have
Because M/M/1 is a special case for birth death process where
1
0 0 10 1
1 0 1
1;1
ik
i ik kk
i k k
p p pllµµ
-
-¥= +
= = +
= =+
ÕåÕ
01 0
1
p plµ
=
i j
i j
l l
µ µ
=
=
Hence by simplification
–In general or traffic intensity. Thus, the probability of being in state i is implicitly determined by traffic intensity –And
–Hence
0
i
ip plµ
æ ö= ç ÷è ø
l rµ
æ ö=ç ÷
è ø
0i
ip pr=
0 2
1 11 ...
p rr r r¥= = -
+ + + +
(1 ) iip r r= -
Example of M/M/1
Consider the following Queue
–The utilization can be calculated:
–E[N] in the system can be calculated as well:
–E[R] can also be calculated
0.30.5
lµ==
0.3 0.60.5
r = =
0.6[ ] 1.51 0.6
E N = =-
1[ ] 5E Rµ l
= =-
[ ] 1.5[ ] 50.3
E NE Rl
= = =
Probability of being in state 5
n Probability of the are number 10 IP Packets or more in the system:
5 55 (1 ) 0.4 0.6 0.03p xr r= - = =
10 0.06r =
M/M/1/m Single Server Queue
Suitable to model M/M/1 in practice
Total of jobs in the queuing station is limited by m capacity
Whenever m has been reached, there are no jobs arrive
Consequently, we have the following flows
Steady state solution for M/M/1/m
1
0 00
0
0 1
1
; 1,2,..., ;
;
1
0,1,2,...,11
;
1
ii
m
ii m
ii
j
p p
p
i m
p
p p
i m
lµ
r
rrr
l lµ
r
rr
µ
-
=
+
+
æ ö= = = =ç
-=
---
÷ø
==
è
=
Õ
Example: M/M/1/5The probability of packet lost equals to p5 –
– See the table below.
55 6
11
p r rr-
=-
The blocking probability pn is a function of m, l,and µ
Cases –(1) a single M|M|1 with arrival λ and service rate Kµ; –(2a) M|M|K with arrival λ and service rate µ; and –(2b) K number of M|M|1 each with arrival λ/K and service rate µ.
Three queuing Cases
Solution(1) E[R] = E[S]/(K-r) where r=λ/µ.
(2)
(3) E[R] = KE[S]/ (K-r) = K. E[R] in Solution (1)
20
)1(!][][][
K
K pKK
SESEREr
r-
+=
Example: Performance measure of HTTP 1.1
An ISP involves two HTTP server A and B which incorporatesversion 1.0 and 1.1, respectively . The average responsetime of Server A < Server B.
Is Server A better than server ?– A is better than B. Since λ is constant, the smallest service rate
E[S] makes the smallest E[R]. Hence, System A has smallestservice time that reflects better service. Service time is timeneeded to serve a packet arrives at a server.
To fulfill real time requirement, what are the performancemeasures ?– E[R] and smallest E[W]
To fulfill reliability service, what are the performancemeasures?– By ignoring the Network delay, the retransmission attempt of
TCP is interested measure.
Simulation
THANK YOU44