CC BY
4ADAPTIVE CONTROL OF TRAFFIC FLOWS AND CONGESTIONS IN COMPUTER
CORPORATE NETWORKS
Lyudmyla Kharlay,
Kiev college of communication Andriy Skrypnichenko,
National Aviation University, aspirant Chang Shu, Ph. D. (Computer Sciences), Northwest University of Politics & Law, People Republic of China
Yaroslav Toroshanko, candidate of science (technic), State University of Telecommunications
АДАПТИВНОЕ УПРАВЛЕНИЕ ПОТОКАМИ ТРАФИКА И СКОПЛЕНИЯМИ В КОМПЬЮТЕРНЫХ КОРПОРАТИВНЫХ СЕТЯХ
Харлай Людмила Алексеевна, Киевский колледж связи, преподаватель
Андрей Анатолиевич Скрипниченко, Национальный авиационный университет
Чжан Шу, кандидат наук, Северо-Западный университет политологии и права, Китайская Народная Республика
Ярослав Иванович Торошанко, кандидат технических наук, Государственный университет телекоммуникаций, г.Киев
АННОТАЦИЯ
Представлен метод адаптивного формирования потоков сетевого трафика и способа настройки структур систем управления с косвенной обратной связью, которые управляют параметрами и структурой формирователя. Математические модели компьютерных сетей, в том числе моделей процессов управления потоками трафика, анализируются. Разработаны рекомендации по выбору параметров формирователей и системы управления ими (необходим порядок системы управления, структура формирователей, измерительных приборов параметров пакетов и т.д.) в зависимости от интенсивности потоков данных, а также их статистических описаний и сети структура. Исследованы характерные свойства процесса формирования потока, который необходимо учитывать при выборе параметров и структуры системы управления многоскоростного формирователя трафика с изменяемыми характеристиками окон загрузки.
ABSTRACT
The method of the adaptive forming of flows of network traffic and method of tuning of the systems managements with a non-direct feedback with control parameters and structure of shaper is presented. The mathematical models of computer networks, including models of processes of management the flows of traffic, are analysed. The recommendations in relation to the choice of parameters of shapers and control system by them are developed (it's necessary order of control system, structure of shapers, measuring devices of parameters of packages et cetera) depending on intensity of flows of data, their statistical descriptions and network structure. Research of characteristic properties of process is conducted forming of stream, which must be taken into account at the choice of parameters and structure of control system by the multi-rate shaper of traffic with variable parameters of load windows.
Ключевые слова: адаптивная система, формирование и регулирование трафика, маркерное ведро, многоскоростной формирователь.
Keywords: adaptive system, traffic policing and shaping, token bucket, multi-rate forming.
I. Introduction
Difference Between Flow Control and Congestion Control and Problem statement
Flow control is a mechanism used in computer networks to control the flow of data between a sender and a receiver, such that a slow receiver will not be outran by a fast sender. Flow control provides methods for the receiver to control the speed of transmission such that the receiver could handle the data transmitted by the sender. Congestion control is a mechanism that controls data flow when congestion actually occurs. It controls data entering in to a network such that the network can handle the traffic within the network.
In the open-loop flow control mechanism, receiver does not send any feedback to the sender and it is the most widely used flow control method. In closed-loop flow control, congestion information is transmitted back to the sender. Commonly used types of flow control are network congestion, windowing flow control and data buffer [1].
Congestion control provides methods to regulate the traffic entering in to a network such that the network itself could manage it. Congestion control prevents a network from reaching a congestive collapse where little or no useful communication is happening due to congestion. Congestion control is implemented in Transmission Control Protocol (TCP) and User Datagram Protocol (UDP) transport layer protocols.
Although, flow control and congestion control are two network traffic control mechanisms used in computer networks, they have their key differences. Flow control is an end-to-end mechanism that controls the traffic between a sender and a receiver, when a fast sender is transmitting data to a slow receiver. On the other hand, congestion control prevents loss of packets and delay caused due to congestion in the network. Congestion control can be seen as a mechanism that makes sure that an entire network can handle the traffic that is coming to the network. But, flow control refers to mechanisms
used to handle the transmission between a particular sender and a receiver.
We consider the methods of developing combine mechanisms of flow and congestion control using adaptive approach to traffic policing and shaping. This goal is the subject of represented paper.
II. Traffic shaping with using adaptive token bucket concept Foremost, we will give some important definitions [2].
1. The token generator TGE throws down in the bucket of E tokens at a speed of EIR in a second. If a bucket is filled, superfluous tokens are cast aside. Time of filling is
tfE = EBs!EIK
2. The token generator TGC throws down in the bucket of C tokens at a speed of CIR in a second. If a bucket is filled, superfluous tokens are cast aside.
3. Tokens accumulate in the buckets of E and C. General length of temporal interval, occupied by a token in the bucket
t =t +t t
of E, is equal e e ge , where te is duration of token
T
in the bucket of E; ge is duration of protect interval. General length of time interval, occupied by a token in the bucket of C,
t =tt + t t
is equal c c gc, where tc is duration of token in
T
the bucket of C; gc is duration of protect interval.
The number of tokens in the bucket of E is equal, in the
bucket of C equal
n
in the bucket of E is equal
T = n • t
is equal c c
c. Then the general total size of tokens
T = n„ • T
e, in the bucket of C Duration of arriving package will
designate through
Ps
4. Adaptation to the change of length and instantaneous intensity of entering packets can be carried out as follows:
by changing length of token at permanent length of guard interval;
by changing length of protective interval at permanent length of token;
- by changing size of «yellow range» [2,3];
by changing size of data and token buffer memory.
Both in that and in other case of speed of E and C will change to the limits, which depend on the maximal carrying capacity of, interconnect knot.
It's reasonable to adapt to the change to middle intensity of packets by the change of speeds of EIR and CIR. The load metering and packets specification procedures are shown on fig. 1.
We consider traffic shaper with variable speeds CIR and E (Fig. 2). They depend from the speed and acceleration of intensity grows.
IR, Kbps
' ^ t, ms
Fig. 1. Load metering and packets specification in traffic shaper. IR - instant intensity of traffic flow. Green packets - high priority, yellow- middle, red - low (first turn dropping)
Ir, Kbps UK.
ElR(i)
ClR(i)
t,mi
Fig. 2. Load metering and packets specification in traffic shaper with changing size of «yellow range»
A counter counts up the number of packets coming on the entrance of shaper. A store, essentially, is discrete integrator, scalar or vector. Frequency of token generator is regulated depending on the number of the accumulated packets, speed of accumulation (and in theory - and higher derivative). At devastation of token buffer (bucket) the growth of speed can be limited, based the parameters of incoming traffic and potential possibilities of destination node.
E C
Anyway speeds of IR and IR will change to the limits which depend on the maximal carrying capacity of switch node. It's expedient adapting to the change of middle intensity
EC
of packets by the change of the speeds of IR and IR and changing size of «yellow range».
The chart of multivariable adaptive token bucket mechanism is shown on fig. 2 (changing size of «yellow range» isn't shown). M-range integrator with weight coefficients
k = k (t),k2 = k2(t)km = km(t)
1 IV /'2 2W m\ ) estimates
speed, acceleration and higher derivatives of packets flow.
Controlsignals (t -t) , (t -t) , ytg (t -t) regulate size of data and token buffers and frequency of tokens series. Those signals are defined through traffic parameters including instant intensity and kind of statistic distribution (for instance heavy-tail distribution if traffic is self-similar).
In practice it's inappropriate to calculate derivatives of accumulation process higher than 2nd (speed and acceleration) due to quick deterioration of precision of results. So those results will be without effect on resulting efficiency of control process.
Besides the multivariable adaptive mechanisms of traffic shaping have a lot of ranges of freedom in contrast with traditional approaches. As a result we get auxiliary options for storing of arriving traffic and excluding packets loss and reduce the quantity of retransmission.
Data buffer
w
Input requests
Dropping with overload
Control signal ydb(t-T)
Signal counter n(t)
Dropping with overload
Control signal ytb(t-i)
Control device
if
Control signal
ytg(t-^)
Outgo] reques
Dropping with overload
Token buffer
V
out
Controlled
token generator
Fig. 2 The chart of multivariable adaptive token bucket mechanism (changing size of «yellow range» isn't shown)
The transfer efficiency depends not only from quality of adaptation but also from key parameters of network. Let consider these dependences.
III. Functional of Transfer Efficiency The general concept of the main and additional key network functions is that network elements and probes, which are used as service resource instances, are placed at certain nodes of the network infrastructure to pick up performance-related data, e.g. cumulative counters of protocol events. In constant time intervals or in near real time this performance-related data is transferred to higher level service assurance and performance management systems.
As the optimised parameters of task the following are chosen:
transfer delay ^dc;
C
- capacity p;
L
- losses of packets p at the data transfer;
D
- security and data protection level sp on network data transfer;
- quality of Web-service;
quality of transmission audio (audio files, plain and IP-telephony);
- speed and reliability of files exchange on the FTP protocol;
- speed and reliability of data transfer through E-mail;
- quality of transmission of video.
Use partial correlation coefficients of optimised parameters
taken from [4] (see Table 1).
Table 1
Parameter
tdc 1,0
Cp 0,98 1,0
Lp S 0,69 0,68 1,0
DSp ie 'o 0,89 0,86 0,69 1,0
Web <D O O 0,75 0,76 0,36 0,77 1,0
Audio ö o ti 0,85 0,64 0,50 0,56 0,30 1,0
FTP 0,27 0,75 0,63 0,61 0,57 0,44 1,0
E-mail o O 0,17 0,22 0,34 0,78 0,30 0,36 0,16 1,0
Video 0,87 0,89 0,84 0,82 0,53 0,67 0,79 0,30 1,0
X Cp Lp DSp Web Audio FTP E-mail Video
Current transfer delay is one of the most important parameters of quality of service. It is clear from data in Table 1 that all another key performance indicators have large correlation with it. So the problem of optimisation of transfer delay in general and its components, especially jitter, is rather urgent. Let's consider the components of transfer delay.
Current delay is measured as difference between moment
of sending and moment of receiving acknowledge tack:
tdc tack
Ai
t
s. We include such notations: <5,
pd and pd - time and standard deviation of packet delivery respectively;
- A ack and ^ack - time and standard deviation of wait of acknowledge respectively.
tdc = At„, + At
Then
pd ack
At„d *Ata.ck, Vid
ack
In general pd ack' pd Let the time of wait of current acknowledge (time-out) is
t
to . The total delay of delivery is random value with average
= 2 2 y/2
tdc j • f ^ total = I ® pd + ® ack ) T ,
dc and variation v ' . In order to
take in consideration the influence of variation of turn time we
use coefficient of variation ^tt ^totalftdc .
Besides we use normalized coefficient of losses
k _ Ntotal
N,
N N ,-N
y rec _ 1 — y total y n
total
N
total
, where Ntotal
N..
systems approach is needed. The criteria of optimisation of key parameters of functioning of network and current management by a network are ambiguous and contradictory. The account of these contradictions and search of compromise decisions is possible at the use of statistical methods, concordance of authenticity and detailed of basic data with physical sense of the decided tasks.
The processes of the key parameters change, from one side, are substantially non-stationary, and, with other are the tendencies of their changes are very similar. Therefore there is research of descriptions of their stochastic intercommunication of interest. This interest has not only theoretical but also practical character. As basic descriptions of stochastic intercommunication the coefficient of plural correlation and multiple step-by-step regression is used [5].
Optimal value of timeout is the functional of parameters
t
dc
t
ack
k
and
k
Vto _
components of vector symbol of transposition):
which can be represented as T
Atpd Atack K kl
(T is
to _ y
to
Atpd Atack
ktt kl
K tto ^ tdc
(1)
Strictly speaking we come to problem of vector
t
dc
t
ack
k
is total number of transmitted packets, rec is number of successfully received packets. Now define the structure of functional for optimal choice of timeout value.
For the decision of tasks of current management by networks
optimisation. However it is clear that k
and l are mutually independent random values. So we can replace the vector optimisation by weighted scalar
v = a At , + a At , + ak„ + a.k,
sum 1 pd 2 ack 3 tt 4 l, where
a i = 14
i are weighting coefficients choosing from
practical considerations [1]:
to a1 Atpd + a2Atack + a3 K + a4 K )
^ min, t. > t
dc
T
to
Fig. 3 shows the variations of normalized time-out 1norm
k
in dependence of turn time for different values 11.
Analysing the results of optimisation we come to conclusion
that the problem of current real-time optimisation of key parameters in independent network segment or in network in general is actually the adaptation problem [6] or, in wide sense, problem of (network) control.
0.75
0.5
0.25
0 5
—♦—ktt=0.01
10 - ktt=0.1
15
20
dc
-ktt=0.05 -*-ktt=0.5
Fig.3. Variations of normalised time-out
Conclusion
Proposed procedure of traffic shaping is rather simple and efficient. The results of modeling shows that it is possible to limit the frequency of token generator till such value, when all input traffic would been received and then transferred without losses and retransmissions. The impacts oa relative frequency of token generator is rather short and quickly decreased. Overload cancellation is achieved in small time, so load swings are comparative small as well. Though both the overloads and losses of traffic controllability may have place if the duration of bursts overcomes the reserve of dynamic stability of shaper. For example if the main and additional buffers are filled in the period higher than acceptable time-out this route or network segment in general become inaccessible. Such situations have to process with the utilities of transport or higher layers.
Variations of normalized time-out are rather smooth because of averaging of turn time on the interval of observing. With increase variation of turn time resulting time-out grows quickly. So permanent control of parameters and state of network for prevention deterioration of delay of delivery and jitter is urgent problem.
References
1. A.S.Tanenbaum, Computer Networks, 5th Ed./ Andrew S. Tanenbaum, David J. Wetherall. - Prentice Hall, Cloth, 2011. - 960 pp.
2. Chang Shu, Nick A. Vinogradov. The Method of Adaptive Shaping of the Traffic Flows of Calculating Networks // Proceedings the Fourth Congress "Aviation in the XXI Century", (Safety in Aviation and Space Technologies), V.1, Kiev, National aviation university, 2010, Sept. 21 - 23. - PP. 18.13 - 18.16.
3. Chang Shu. Adaptive control traffic rate by multivariable regulating token shaper // Scientific Notes of Ukrainian Scientific and Research Institute of Telecommunications, Nr 6 (34), 2014. - pp. 43 - 48.
4. Andriy Skripnichenko. Control of Quality of Service in High-Speed Computer Networks // Proc. of Int'l Conf. «Computer Science & Information Ttchnokogies" (CSIT'2013), 14-16 Oct. 2013, Lviv, Ukraine. - 2 PP.
5. A.A. Afifi, Statistical Analysis: A computer Oriented Approach. - 2nd ed. / A.A. Afifi, S.P. Azen. - Academic Press, New York, San Francisco, London, 1979. - 442 pp.
6. B. Widrow, Adaptive Signal Processing / Bernard Widrow, Peter N. Stearns. - Prentice-Hall, Inc. Englewood Cliffs, N.J., 1985. - 528 pp.
t
0
