Conference PaperPDF Available

Network traffic characterization for high-speed networks supporting multimedia

Authors:

Abstract

Continuously growing needs for distributed applications that transmit massive amounts of data has led to the emergence of high-speed networks that require broadband and multimedia capabilities. Such networks are supposed to have the ability to handle heterogeneous traffic and to manage large span of resources and services effectively. A single server G/D/I queuing system with infinite buffer is simulated with the consideration of three input traffic sources: exponential, Weibull, and normal distributions. The upper bounds on buffer size are evaluated for the given distributions
Network Traffic Characterization for High-speed Networks Supporting Multimedia
Khaled
M.
Elleithy
Computer Science and Engineering Department
University of Bridgeport
Bridgeport, CT 06601
clleithy
@
bridgcport.cdu
Ali S.
Al-Suwaiyan
Computer Engineering Department
King Fahd University
Dhahran 31261, Saudi Arabia
Abstract
Continuously growing needs for distributed
applications that transmit massive amount of data has led
to the emergence of high-speed networks that require
broadband and multimedia capabilities. Such networks
are supposed to have the ability to handle heterogeneous
trajjic and to manage large span of resources and services
effectively.
In
this paper, a single server G/D/l queuing
system with infinite buffer is simulated with the
consideration of three input traffic sources: exponential,
weibull, and normal distrbutions. The upper bounds
on
buffer size are evaluated for the given distributions.
1.
Introduction
Without the knowledge of traffic characteristics, we
would not meet what networks are supposed to achieve.
For example, without accurate traffic characterization, the
network may be forced to use overly conservative resource
schemes leading to underutilized servers. Traffic
characterization
is
an integral part of queuing systems
employed in the study
of
network, protocol, and switch
design performance. It serves the following purposes
[I-
21
:
Helps in specifying critical
QoS
parameters such
as
buffer
size
Predicting bandwidth requirements which allows
for better capacity assignment and congestion
control in communication networks
Estimation of statistical multiplexing gains
of
VBR
transmission over B-ISDNs
Mathematical analysis and simulation of traffic
signals models in the process of designing
communication networks.
data source (e.g., video, voice, multimedia,
...).
Then the
sample is studied well to see which distribution typifies its
first and second order statistics. When
a
distribution is
found, we must calculate the best fitting curve by choosing
the distribution parameters carefully. After that, the
optimal distribution is an accurate characterization of the
traffic. Usually, secondary steps come after that which are
dependent on the characteristics of the traffic, such as
simulation. In this paper, we have conducted a simulation
study to get the steady-state probabilities.
The process of traffic characterization needs an
empirical study of the source traffic to determine its
distribution
[
1-6,10,13-
141.
We have assumed the
distribution of the traffic source and we have studied the
effects of this assumption by measuring the steady-state
probabilities assuming a
G/D/l
system with infinite buffer.
The result of this study can help in determining the needed
buffer size and several useful metrics (e.g., throughput,
response time,
.
.
.
etc).
This paper is organized as follows. The problem under
study is defined in the next section. Then, we show
simulation results after describing the simulation
mechanism. Finally, the paper offers conclusions.
2.
Problem Definition
We have
a
single server model of type
G/D/l
with
infinite buffer, where
G
(input distribution) can be one of
the following distributions:
1
)Exponential
2)Weibull
3)Normal
These models are used with a single server queue, which
can be considered
as
a router with infinite buffer
as
shown
in Figure
1.
These input models are evaluated one at
a
time independently, not
all
together, and for each input
model, we approximated the steady-state probabilities.
The objective is
to
._
probabilities using simulation, given the above input
distribution, one at a time.
In general, traffic Characterization goes through the
following steps. First,
a
traffic sample is generated from
a
200
0-7695-1092-2/01
$10.00 0 2001 IEEE Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply. I 1 1 Exponential], infinite Queue I’ I Figure 1. System Model 3. Simulation Model In this section, we develop a discrete event simulator that simulates the behavior of the mentioned single server G/D/l queuing system [15]. We first describe the simulation mechanism used in the simulator, and then we give some simulation results. 3.1. Simulation Mechanism We have used an event-driven simulator programmed using the C language. The simulator is divided into the following components: External definitions: this part includes the “#include” directives, the “#define” directives, the global variable declarations and functions prototypes. Main function: this part controls the overall behavior of the simulator. Initialization routine: in this routine, we initialize some variables including simulation clock, state variables, the event list . . . etc. Timing routine: this routine determines the type of the next event and advances the simulation clock. New arrival routine: this is executed whenever we have an arrival event. It schedules next customer arrival and current customer departure times. Departure routine: this is executed whenever we have a departure event. If there is a customer waiting in the queue, the routine will schedule its departure. Statistics calculation routine: This is a routine to calculate some statistics, e.g., the frequency of having i customers in the system, maximum number of customers during the simulation period . . . etc. Random variate generator: this could be one of the following depending on the input distribution: a) Expon: this routine is used to generate exponentially distributed interarrival times. b) c) Weibull: this routine is used to generate weibull distributed interarrival times. Normal: this routine is used to generate normally distributed interarrival times. Cl P(i) = ~ 2 Ck k=O In order to approximate the steady-state probabilities P(i), we have used the following formula: Where C, is the number of times in which the system has i customers and m represents the maximum number of customers that the system has got during the simulation period. For the purpose of approximation, we can neglect P(n) for n > m. 3.2. Simulation Results In this section we present some experimental results. For each input distribution, we have run three experiments with different parameters, but all the experiments have the same service time, which is fixed at 0.5 time units. We have fixed the service time to isolate its effect and see only the effects of changing the input distribution. Figures 2-10 Show charts of P(N), which are the steady-state probabilities, versus N, which represents the number of customers in the system. These charts give us hints about the minimum amount of buffer needed to handle customers in the queue. For example, Figure2 shows us that we should have at least a buffer of size seven, because, as seen from the figure, the probability of having above seven customers in the system can be neglected. Table 1 shows upper bounds on buffer size associated with each input traffic type. 201 Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply. 4. Conclusion Interarrival distribution Traffic modeling or characterization describes the random flow of traffic associated with network sources in terms of stochastic models. Network sources might be a VBR (Variable Bit-Rate) video or LANWAN data. In this paper, it was shown how traffic characterization helped in determining critical QoS parameters, such as buffer size. In addition, we have seen how the upper bound on buffer changes as the input distribution changes. These results on the importance of traffic characterization as a necessary step on evaluating the performance of a network system that supports multimedia applications. Upper Bound on Buffer Size 5. References [I] T. Taralp, M. Devetsikiotis, and I. Lambadaris, “Traffic characterization for QoS provisioning in high-speed networks” Proc. of the 31”’ Hawaii international col$
on
System Sciences,
1998, pp. 485-492 vol. 7.
[2] C. Nayak and
M.
Ilyas, “Characterization of video
conferencing traffic in ATM networks”
Proc.
of
the IEEE,
Southeastcon ‘96,
1996, pp.344-347.
[3]
A.
Tarraf,
I.
Habib, and T. Saadawi, “Characterization of
packetized voice traffic
in
ATM networks using neural
networks”
IEEE GLOBECOM’93,
1993, pp. 996-1000
v01.2.
[4]
P.P.
Tang and T. Tai, “Network traffic characterization using
token bucket model”
IEEE INFOCOM’99,
1999, pp. 51-62
vol. 1.
[5]
C. Khunbo and
S.
Banerjee, “Network traffic characterization
of distributed database applications”
Proc.
Of
the
31“‘
Annual Simulation Symposium,
1998, pp. 98-105.
Exponential(
1
.O)
[6] W. Bin and A. Gupta, “Traffic characterization
of
VBR video
streams using uniform transmission
in
ATM networks”
Proc.
of
I997 international con5 on
ICICS.,
1997, pp. 1647-
1651 vol. 3.
[7]
B.
Mark,
D. Jagerman, and
G.
Ramamurthy, “Peakedness
measures for traffic characterization
in
high-speed
networks”
IEEE INFOCOM’97,
1997, pp. 427-435 v01.2.
[8] C. Wu, J. Jiau, and K. Chen, “Characterizing traffic behavior
and providing end-to-end service guarantees within
ATM
networks”
IEEE INFOCOM’97,
1997, pp. 336-344 vol.
1.
[9] W. Smith
and
A.
Zaghloul, “Network traffic characterization
in
a
transaction based modeling environment”
Zfld
IEEE
Symposium
on
Computers and Communications,
1997,
[IO]
Y.
Lin,
T.
Huang, and
Y.
Lai, “Characterization and control
of highly correlated traffic
in
high speed networks”,
Proc.
of
21”
IEEE
con5
on
Local Computer Networks,
1996, pp. 19-
27.
[I I]
E. Knightly, R. Mines, and H. Zhang, “Deterministic
characterization and network utilization for several
distributed real-time applications”
First workshop on
Object-Oriented Real-Time Dependable Systems,
1995, pp.
[
121 T. Ndousse, “Fuzzy characterization and control of
QoS
in
multimedia cell-based networks”
IEEE GLOBECOM’97,
[I31
C. Bisdikian,
E.
Spiegel, and A. Tantawi, “Traffic
characterization on high-speed Token-Ring networks”
IEEE
workshop
on
the
HPCS’92,
1992, pp. 1-4.
[
141 D. Wrege and J. Liebherr, “Video traffic characterization for
multimedia networks within a deterministic service”
IEEE
[I51 A.
M.
Law and W. D. Kelton, Simulation Modeling
&
pp.28-33.
63-70.
1997, pp.1662-1667 ~01.2.
INFOCOM’96,
1996, pp.537-544 ~01.2.
Analysis,
2”d
edition, McGrawHill, 1991.
7
Exponential(0.8)
9
Exponential(0.6)
Weibul( 1,l)
17
7
Weibul( 1,1.2)
I I
Weibu1(0.2,8.2)
8
3
I
14
I
Normal(
1,
1.69)
Normal(0.8, 1.69)
Normal(0.6, 6.25)
30
308
202
Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply.
45
40
35
0
30
2
*
25
E
20
15
10
5
0
0
1
2
3 4 5
6
7N
Figure
2.
Expon(1
.O)
Interarrivals
35
30
25
za
x
-
r
15
10
5
0
0
1
2
3
4
5
6
7
8
9
N
Figure
3.
Expon
(0.8)
lnterarrivals
203
Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply.
0
2
P
-
18
16
14
12
10
8
6
A
0
12 3 4
5
6 7 8
9
1011 121314151617
N
Figure
4.
Expon(O.6)
Interarrivals
0
1
2
3 4 5
6
7
N
Figure
5.
Weibul(1
,I)
Interarrivals
204
Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply.
60
50
0
1
2
N
3
Figure
6.
Weibull(2, 1.2) Interarrivals
.o
1
2
3
4
5
6
7
a
N
Figure
7.
Weibull (0.2, 8.2) lnterarrivals
205
Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply.
35
30
25
20
8
-
z
15
10
0
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18
19
20
21
22
23 24
N
Figure
8.
Normal(1, 1.69) Interarrivals
25
20
15
10
5
Figure
9.
Normal(0.8,1.69) lnterarrivals
206
Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply.
4.5
4
3.5
3
Figure
10.
Normal
(0.6,6.25)
Interarrivals
207
Authorized licensed use limited to: University of Bridgeport. Downloaded on February 24,2010 at 12:45:34 EST from IEEE Xplore. Restrictions apply.
... Also, based on the arrival rate(λ) the service rate((µ) and traffic rate(Y) are calculated. For each flow, the calculations were made and at the same time, sender knows the status of network capacity, traffic rate and the sending rate at the time of congestion [16]. Here the congestion bit value is sent to the sender and the capacity of the network and window size is updated based on the arrival rate, service rate and traffic rate. ...
Article
Full-text available
Continuously growing needs for Internet applications that transmit massive amount of data have led to the emergence of high speed network. Data transfer must take place without any congestion and hence feedback parameters must be transferred from the receiver end to the sender end so as to restrict the sending rate in order to avoid congestion. Even though TCP tries to avoid congestion by restricting the sending rate and window size, it never announces the sender about the capacity of the data to be sent and also it reduces the window size by half at the time of congestion therefore resulting in the decrease of throughput, low utilization of the bandwidth and maximum delay. In this paper, XCP protocol is used and feedback parameters are calculated based on arrival rate, service rate, traffic rate and queue size and hence the receiver informs the sender about the throughput, capacity of the data to be sent and window size adjustment, resulting in no drastic decrease in window size, better increase in sending rate because of which there is a continuous flow of data without congestion. Therefore as a result of this, there is a maximum increase in throughput, high utilization of the bandwidth and minimum delay. The result of the proposed work is presented as a graph based on throughput, delay and window size. Thus in this paper, XCP protocol is well illustrated and the various parameters are thoroughly analyzed and adequately presented.
Article
In general, network traffic data has a heavy-tailed probability distribution. The Entropy-Based Heavy Tailed Distribution Transformation (EHTDT) has been developed to convert the heavy tailed network traffic data distribution into a transformed probability distribution. In practice, the entropy distribution of the transformed probability distribution exhibits a type of linearity that gives rise to an eigenstructure that allows the characterization of network traffic data to effectively lossily compress network traffic data via the Rate Controlled Eigen-Based Coding. The aforementioned eigenstructure is motivated by singular value decomposition theory. A very high compression ratio can be achieved by the proposed method. Results of applying the methods to real network traffic data network traffic data are presented.
Article
Internet traffic has been shown to exhibit characteristics that are very different from traditional telephone network traffic such as long-range dependence (LRD). Therefore, performance analysis of queueing system in LRD traffic is needed to design reliable networks under the tremendous growth of Internet traffic. In this paper, we evaluate the sensitivity of input distribution on the output performance measures. We conduct a Monte Carlo simulation based on a designed experiment in order to relate four performance measures such as the mean and the expected maximum number of requests in queue, and the mean and the expected maximum time in queue to the degree of LRD in arrival, traffic intensities and the types of distributions in G/G/1 queue. Experimental results show that mis-identification of input distributions has a significant effects on the output performances especially when one of the input distributions has LRD property. For such cases, Weibull distribution overestimated the output performances while lognormal distribution underestimated those of the LRD queues.Scope and purposeIn the era of information technology, understanding the nature of queueing system associated with a network traffic is crucial for effective network management. The main feature of arrival and service times in information network traffic is LRD property. Many long-tailed distributions have been applied to the fitting of arrival and service time distributions with LRD property. The main purpose of this paper is to see the exchangability of those long-tailed distributions to LRD distributions. We generate both arrival/service time distributions from Pareto, one of the representative distributions with LRD and fit these distributions with long-tailed distributions such as Weibull and lognormal. The result of our sensitivity analysis is expected to provide the case where careful input data analyses are specially needed for the queueing network management.
Article
Consider various host systems that consist of both hardware and software. There is an IT center which governs the activities of all of those host systems including the reported failures. The center's major concern is deciding the proper level of repair capacity for various kinds of failures due to hardware, software, and operator related problems where corresponding failure and repair times may exhibit fairly large variations. In this paper, we use Monte Carlo simulation in order to decide the optimal repair capacity when both failures and repair times of individual host systems exhibit long range dependent (LRD) property. In terms of both utilization rate and the maximum waiting time, we suggest optimal capacity. Various sensitivity analyses follow.
Article
We propose a rate-controlled service discipline for supporting B-ISDN services. According to our approach, traffic streams from different connections can be well regulated at the output of each node based on their rate requirements. Moreover, the traffic envelope of a connection inside the network can be effectively characterized. By assuming a leaky-bucket constrained input source, we further prove that the proposed scheme can provide end-to-end delay and jitter bounds for each connection passing through a multi-hop network. Finally, we make a comparison of related works and show the effectiveness of our algorithm.
Article
Workload characterization is difficult yet vital in the construction of realistic and useful performance models. Previously, a Poisson process has been used to represent the arrival of traffic to nodes in a network model. Traffic could be constructed as groups of related messages called transactions. Letting these transactions represent basic units of work, a hierarchy is described here that allows for the construction of workloads representing traffic at the network user level. These user workloads are then decomposed into independent Markov Modulated Poisson Processes for use as input to analytic solution techniques based on the GI/G/m two-moment decomposition method.
Article
We are developing a new approach to traffic characterization motivated by the traffic-queue analysis techniques of effective bandwidth and the use of log moment generating functions. A generic representation of traffic exhibiting non-Gaussian marginal and long range dependence is modeled by three dissimilar stochastic models. Generated traces of the model traffic are put to test in a G/D/1/$\infty$ system by computer simulation. Experimentation demonstrates that the marginal and autocorrelation pair of descriptors do not adequately characterize traffic such that different input models with similar marginal-autocorrelation produce vastly different queuing behavior. The ability of the new approach to describe and differentiate stochastic traffic is shown to be robust with regard to queuing results, whereas marginal and autocorrelation prove insufficient.
Conference Paper
Asynchronous transfer mode (ATM) broadband networks support a wide range of multimedia traffic (e.g. voice, video, image and data). Accurate characterization of the multimedia traffic is essential in ATM networks, in order to develop a robust set of traffic descriptors. Such a set is required by the ATM networks for the important functions of traffic enforcement (policing) and bandwidth allocation utilizing the statistical multiplexing gain. In this paper, we present a novel approach to characterize and model the multimedia traffic using neural networks (NNs). A backpropagation NN is used to characterize the statistical variations of the packet arrival process resulting from the superposition of a number of packetized voice sources. The NN is trained to characterize the arrival process over a fixed measurement period of time, based upon sampled values taken from the previous measurement period. The accuracy of the results were verified by matching the index of dispersion for counts and the variance of the arrival process to those of the NN output. The results reported show that the NNs can be successfully utilized to characterize the complex non-renewal process resulting from the aggregate voice-packet arrival process with extreme accuracy. Hence, NNs have excellent potential as traffic descriptors for the usage parameter control algorithm used in admission control and traffic enforcement in ATM networks
Conference Paper
In high-speed networks based on asynchronous transfer mode (ATM), variable bit rate (VBR) sources generate bursty cell streams which share the link bandwidth wherever multiplexing occurs. In theory, the bandwidth requirement per stream to support a given quality-of-service at a multiplexer should generally decrease as the number of streams increases. In practice, high statistical multiplexing gain is difficult to achieve because good traffic characterizations are usually not available. This paper proposes the use of peakedness-based measures to capture the statistical information from traffic streams needed to determine bandwidth allocations. The standard definition of peakedness applies only to point process models of traffic. Yet fluid models of traffic have certain advantages in terms of tractability. Hence, we introduce a new measure called modified peakedness, which encompasses point process and fluid models under a common framework. We develop some of its properties and specialize it to several common traffic models. Finally, we study the effectiveness of the peakedness/modified peakedness as burstiness measures for real-time traffic
Conference Paper
First Page of the Article
Conference Paper
This paper investigated the problem of deriving token bucket parameters from the observed traffic patterns of computer network flows in two cases. In the basic case, we identified the set of token bucket parameters for any given flow so that all the packets in the flow will be delivered immediately without incurring any delay or loss. Then, we extended the basic case by adding a queue to the token bucket model to temporarily store packets that cannot be delivered right away until enough tokens get accumulated in the bucket. The queue has the effect of smoothing the traffic so that the resulting token bucket parameters are less demanding. It also makes the derived token bucket parameters less fluctuated over time. Relationship among the queue size and token bucket parameters for a given flow are analyzed rigorously. Queuing delay for each packet in the flow is calculated and used to describe the adjusted post-queuing traffic pattern. Simple and efficient algorithms for the derivation of measurement-based traffic specification (MBTS) are presented, along with empirical results. This MBTS technology alleviates the need for users to explicitly characterize the traffic a priori in order to reserve network resource in an integrated services network
Conference Paper
The authors' efforts in characterizing the traffic of distributed database systems are described. This is an important class of applications considering that database or transaction processing systems form the core of several key commercial applications. Specific traffic source models for this application class have not been developed which is imperative for their successful deployment in emerging networks. The main contribution of the paper is to present the relationship between a distributed database application scenario and the corresponding network traffic characteristics, thus providing guidelines to these applications to compute their traffic descriptors accurately
Conference Paper
We propose a simple method which can be used to characterize any pre-compressed variable bit rate (VBR) video stream in terms of ATM traffic parameters: peak cell rate (PCR), sustainable cell rate (SCR), and maximum burst size (MBS). Though it is simple to find the PCR parameter, there is generally a very large set of SCR/MBS pairs that can be used to describe a given video stream. An SCR/MBS pair is considered most suitable which results in efficient utilization of the network resources provided the quality of service (QoS) commitments can be met. We suggest that the resources should be allocated based on the actual traffic rather than the worst case traffic corresponding to the selected SCR/MBS pair. The proposed scheme will benefit both, the network operator and the user, since it results in more efficient utilization of the network resources