Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstract
In TDMA (time division multiple access), stations transmit their messages on a shared communication channel using their dedicated time slots. All the previous delay analyses of TDMA have been based on the assumption that either the interarrival times of the traffic is exponential or the
message lengths are geometrically distributed. This paper
presents a generalized model in which the above assumptions are relaxed. This model allows us to compute the exact
performance characteristics in closed forms for the mean
and the variance of the queue size and the message delay
for the TDMA. The model is used to define and compare
five bursty traffic distributions. These distributions are used
to study their effects on the buffer size and the end-to-end
delay for the Mars Regional Network.
1. Introduction
TDMA (time division multiple access) with its variations have been widely used in various network architectures including time slot interchange switches, multiplexing, multiple-access, polling, fixed assignment and demand
assignment multiple access (MAC) protocols. These architectures have been implemented for transmission of digital
information over wired or wireless communication channels. TDMA schemes belong to the contention-free MAC
protocols. MAC protocols can be classified into fixed assignment, demand assignment (reservation), random access, hybrid of random access and adaptive protocols. Fixed
assignment protocols that are based on TDMA schemes are
easy to implement with deterministic control and deterministic channel utilization. A survey and classification of the
MAC protocols are presented in [7].
In TDMA, the time axis is equally divided into successive frames, each consisting of a fixed number of indexed
slots. In the basic TDMA scheme, each station has exactly
0 This
700
Message length, x
Density, f(x)
600
500
400
300
200
100
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Throughput, S
2. Traffic Descriptors
The traffic arriving to a multiaccess channel consists of
two distributions: one describes the interarrival process of
messages, and the other describes the message lengths. All
known mathematical analyses are based on the assumption
that either the interarrival component of the traffic is a Poisson process or that the message length component of the
traffic is geometrically distributed. The assumption of exponential interarrivals provides analytical simplicity. As we
will see in Section 4, the average delay becomes sensitive to
the message length distribution as the throughput increases.
Figure 2 shows the throughput versus delay performance of
a network of four stations and two orbiting satellites which
was simulated for the Mars Regional Network [8].
This figure illustrates the effect of the message lengths on
the delay when two distributions, with the same mean, are
used to generate messages. The solid line indicates the
delay when all messages are of fixed length of ten packets. The dashed line indicates the delay when the message lengths are geometrically distributed with a mean of
ten packets. In both cases, the interarrival distribution was
chosen to be exponential. We also observed the same behivior when the interarrival distribution was chosen to be
geometric.
To complicate the matter, the average message length
alone does not provide sufficient information about the message sizes. Two or more parameters may be required to realistically model actual message lengths. Since the geometric
distribution is completely characterized by one parameter, it
may be a poor candidate to reflect a realistic message length
distribution.
The main weakness of the Poisson process model
(M/G/1) is that the process is completely characterized by
duced a new model known as the G/Geom/m. Both approaches lead to a closed form solution for the generating
function of the limiting queue size. Using the limiting generating function, one can compute the expected queue size
and, furthermore, deduce the expected delay behavior of the
protocol. In this section we describe how the G/D/1 model
can be used to study the performance of the TDMA protocol
with an arbitrary traffic.
The main idea behind the G/D/1 approach is as follows.
The arrival process is taken to be any renewal process so that
the interarrival times will be independent and identically
distributed (iid) random variables. The messages arrive at
the renewal times. We assume that the message lengths are
independent and identically distributed according to any realistic probability distribution.
Let B represent the packet size in bits, and let C represent the channel speed in bits per second, then the packet
transmission time is Ts = B=C: We assume that the duration of each slot is exactly the same as the transmission
time of one packet. The delay that a packet suffers has three
components:
, )(z , 1) :
G(z ) = g(z )(1z ,
g(z )
This gives that the expected limiting queue size is
G0 (1) = g
00 (1) + 2(1 , )
2(1
, )
L2
=
=
s +W +T ;
D = MT
q
s
2
The Poisson traffic model (M/G/1, see [9]) can be described as a special case of the above G/D/1 model when the
service time is deterministic and the arriving traffic forms a
Poisson process. In this case, the station under consideration generates independent and identically distributed (iid)
Poisson random variables with parameter Ts , where Ts is
the duration of the slot. In other words, the arrival times
of the packets form a Poisson process with rate . Hence,
the random variable Nn representing the number of packets
arriving during the n-th frame time (frame consisting of M
slots) is
Nn = Z1 + Z2 + + ZM ;
That is,
Zi iid
Poisson(Ts ):
Nn Poisson(Tf ):
In this case = E (Nn ) = Tf and
g00 (1) = E (Nn (Nn , 1)) = (Tf )2 :
Hence,
G0 (1) = g
00 (1) + 2(1 , )
2(1
, )
(2 , ) :
2(1 , )
L
D1 =Ts , is
M
Db 1 = 12 M + 2(1MS
, S ) + 1 = 2(1 , S ) + 1;
where the throughput S = , is the fraction of time the
server is busy. To compute the limiting variance of the
queue size, we have g (z ) = e,Tf (1,z) and Tf = . Differentiating three times; g 0 (1) = , g 00 (1) = 2 , g 000 (1) =
3 , and using equation (1), we get the limiting variance of
the queue size;
18 + 102 , 3 )
12 = (12 , 12(1
:
, )2
P (Y
k) = (1 , )k,1 ;
k = 1; 2; :
Nn =
M
X
i=1
Xi Yi ;
Xi iid
B (1; ):
= g0 (1) = M
:
Similarly, we get
g00 (1) =
=
2
,1
1
, M :
(2)
Lq = G0 (1) , =
2(
, 1) + , M
:
2(1 , )
MTs
Wq = Ts
Lq =
, 1) + , M
:
2(1 , )
1
2(
D2 = 21 MTs +
MTs
, 1) + , M
2(1 , )
1
2(
o
+
Ts :
,
(3)
g(z ) =
, , z (1 , ,
) M :
1 , z (1 ,
)
500
400
,1
1
:
300
g00 (1)
Lq = 2(1 , ) =
200
100
, 1) +
:
2(1 , )
1
2(
0
1
0.5
0
10
Using g 000 (1) and equation (2), in the limiting variance formula of equation (1),
f(6M (M + 1) , 2 (M , 4)(M + 1) ,
(M + 5)(M + 1))g + 12M (2 , )
1 ,
2
D3 = 2 MTs +
Db 3 = 12 M +
M (M + (M + 2) , 2 (M + 1))
=
,
2
2
12M (1 , ) :
1
Nn =
j =0
K1;j +
j =0
K2;j + +
ZM
X
j =0
KM;j ;
Zi iid
Poisson(Ts );
iid
where Ki;j Geometric(
) which are independent of Zi
and Ki;0 = 0. This is a particular case of the compound
g(z )
=
=
,Tf (1 , E (z K1 1 ))
exp f,
(1 , z )=(1 , z (1 ,
))g :
exp
g000 (1) = 2 + 6
In this model, during each slot time packets are generated by iid Poisson random variables with parameter Ts ,
where Ts is the duration of the slot. That is, arrival times
of messages form a Poisson process with rate . This is a
generalization of Model 1 since messages constitute a random number of packets. For the comparison purposes we
will assume that the message lengths are geometrically distributed. Hence, the random variable Nn , representing the
number of packets arriving during the n-th service time, is
Z2
X
MTs
, 1) +
2(1 , )
1
2(
Ts :
, 1) + S
2(1 , S )
1
2(
+1
where the throughput S = . In order to evaluate the limiting variance of the queue size we use equation (5) and
Z1
X
, 1) +
:
2(1 , )
1
2(
12
MTs
Wq = Lq =
Throughput
22
f
g00 (1) = T
2 (2(1 ,
) + Tf ) =
600
(4)
,1
1
+6
2 )
,1
1
32 =
12(2
,
) , 6 2+2
,
2 + 22 6,
12(1
, )
,
Nn = `
M
X
i=1
Xi ;
Xi iid
B (1; );
(5)
(6)
= 1 + (` , 1)T , where T B (1; p). Since, our service time consists of M such slots, the arrival random variable Nn of the G/D/1 system is the sum of such independent
random variables. That is,
Nn =
G0 (1)
`+
M ,1) , 1
M
2(1
, )
, )
+ 2 (1
` + (MM,1) , 1
0
:
Lq = G (1) , =
2(1 , )
2 :
g00 (1) = 2 + (`p +`p1 , p) (` , 1) , M
, )
`p (` , 1) ,
MTs + (`p+1
,p)
M
Wq =
:
2(1 , )
(7)
M ` + (MM,1) S , 1
b
+ 1;
D4 = 2 M +
2(1 , S )
where the throughput S = . In order to evaluate the limit1
2(1
(9)
3 1 , M3 + M2 2
+(` , 1)(` , 2);
`p (` , 1) ,
2 + (`p+1
,p)
M:
Lq = G0 (1) , =
n
(8)
MTs ` + (MM,1) , 1
T
s
:
Wq = ` Lq =
2(1 , )
g000 (1)
Xi iid
B (1; );
i=1
Xi Yi ;
M
X
2 (` , 1)
+3
1
, M1
300
250
200
150
100
50
0
1
0.5
p
0.2
0.4
0.6
0.8
Throughput
That is,
g (1) , g (1)
Lq [modeli ] , Lq [modelj ] = i 2 , 2j ;
300
00
250
200
150
100
50
0
1
0.5
0
10
Throughput
g000 (1)
00
2
p`(` , 1)(M , 1)
3 1 , M3 + M2 2 + 3 (1
, p + p`)M
p`(` , 1)(` , 2) ;
+
1 , p + p`
5. Analytical Comparisons
Burstiness of the arrival traffic adversely effects the behavior of almost all communication protocols. There are
several definitions which attempt to quantify the burstiness
of traffic. Among these include the coefficient of variation
of the interarrival times of the packets, the square of the
coefficient of variation etc. [1]. However, there is no commonly accepted definition for the burstiness of the traffic.
Instead of defining the burstiness of the traffic, one should
ask the question which aspect of the traffic distribution adversely effects the delay of the TDMA protocol. Identifying such factor(s) allow us to accurately predict the performance of the protocol. To do this, we provide some comparisons of the delay Wq as a function of the throughput S .
Since Wq = Tf Lq = and will be kept the same for different traffic distributions, it suffices to compare Lq for various
traffic distributions, where
Model
1
2
3
4
5
g00 (1)
2
2 + 2
1, 1 , M
2 + 2
1 , 1
2
2 + (` , 1) , M
`p , 2
2 + (` , 1) `p+1
,p M
2
In order to compare the performance of the TDMA protocol for various traffic distributions, we assume that the
average number of packets arriving during a frame time is
the same. As Table 5 indicates, for almost all traffic scenarios studied here, the classical Poisson process assumption
on the arrival of packets (Model 1) will underestimate the
delay performance of the TDMA protocol. By making
small in models 2 and 3, or making ` large in models 4 and
5, we can make g 00 (1) arbitrary large without effecting the
average traffic, . This clearly shows that the TDMA protocol can perform poorly at any average traffic arrival rate
when large size messages are allowed.
The above approach indicates that the average number
of packets arriving during a frame time is an important parameter of the traffic. However, alone this parameter fails to
capture the `burstiness' of the traffic. For predicting the average performance of the TDMA protocol, one needs both,
the first factorial moment (g 0 (1) = = S ) as well as the
second factorial moment (g 00 (1)) of the number of packets
that arrive during a frame time. Figure 6 illustrates the effect
of g 00 (1) on the delay performance of the TDMA protocols.
Besides the average performance of the TDMA protocol,
it is important to quantify the extreme behavior of the queue
size. For this purpose one needs to evaluate the standard
deviation and then provide confidence interval bands around
the average performance line. Instead of focusing on the
delay, we equivalently focus on the queue size since it is
algebraically a bit easier to manipulate.
The standard deviation of the models studied in this paper are compared in Figure 7. Equation (1) shows that the
variance is effected by g 00 (1) as well as g 000 (1). The magnitude of the effect of these two factorial moments is illustrated in Figure 7. Note that the usual Poisson process
model (Model 1) may fail to predict the standard deviation
500
Model 1
Model 2, =0.1
Model 3, =0.1
Model 4, l=10
Model 5, p=0.1, l=10
450
400
Normalized Delay
350
300
6. Simulation Results
250
200
150
100
50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Throughput
of the queue size when the arriving traffic cannot be described by a Poisson process.
Source
Rate
Model 1
Model 2, =0.7
Model 3, =0.7
Model 4, l=10
Model 5, p=0.1, l=10
45
40
No. of packets
35
30
25
20
15
10
5
0
0
0.1
0.2
0.3
0.4
0.5
Throughput
0.6
0.7
0.8
0.9
6
0.116
The simulation was performed based on three parameters: 1) the number of stations, the round trip propagation
delay between two Martian surface stations (0.113 sec), and
the channel speed (400 Mbps). Other parameters have been
calculated accordingly to obtain the maximum utilization.
The arrival process for each station were taken to be geometric with geometrically distributed message lengths. The
average message lengths were assumed to be 1, 5, and 10
packets per message.
The delay versus throughput as well as the buffer performance were measured for different traffic loads. Figure
8 illustrates the average delay of the six stations for the
TDMA protocol. It clearly shows that the larger average
message length has adverse effect on the overall network
performance. This agrees with the analytical results of the
previous sections.
The analytical models of the earlier sections provide
closed form expressions for the average queue size, the average delay, and the standard deviation of the queue size.
However, they fail to predict the maximum buffer size required to accommodate certain traffic loads. Figure 9 illustrates the simulation results for the average and maximum
buffer sizes using different message lengths. Once again,
we see that the message length distribution has a significant
effect on the maximum buffer size.
10
10
10
Burst=1
Burst=5
Burst=10
0
10
0.18
0.2
0.22
0.24
0.26
0.28
Througput, S
0.3
0.32
0.34
approach shows that the main parameter that effects the average delay of the TDMA protocol is the second factorial
moment of the number of packets arriving during a frame
time. The variance of the queue size, on the other hand, is
effected by the second as well as the third factorial moments
of the number of packets arriving during a frame time. By
estimating these parameters, and using the expressions provided in sections 3 and 4, we may accurately predict the performance of the TDMA protocol for general traffic. Since
the Poisson process can be completely characterized by a
single parameter, it lacks the ability to accurately represent the arrival traffic. The simulation results indicate that
bursty traffic has similar effects on the maximum buffer size
needed to accommodate the traffic.
0.36
References
Figure 8. Throughput vs delay performance
with TDMA.
800
700
600
500
400
300
200
100
0.4
0.35
0.3
0.25
10
0.2
Throughput S
0.15
5
1
Burst
7. Conclusions
In this paper, we described how the G/D/1 model of
queuing theory may be used to study the effect of arbitrary
traffic on the performance of the TDMA protocol. The approach is simple and provides closed form expressions for
the average queue size, the average delay as well as the standard deviation of the queue size.
We also provided five different traffic distributions to
study their effect on the delay and the queue size. For almost all cases, the usual Poisson process model underestimates the queue size as well as the delay. Furthermore, our