Professional Documents
Culture Documents
Исследование импульсного интернет-трафика
Исследование импульсного интернет-трафика
Abstract of CBR traffic, while compressed audio and video, and asyn-
chronous transfer of files are examples of VBR traffic.
We study the effects of bursty Internet Traffic through simu- In the case of CBR traffic, it is necessary to allocate constant
lations. Both short-range dependency (SRD) traffic and long- bandwidth equal to their bandwidth demands to avoid lost and
range dependency (LRD) traffic are simulated over different overdue data. However, for VBR traffic if bandwidth is allo-
burst parameters. The results are collected for 10 different 24 cated to the VBR sources based on their peak demands, then
hour simulated periods in order to study and measure day-to- most of the time it will be under-utilized. Therefore, there is
day statistical fluctuation. Effects of employing different traffic a need in high speed switches to multiplex various VBR traf-
admission constraints are examined. An alternative for improv- fic streams in order to cost-effectively utilize overall capac-
ing network throughput and utilization is proposed. Finally, a ity. Nevertheless, multiplexing various sources decreases the
case when arrival patterns of traffic are correlated is explored. chance of a network ensuring performance standards (Quality
of Service or QoS) of each traffic source. Therefore, it is impor-
tant for a network to acquire a mechanism to cost-effectively
manage its limited resources to meet diverse performance re-
1 Introduction quirements of different traffic sources yet maintain reasonable
resource utilization.
The goal of multiservices wide-area networks is to serve a Currently, the IETF (Internet Engineering Task Force) has
variety of applications, from simple data and voice, to advanced developed two service architectures to support quality of ser-
multimedia applications such as CD-quality music, video-on- vice for real-time applications on the Internet. They are the in-
demand and videoconferencing, using the same underlying in- tegrated service (IntServ) and differentiated service (DiffServ)
frastructure (e.g., ATM and Internet). With the improvements [10] [11]. IntServ is a fine-grained approach which provides
of high speed and reliability of transmission media (e.g., fiber QoS to individual applications or flows, while DiffServ is a
optics), including the development of powerful and inexpen- coarse-grained approach which provides QoS to classes of ag-
sive workstations, providing this service is feasible. Such ap- gregated flows. However, the major drawback of IntServ is
plications can be classified as either time-sensitive or time- scalability since every router involved needs to store state in-
insensitive applications [10]. For time-sensitive applications formation per flow basis. The major drawback of DiffServ is
(e.g., interactive video browsing or video conferencing), the unpredictable end-to-end service quality since a gold class ser-
value of the communication relies on the time it takes for the vice of one network could be a bronze class service of another
information to be successfully delivered to the recipient and network.
typically can afford a small amount of data lost. On the other These issues are very complex and involve large scale cor-
hand, time-insensitive applications (e.g., file transfer) do not porations of many network entities [2] [5]. There are many pro-
have a time constraint and usually can tolerate a certain amount posals in the literature using either analytical based approaches
of delay. However, any data loss may be intolerable because the [1] [3] or simulation based approaches [6] [7] to provide QoS
value of the data depends on correctness of its content rather for real-time applications. Most analytical approaches are lim-
than the timeliness of delivery. ited by complexity which makes it difficult to derive generic so-
In this research, we are interested in time-sensitive appli- lutions for a real network. Although simulations allow greater
cations. Such traffic can be classified into either Constant Bit flexibility to study a system representing a real network, most
Rate (CBR) or Variable Bit Rate (VBR). The CBR traffic re- simulation based approaches lack a well defined set of parame-
quires constant bandwidth over time, while a VBR source, or ters accurately capturing burstiness of Internet traffic. As a re-
so-called bursty traffic, dynamically changes its bandwidth de- sult, the studies may not be meaningful enough to comprehend
mand over time. Uncompressed audio and video are examples the complex nature of a system. In addition, they are likely to
become untraceable when a system gets more complicated. OFF period
In this work, we aim to gain insight into how Internet traffic Source 1
ON
The N -Flow model is a variant of many ON-OFF models κ= λp (1)
presented in the literature [9]. The ON-OFF arrival process ON + OF F
consists of two states called the ON state and the OFF state. The 1-Flow model depends on four separate distributions. Each
During an ON state packets are generated with peak rate, λp . one governs a different sub-process which all together describe
During an OFF state no packets are generated. ON-OFF traffic characteristics of the 1-Flow model. They are:
sources generate packets according to this scenario. Durations
of ON and OFF state can be described by proper distribution • SV: Packet Service Time Distribution with mean 1/ν.
functions. In this study the term burst is used to describe a se- This distribution depends on packet size distribution and
quence of packets generated during an ON period and the term router speed.
flow or session is used to describe a sequence of ON and OFF • OFF: OFF Time Distribution with mean OF F . This
periods produced by the same source. Flow transmission peri- distribution depends on how bursts are generated and how
ods can be either infinite or finite. The N -Flow traffic model often.
is a superposition of N independent and statistically identical
streams of ON-OFF traffic type, as shown in Figure 1. There • ON: ON Time Distribution with mean ON . This dis-
are good reasons to believe that the N -Flow ON-OFF model tribution depends on packet size distribution and nature of
an application. For example, the ON-time distribution of where R(x) = P r(X > x), is a reliability function, α is a
file transmission depends on the distribution of file sizes. shape parameter, whose value is between 1 and 2, and np is the
While the ON-time distribution of voice transmission de- mean of the distribution, (i.e., E(X) = np ).
pends on the talk period of a speaker between breathing EXP produces short-range dependency (SRD) traffic, while
pauses. ON can be computed from np /λp . PT yields long-range dependency (LRD) traffic. Time between
bursts (OFF-time distribution) is exponentially distributed. The
• IN: Inter-Packet Time Distribution during a burst with router has speed, ν = 104 packets per second and transmits
mean 1/λp . This distribution depends on how packets are packets in a first-in first-out, (FIFO) manner.
generated and how often. We assume that all packets are for real-time applications and
have a pre-defined deadline, D. The usefullness of the packets
A useful shape parameter for describing a 1-Flow arrival pro- critically depends on time delivery to their destinations. Our
cess is the so-called Burstiness parameter, b, which is defined goal is to study the behavior of ON/OFF traffic using the N -
as followings. Flow model and its impact on network performance such as
percentages of packet loss and router utilization. We evaluate
OF F κ our study through simulations under three case studies below.
b= =1− (2)
ON + OF F λp
1. Infinite Sessions (Constant number of flows): Each
Thus κ and λp are related by ON/OFF source sends data according to some ON-OFF
pattern for the entire simulation. The number of flows ac-
κ = λp (1 − b) (3) tive at any time is constant. Consider a situation when
people transfer their files over the internet. These appli-
The value of b varies over a range of [0, 1] inclusively. If b = 0, cations are created every day and can be viewed as a sin-
then OF F = 0 which means that bursts abut each other and gle ON/OFF flow whose session is infinite with transfer-
packets are continuously transmitted with rate κ. The 1-Flow ring and idle periods as ON and OFF periods, respectively.
(ON/OFF) process is then reduced to a renewal process. On Since it is impossible to simulate a flow of infinite length,
the other hand, if b = 1, then ON = 0, which means that all we truncate our simulations after a 25 hour period and col-
packets in a burst are transmitted simultaneously with infinite lect relevant statistics after the warm up period of the first
rate, λp = ∞. The 1-Flow (ON/OFF) process becomes a bulk hour. Both SRD and LRD traffic patterns are simulated at
arrival process. When packets are produced by more than one different values of b and their impact on network perfor-
ON/OFF source (i.e., N > 1), an effective arrival rate, λ, of mance is studied.
the N -Flow model is the sum of individual mean rates, i.e.,
λ = N κ. 2. Finite Sessions (Variable number of flows): Each ON-OFF
source independently generates a flow that exists for a fi-
nite time (chosen as 300 seconds here). Flow-starts arrive
2.2 . Model Setting and Case Studies in a Poisson manner. If no traffic admission contraints
are applied, the number of flows active at any time is a
To clearly understand system behavior of a complicated traf- Poisson process. Otherwise, it changes based on the con-
fic model such as the N -Flow model it was necessary to first straint chosen. Consider a scenario when people make
begin our study on a simple network of a single router. Here telephone calls over an internet. Each call can be viewed
packets generated by the N ON-OFF sources arrive at a router as an ON/OFF flow with talk and silence periods.
where they are queued in an infinite buffer until they can be
sent to their destination (see Figure 1). 3. Correlated Sessions: This case is similar to the finite ses-
We assume that the packets are variable in size according sion case except that some (two or more) flows are allowed
to an Exponential distribution with mean of 1Kb (1000 bits). to start and end their sessions simultaneously. This corre-
There are on average np = 50 packets transmitted with con- lated session corresponds to a situation when a user makes
stant peak rate during a burst. Based on the value of b we clas- an attempt to increase the data rate that he or she is receiv-
sify traffic into two types: smooth traffic when b ≈ 0 and bursty ing by getting half of the data from one site and the other
traffic when b > 0.3. Two distributions: Exponential (EXP) half from another site at the same time.
and Pareto (PT) were chosen to represent the distribution of
In all cases two different packet managing disciplines are used
the number of packets in a burst. There are many versions of
to study the effects of N ON-OFF sources (i.e., N ≥ 1) on
Pareto distributions presented in the literature. The one used
network performance:
in this study is defined as follows [9]. Let X be the random
variable denoting the number of packets during an ON state. • No-Drop: overdue packets are transmitted as normal in a
Then FIFO manner.
1 • Drop: packets that are expected to miss their deadlines are
R(x) = x α
(4)
(1 + np (α−1) ) dropped at the router and considered lost.
For both cases, any packet that arrives late at its destination is 3.1 . Simulation Results and Performance
counted as lost and dropped. Analysis
3 Infinite Session EXP, Rho 0.5, 1 flow, Target Delay 0.01 seconds
0.01
MAX
0.006
0.004
We designed this case study such that a router sees the same 0.002
0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
30
MAX
AVG
20
15
10
0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
even though the mean packet rate of individual flows, κ, can 100
80
MAX
AVG
Good Throughputs
MIN
40
20
λ/N . Thus, the more flows being active, the smaller packet 0
0.5 0.6 0.7 0.8 0.9
burst parameter
1 1.1 1.2
rate each flow carries. The key idea here is to test if a router
could differentiate the differences in number of flows and how
Figure 2. fluctuation of Performance Metrics for
these changes affect the network performance.
EXP when N = 1, ρ = 0.5, d = 0.01
In addition, for a given N , we complicate the scenarios even
more by varying λp of each flow while keeping its κ the same.
Therefore, even though the aggregated flow seen by the router
consists of the same number of individual flows, each with the 5
PARETO, NO Drop, Rho 0.5, 1 flow, Target Delay 0.5 seconds
MAX
0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
also changes the burst parameter (see Equation 3). The goal 20
MAX
AVG
80
MAX
AVG
Good Throughputs
MIN
60
20
MAX
Average Delays at Sink
AVG
40
20
0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
15 MIN
network that can keep packet delays lower than the target. 0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
100
MAX
80 AVG
Good Throughputs
MIN
60
2. Percent Packet Loss: Packets that are past due at their des- 40
20
0.004 0.004
0.002 0.002
0 0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
1flow 1flow
30 30
2flows 2flows
5flows 5flows
10flows 10flows
20 20
10 10
0 0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
100 100
1flow 1flow
Good Throughputs
Good Throughputs
80 2flows 80 2flows
5flows 5flows
60 10flows 60 10flows
40 40
20 20
0 0
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
burst parameter burst parameter
Figure 5. EXP Performance Metrics using NO- Figure 7. EXP Performance Metrics using DROP
DROP policy for ρ = 0.5 policy for ρ = 0.5
0
0 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
20
20 1flow
2flows 5flows
15
5flows 10flows
10flows 10
10
5
5
0
0 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
100
100 1flow
Good Throughputs
1flow 80 2flows
Good Throughputs
80 2flows 5flows
5flows 60 10flows
60 10flows
40
40
20
20
0
0 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 burst parameter
burst parameter
Figure 6. PARETO Performance Metrics using NO- Figure 8. PARETO Performance Metrics using
DROP policy for ρ = 0.5 DROP policy for ρ = 0.5
Cut Off at 2
Average Delays at Sink
Put in Line
1
n = T λf (7) 0.5
0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
50
Cut Off at 2
Put in Line
Percent Total Loss
as
40
30
20
10
T λf κ 0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
ρ= (8) 100
Cut Off at 2
ν Put in Line
Good Throughputs
80
60
40
0
0.3 0.4 0.5 0.6 0.7 0.8
burst parameter
0.9 1 1.1 1.2
ρν
λf = (9)
Tκ
Figure 9. Performance Metric of NO-DROP EXP
In this study, we set ρ = 0.5, κ = 5000 packets/second, when ρ = 0.5, d = 0.5, m = 2
ν = 104 packets/second, and T = 300 seconds. Therefore,
m = 2, N = 2, and λf = 1/300 flows/second. We allow
the last burst of each flow to complete its transmission. There- Initially, we attempted to simulate ON/OFF traffic under the
fore, transmission duration may not be precisely 300 seconds NO-DROP policy using the Free-for-all criterion. However,
for every flow. The last burst of the last flow may complete its without any restriction on traffic admission and dropping, the
transmission after the simulated 25-hour period. In all cases, router was quickly overrun and packets were queued in a buffer
both EXP and PT traffic have D = 0.5 seconds. for an unboundedly long time for both EXP and PT traffic. No
results could be collected for this case. This indicates that an
4.1 . Performance Metrics admission constraint is necessary for the network in order to
provide acceptable services. Therefore, for the NO-DROP pol-
In addition to an average delay, percent packet loss, and icy only the results of EXP and PT traffic using the Cut-Off and
goodput defined in Section 3, new performance metrics are Put-in-Line criteria are presented as seen in Figures 9 and 10.
used to evaluate the effects of the three admission criteria as Clearly, overall PT traffic received worse service than EXP
followings. traffic did. Although the Put-in-Line yields higher % packet
NO−DROP PARETO, Rho 0.5, Router saturated at 2 flows, %Rej 19.71, Pr(HeldUp) 0.32, Avg HeldUp Delay 166.58s PARETO, Rho 0.5, Router saturated at 2 flows, %Rej 19.35, Pr(HeldUp) 0.306, HeldUp Delay 158.902 s
6
0.15
2 0.1
0.05
0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
25
50 Cut Off at 2 Free for All
Percent Total Loss
15
30
10
20
10 5
0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
100 100
Cut Off at 2 Free for All
Good Throughputs
Cut Off at 2
80 Put in Line 80
Good Throughputs
Put in Line
60 60
40 40
20 20
0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
burst parameter
burst parameter
Figure 10. Performance Metric of NO-DROP Figure 12. Performance Metric of DROP PARETO
PARETO when ρ = 0.5, d = 0.5, m = 2 when ρ = 0.5, d = 0.5, m = 2
loss and average packet delay than the Cut-Off, they perform serve that the graphs of DROP policy have much smaller slopes
comparably in terms of goodput. Notice that EXP traffic tends than ones of NO-DROP policy. This indicates that for variable
to perform better when b increases. This is because EXP nor- number of flows, performance is dominated by the number of
mally produces much smaller burst sizes compared to ones of active flows at any given time but not by bursty behavior of
PT. According to the results of this study, the maximum burst flows. Whenever this number exceeds m, performance col-
size measured over 10 days of EXP traffic is ≈ 800 packets, lapses.
while PT traffic could be > 2 × 106 packets. At smaller b, Figures 11 and 12 confirm that EXP traffic receives better
bursts are transmitted at a lower rate. Thus, it takes a longer service when b increases as opposed to PT traffic for the same
time to complete an ON period. Since a cycle length is fixed reasons described above. The graphs in these figures show
in this study (i.e., C = 0.01), an OFF period becomes shorter clearly that the Free-for-all criterion yields the most % packet
at smaller b. As a result, bursts are likely to abut on one an- lost and packet delays since no restriction on flow admission is
other and eventually jam a router. While, at higher b, bursts applied. In fact, % packet lost of the Free-for-all is more than
are transmitted with higher rate and possibly avoid each other. twice as much as ones of the other two criteria. However, it
Consequently, they can get through a router easier. However, achieves slightly better good throughput than the Cut-off cri-
terion, which is due to flow rejection. This indicates that if
EXP, Rho 0.5, Router saturated at 2 flows, %Rej 19.34, Pr(HeldUp) 0.304, HeldUp Delay 158.28 s
0.15
0.2
0.1
the network still can achieve comparable goodput and profits
as when such admission constraint is applied. Therefore, drop-
0.05
0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
25
20
Free for All
Cut Off at 2 ping alone has greater impact on performance than admission
Percent Total Loss
Put in Line
15
10
5
constraint. It is up to the network to decide whether it is better
100
0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
flows to be rejected.
Good Throughputs
Put in Line
60
40
20
0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
The Put-in-Line criterion yields the highest % goodput
burst parameter
of the data from one site and the other half of the data from
Cut Off at 2
60
another site at the same time. In this case, the illegitimate users 40
0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
40
ON/OFF source independently generates one flow at a time, 0.3 0.4 0.5 0.6 0.7 0.8
burst parameter
0.9 1 1.1 1.2
20
40
In order to keep overall flow arrival rate seen by the router the 20
References