Documente Academic
Documente Profesional
Documente Cultură
OF NETWORKS PERFORMANCE
Merlo, G. 1, Britos, P. 2, Rossi, B. 2 & Garca Martnez R.2
Abstract
The present study analizes the usage of
backpropagation neural networks for the determination
of average response time of an SNA network.
In order to evaluate the method, this was applied to
an existent SNA telecommunications network, to
which measurements of response time were taken
against users load. Moreover, a backpropagation neural
network was implemented in PASCAL programming
language, asumming the total number of users as
inputs (in binary state representation indicating
whether they are active 1 or inactive 0 ), and
giving the estimated response time as output of the
network. It is beyond the scope of the present article to
give a detailed description of a backpropagation neural
network algorithm and its implementation, for which
references are included. We asumme the reader is
familiarized with neural networks techniques, and we
emphasize on its practical utilization and the
convenience of its usage in the determination of
communication parameters in a computers network
(specifically, an SNA network in the present case).
At the same time, we modelized the SNA response
time behaviour based on mathematical expresions
taken from queueing theory, in order to evaluate and
contrast both response time estimations, the one
achieved by neural networks usage and the one from
classical mathematical models.
Finally, we detail the conclusions of the study, in
where we analyze the convenience of the usage of each
model, with its pros and cons.
Token Ring
Control Unit
Workstations
Desktop computer
Desktop computer
Satellite
Desktop computer
Desktop computer
Satellite
Satellite Dish
Satellite Dish
Communication
Controller
Communication
Controller
Desktop computer
Desktop computer
ame
Desktop computer
Microwave
Radio Tower
Desktop computer
Radio Tower
1. Introduction
The determination of response times in a remote
telecommunicactions network is inherently a queue
problem. For this reason, in the case of a simple
network with a small number of traffic variables, the
average response time of a terminal could be calculated
by means of mathematical expressions from the
queueing theory.
However, in reality this does not happen. In a
corporative network with hundreds of terminals where
users access to different types of applications with
their corresponding activities times, and where there
exist alternative communication links, each with their
characteristical resonse time it becomes impractical
Desktop computer
Desktop computer
1,6
1,4
1,2
0,8
0,4
0,2
0
0
14 21
28 35 42
49 56 63 70
77 86 104 112 120 128 136 145 152 160 168 175 182 190 198 205 213 220 229 237 246 255 264 272 280 288
Active Users
Resource consumption:
Measuring systems like this are not transparent to
the hole communication system, in the sense that it
usually requires the allocation of resources such as
CPU processing time, bandwidth, among other
resources. Hence, it is recommended not to measure at
peak hours, when the measuring system could help to
cause the saturation of the communication channel,
8users
input Nr. 1
8users
input Nr. 2
8users
input Nr. 3
8users
input Nr. 4
8users
input Nr. 5
8users
input Nr. 6
8users
input Nr. 7
8users
0.2- 0.3sec.
input Nr. 8
input Nr. 1
8users
input Nr. 2
8users
input Nr. 3
8users
input Nr. 4
8users
input Nr. 5
8users
input Nr. 6
8users
input Nr. 7
8users
input Nr. 8
Output 2
Input 2
Hidden
Unit 2
0.3- 0.4sec.
ResponseTime
Range
Output 1
Hidden
Unit 1
0.1- 0.2sec.
Control Unit 1
.
.
.
8users
Input 1
0- 0.1sec.
0.4- 0.5sec.
0.5- 0.6sec.
Hidden
Unit N
0.6- 0.7sec.
Input 40
Output 10
0.7- 0.8sec.
Control Unit 5
HiddenLayer
0.8- 0.9sec.
0.9- 1.0sec.
0,7
0,5
0,4
0,3
Time Response (seconds)
0,2
0,1
0
0
7 14 21 28 35 42 49 56 63 70 77 86 104 112 120 128 136 145 152 160 168 175 182 190 198 205 213 220 229 237 246 255 264 272 280 288
Active Users
0,078
Infinite Buffer
Both and of a Poisson distribution
2 output channels (simultaneous sessions running
in the mainframe)
both channels with the same
Stationary process
0,076
0,074
0,072
0,068
0,066
0
=/
Po =
/ n! ) +
1
/(( M 1)!( M ))
L=
Po
+
( M 1)!( M ) 2
M
20
40
60
80
100
120
140
160
180
200
220
300
320
340
360
380
400
420
440
460
480
500
5. Method comparison
In order to compare the overall tendencies of the
three sets of results, we present them in an unique
picture, figure 8.
1,8
M Po
Lc =
( M 1)!( M ) 2
Wc = Lc /
W = Wc + 1 /
1,6
1,4
1,2
1
Real values
Neural Networks
Mathematical Model
0,8
0,6
Channel Speed
0,4
0,2
Channel Speed
28
8
27
2
25
5
23
7
22
0
20
5
19
0
17
5
16
0
14
5
12
8
86
11
2
70
56
a.
b.
42
(bits/second)
28
Active Users
(bits/second)
Buffer
3745
14
0
0
Being,
Po :
Probability of non-existing bits in the system
(n can take values from 0 to 2)
L:
Average quantities of bits in the system
Lc:
Average quantity of bits in the queue, waiting
channel attention.
Wc:
Average permanence time of each bit in the
queue.
W:
Average permanence time of each bit in the
system.
The following picture shows the scheme of this
mathematical model.
Sessions ( Attention Channels)
c.
a.
b.
c.
d.
6. Conclusion
From the facts exposed formerly, we can conclude
that:
1.
2.
a.
b.
c.
References
[1]. Sackett, G. Introduction to SNA Networking. J.
Ranade IBM Series, 1989.
[2]. Sackett, G. Advanced SNA Networking . J.
Ranade IBM Series, 1991.
[3]. Freeman, J., Skapura, D. Neural Networks.
Addison-Wesley, 1993.
[4]. Kosko, Bart. Neural Networks and Fuzzy
Systems. Prentice-Hall, 1992.
[5]. Clifford Lau. Neural Networks. IEEE Press,
1992.
[6]. Schwartz, M. Telecommunication Networks.
Addison-Wesley, 1998.
[7]. King, P. Computer and Communication System
Performance Modelling. Prentice Hall, 1990.