Survey
of
Traffic Control
Schemes
and Protocols in ATM Networks
In the past few years, Broadband ISDN (BISDN) has received increased attention as a communication architecture capable of support ing multimedia applications. Among the techniques proposed
to
imple ment BISDN, Asynchronous Transfer Mode (ATM) is considered
to
be the most promising transfer technique because
of
its ejiciency andjle.ri biliiy. In ATM networks. the performunce bottleneck
of
the network, which was once the channel transmission speed, is shifred
to
the processing speed at the network switching nodes
and
the propagation dela
of
the channel. This ship is because the highspeed channel increuses the ratio ofpropagation delay
to
cell
transmission time and the ratio of processing time
to
cell transmission time. Due
to
the increased ratio
of
propagation delay
to
cell transmission time. a large number
of
cells
can be in trunsit between
two
ATM switching nodes. In addition, the increased ratio
qf
processing time
to
cell transmission time makes it difjirult
to
implement hopbyhop control schemes. Therefore,
trufjic
control in ATM networks is a challenge, and new network urchitectures
(jlow
control schemes, error control schemes, etc.) are required in ATM networks. This paper sun>eys number of importunt research topics in ATM net works. The topics covered include marhemutical modeling
of
various
types
of
trafic sources, congestion control and error control schemes for ATM networks, and priori5 schemes to support multiple classes
of
traflc.. Standard activitv for ATM networks and future re5earch problems in ATM are also presented.
I.
INTRODUCTlON
Due to the increased demand for communication services of
all
kinds (e.g., voice, data. and video), Broadband ISDN (BISDN) has received increased attention in the past few years. The key to
a
successful BISDN system is the ability
to
support
a
wide vari ety of traffic and diverse service and performance requirements. BISDN is required to support traffic requiring bandwidth ranging from
a
few kilobits per second (e.g..
a
slow temiinal) to several hundred megabits per second (e.g., moving image data). Some traffic, such
as
interactive data and video. is highly bursty; while some traffic, such
as
large files, is continuous. BISDN is
also
required to meet diverse service and perfomiance requirements of multimedia traffic. Realtime voice, for instance, requires rapid transfer through a network. but the
loss
of
small
amounts of voice information is tolerable. In many data applications, realtime delivery is not of primary importance, but high throughput and Manuscript received March
IO.
1990.
revised Augu5t
16,
1990.
This
work
was
supported
in
part by
the National Science Foundation under Grant NCR8907909 and
by
the University
of
California MICRO Pro The authors are with the Department of Infomiation and Computer Sci
IEEE Log
Number
9040850.
gram.
ence. Univer5ity
of
California. Imine.
CA
92717. strict error control are required. Some services, such
as
realtime video communications, require errorfree transmission as well
as
rapid transfer
[I].
BISDN should
also
be able to facilitate expected (as well
as
unexpected) future services in
a
practical and easily expanded fashion. Examples of expected future services include highdefi nition TV (HDTV). broadband videotex, and videoidocument retrieval services
[2],
3].
To meet the previously stated requirements for
a
successful BISDN, several techinques have been proposed for the switching and multiplexing schemes (“transfer mode”). These schemes include circuitswitching based Synchronous Transfer Mode (STM) and packetswitching based Asynchronous Transfer Mode (ATM). STM, a circuit switching based technique, was initially con sidered an appropriate transfer mode for BISDN because of its compatibility with existing systems. In STM, bandwidth is orga nized in
a
periodic frame, which consists of time slots (Fig.
I @).
A framing slot indicates the start of each frame. As in traditional circuit switching, each slot in an STM frame
is
assigned to
a
particular call, and the call is identified by the position of the slot. In STM, slots are assigned based on the peak transfer rate of the call
so
that the required service quality can be guaranteed even at the peak load. Because of its circuitlike nature, STM
is
suitable for fixedrate services; however, STM cannot support traffic etf ciently since, in STM, bandwidth is wasted during the period in which information is transported below peak rate. ATM eliminatcs the inflexibility and inefficiency found in STM. In ATM, information flow is organized into fixedsize blocks called ‘‘cells,” each consisting of
a
header and an information field. Cells are transmitted over
a
virtual circuit, and routing is performed based on the Virtual Circuit Identifier (VCI) contained in the cell header. The cell transmission time is equal
to
a
slot length, and slots are allocated
to
a
call on demand (Fig. I(b)). ATM’s fundamental difference from STM is that slot assignments are not fixed; instead, the time slots are assigned in an asynchro nous (demandbased) manner. In ATM, therefore, no bandwidth is consumed unless information is actually being transported. Between ATM and STM. ATM is considered to be most prom ising because of its efficiency and flexibility. Because slots
are
allocated
to
services on demand, ATM can easily accommodate variable bit rate services. Moreover, in ATM,
no
bandwidth is consumed unless information is actually being transmitted. ATM can
also
gain bandwidth efficiency by statistically multiplexing
I70
Time Slot
_
eriodic Frame
(a)
Time Slot (Cell)
1
Overhead
Informanon
H Hedder
Fig.
1.
STM and ATM principles. (a) STM Multiplexing
(b)
ATM Multiplexing bursty traffic sources. Since bursty traffic does not require contin uous allocation of the bandwidth at its peak rate, a large number of bursty traffic sources can share the bandwidth. ATM can also support circuitoriented and continuous bitrate services by allo cating bandwidth based on the peak rate (given that sufficient resources are available). Because of these advantages, ATM is considered more suitable for BISDN. This paper therefore focuses on ATM and surveys a number of important research top ics related to ATM networks. The organization of this paper is as foliows. In Section
11,
var ious mathematical models proposed for data, voice and video are surveyed. In Section
111
congestion control schemes suitable for ATM networks are examined. In Section IV, effective error con trol schemes for ATM networks are examined. In Section V, var ious priority schemes proposed to support multiple service classes are discussed. In Section VI, ATM standardization activities are presented. In Section
VII,
a summary of this paper is given, and possible future research problems are discussed. Finally, in Sec tion VIII, brief concluding remarks are given.
11.
MODELING
F
TRAFFIC
OURCES
As
mentioned earlier, ATM networks must support various communications services, such as data, voice, and video, each having different traffic characteristics. To evaluate the perfor mance of such networks, accurate source modeling is required. The purpose of this section
is
to examine several traffic models proposed for data, voice, and video sources. The various math ematical models described below have been examined against actual measured data, and their accuracy has been validated.
A.
Input Trafic
Models
for Data Sources
It is wellknown that generation of data from a single data source is well characterized by a Poisson arrival process (contin uous time case)
or
by
a
geometric interarrival process (discrete time case). For interactive data transmission,
a
single cell may be generated at a time. For
a
bulk data transmission, such
as
a file transfer, a large number of cells may be generated at
a
time (batch arrivals). In existing packet networks, packets could be either of variable
or
constant length. In ATM networks, however, the cell size is fixed. Furthermore, because the size of
a
cell is relatively short compared to the length of a packet in existing networks, multiple cells may be created from one data packet.
BAE
AND
SUDA SCHEMES
AND
PROTOCOLS
IN ATM
NETWORKS
~ ~~
B.
Input Trafic
Models
for
Voice
Sources
An arrival process of cells from
a
voice source (and a video source) is fairly complex due to the strong correlation among arrivals. In this subsection, input traffic models proposed for
a
voice source are examined. The arrival process of new voice calls and the distribution of their durations can be characterized by a Poisson process and by an exponential distribution, respectively. Within a call, talkspurts and silent periods alternate. During talkspurts, voice cells are generated periodically; during silent periods, no cells are gener ated. The correlated generation of voice cells within a
call
can be modeled by an Interrupted Poisson Process (IPP)
[4][8].
In an IPP model, each voice source is characterized by
ON
(correspond ing to talkspurt) and
OFF
(corresponding to silence duration) periods, which appear in turn. The transition from
ON
to
OFF
occurs with the probability
6,
and the transition from
OFF
to
ON
occurs with the probability
a.
In a discrete time case,
ON
and
OFF
periods are geometrically distributed with the mean 1
p
and
1
/CY,
respectively. Cells are generated during the
ON
period according
to
a Bernoulli distribution with the rate
A;
no cell is generated during the
OFF
period (Fig.
2).
(The continuous time analog is an exponential distribution using a Poisson process.)
Fig.
2.
IPP
model. When
N
independent voice sources are multiplexed, aggre gated cell amvals are governed by the number of voice sources in the
ON
state. Assuming a discrete time system, the probability
P,,
that
n
out of
N
voice sources are in the
ON
state
(n
voice cell arrivals in
a
slot) is given by The continuous time analog represents the number of voice sources in the
ON
state as a birthdeath process with birth rate
X
(
n
)
and death rate
p
(n
,
where
h(n)
=
(N
n)a,
p(n)
=
np,
for0
5
n
5
N.
(2)
171
Na
(~1)a
2a
a
+if3C
'
zr3
P
28
W1)B
NP
Fig.
3.
Birthdeath model
for
the number
of
active voice
sources.
Figure 3 shows the birthdeath model. For this continuous time case, the probability
P,
that
n
out of
N
voice sources are
in
the
ON
state is also given by
I)
[6]. Another common approach for modeling aggregate arrivals from
N
voice sources is to use a twostate Markov Modulated Poisson Process (MMPP)
[9],
[lo]. The MMPP is a doubly sto chastic Poisson process where the rate process is determined by the state of a continuoustime Markov chain [9]. In the twostate MMPP model, an aggregate amval process is characterized by two alternating states. It is usually assumed that the duration of each state follows
a
geometrical (discrete time case)
or
an expo nential (continuous time case) distribution, and cell amvals in each state follow a Bemoulli
(or
a Poisson) distribution with dif ferent rates. Therefore, four parameters are necessary to describe an MMPP: the mean duration of each state and the amval rate in each state. Note that an IPP. a process used to describe a single voice source, is
a
special case of the MMPP in which no cell amves during an OFF period. To determine the value of these four parameters, the following MMPP statistical characteristics are matched with the measured data [9]: 1) the mean arrival rate; 2) the variancetomean ratio of the number of arrivals in a time interval
(0,
t,
1;
3) the long term variancetomean ratio of the number of amvals;
4)
the third moment of the number of amvals
in
0,
2).
Note that the analytical models described in Sections 11A and 11B can model only constant bit rate traffic. Analytical models which can adequately model variable bit rate traffic are not yet available.
C.
lnput Trajic Models for Video Sources
Video traffic requires large bandwidth.
For
instance, in TV applications a frame of 512
X
512 resolution is transmitted every 1/30 second, generating 512
X
512
X
8
x
30
bits per second (approximately 63 Mbits/s), if a simple PCM coding scheme
is
used. Therefore, video sources are usually compressed by using an interframe variablerate coding scheme which encodes only significant differences between successive frames. This intro duces
a
strong correlation among cell amvals from successive frames. Like a voice source, a video source generates correlated cell amvals; however, its statistical nature is quite different from a voice source. Two types of correlations are evident in the cell generation process of a video source: shortterm correlation and longterm correlation. Shortterm correlation corresponds to uni form activity levels (i.e., small fluctuations
in
bit rates), and its effects last for a very
short
period of time
(on
the order
of
a few hundred milliseconds). Longterm correlation corresponds to sud den scene changes, which cause a large rate of amvals, and its effects last
for
a relatively long period of time (on the order of a few seconds) [l
11
In
Section 11CI), models which consider only shortterm correlation (i.e., models for video sources without scene changes) are examined.
In
Section 11C2). models which consider both shortterm and longterm correlation (i.e., models for video sources with scene changes) are examined.
I)
Models for Video Sources Without Scene Changes:
In this section, models proposed for video sources
without
scene changes are examined. These models are applicable to video scenes with relatively uniform activity levels such as videotele phone scenes showing a person talking. Two models have been proposed. The first model approximates a video source by an autoregressive (AR) process [12], [13]. This model describes the cell generation process of a video source quite accurately. How ever, because of its complexity, queueing analysis based on this model is very complicated and may not be tractable. This model is more suitable for use in simulations. The second model approx imates a video source
or
video sources) by a discretestate Mar kov model [13]. This model is more tractable in queueing anal ysis than the first model, and yet describes the cell generation process of
a
video source
(or
video sources) well.
a)
Model
A:
Continuousstate
AR
Markov Model
[13]:
Here, a single video source is approximated by an autoregressive (AR) process. The definition of an AR process is
as
follows:
M
h(n)
=
c
a,h(n

m
+
bw(n)
(3)
m=l
where
h(n)
epresents the source bit rate during the nth frame;
M
is the model order;
w
(n
is
a
Gaussian random process; and
a,(m
=
1,
2,
.,
M)
and
b
are coefficients. It
is
shown that the firstorder autoregressive Markov model
h n)
=
a,X n
1)
+
bw(n)
(4)
is sufficient for engineering purposes. Assuming that
w(n)
has the mean
7
and the variance
1,
and that
I
a,
is less than
I,
the values
of
coefficients
a,
and
b
are determined by matching the steadystate average
E(
A)
and discrete autocovariance
C
n)
of the AR process with the measured data.
E X)
nd
C(n)
of
the
AR
process in
(4)
are given by
[14]
b
b2
E(h)
=
C(n)
=
4
Z
2
0.
5)
This model provides a rather accurate approximation of the bit rate of a single video source without scene changes. However, as stated above, analysis
of
a queueing model with the above amval process can be very complex and may not be tractable; therefore, this model is suitable for use in simulations.
b) Model
B:
Discretestate, continuoustime Markov Pro cess
[13]:
The process
X t)
describing the bit rate of a video source at time
I
is a continuoustime, continuousstate process. In this model, process
h(t)
is sampled at random Poisson time instances and the states are quantized at these points (Fig.
4 .
n
1
a
time
Fig.
4.
Poisson sampling
and
quantization
of
the source rate
112
PROCEEDINGS
OF
THE
1tF.E.
VOL
71).
NO
?.
FEBRUARY
1Y91
other words, the process
h(t)
s approximated by a continuous time process
x(t)
with discrete jumps at random Poisson times. This approximation can be improved by decreasing the quanti zation step
A
and increasing the sampling rate. The state transition diagram
x(t)
s shown in Fig.
5.
The pro
Ma
(M1)
a
Fig.
5.
State transition diagramModel
B
cess
x(t)
an be used to describe a single source, as well as an aggregation of several sources. The aggregated amval process from
N
video sources can transit between
M
+
1
levels. The label in each state indicates the data rate in that state
A
s a constant). To determine values of the quantization step
A
and the transition rates
a
and
0,
he steadystate mean
E(&),
variance
EN
0)
and autocovariance function
cN
7)
of the process
x
t)
(describing an aggregate of
N
independent sources) are matched with the measureddata.
(7isatimeparameter.)E(XN),
c,(O)
andFN(7) are given by
E(X,)
=
MAL
ff
+
0
The number of quantization levels
M
is chosen arbitrarily, but it should be large enough to cover all likely bit rates. The process in Fig.
5
can be decomposed into a superposition of simpler processes. It can be thought of as a superposition of
M
independent identical
ONOFF
minisources, each being mod eled as in Fig.
6.
Each minisource altemates between
ON
and
OFF
a
Fig.
6.
Minisource model. states. The transition from
ON
to
OFF
state occurs with the rate
0
and the transition from
OFF
to
ON
state occurs with rate
a
Thus both
ON
and
OFF
periods are exponentially distributed.) The data rate of a minisource in the
ON
state is
A;
a minisource does not generate bits during the
OFF
state (data rate is
0).
(Note that in Fig.
5
a label associated with the state represents the data rate
of
a minisource in the state.) The state of the aggregated arrival process can thus be represented by the number of minisources which are in the
ON
state.
2) Models for Video Sources with Scene Changes:
In this sec tion, models proposed for video sources
with
scene changes are examined. These models capture both shortterm and longterm correlations explained at the beginning of Section 11C and thus these models are suitable to describe a cell generation process from video scenes with sudden changes, such as videotelephone scenes showing changes between listener and talker modes,
or
scene changes in broadcast TV
[l
I].
Two models have been
pro
posed: the first model is an extension
of
Model
B
explained above; the second model approximates a video source by the discrete state continuoustime Markov process (Model B) with batch amvals.
a) Model
C:
An
extension of Model
B
1111:
The state tran sition diagram of the cell generation process from an aggregation of
N
video sources is shown in Fig.
7.
(This process can also be
2
PIb
I
4
It
Fig.
7.
model
(with
scene changes). State transition rate diagram for the aggregate source used to describe a single video source with scene changes.) The label in each state indicates the data rate in the state. There are two basic data rate levels: a high data rate
Ah,
which represents a sudden scene change, and a low data rate
A,,
which represents a uniform activity level. If scene changes do not exist (i.e., if we delete the states which contain a high rate
A,,),
the process in Fig.
7
reduces to the one used in Model
B.
The aggregated process of
N
video sources can transit between
(M,
+
1)
(M2
+
1)
levels, where
MI
=
NM,
M2
=
N.
Here,
M
is chosen arbitrarily. To determine the values of system parameters
c
and
d
(the tran sition probabilities between uniform activity level and high activ ity level), the fraction of the time spent in the high activity level
(c/(c
+
d))
nd the average time spent in the high activity level
(
1
/d)
are equated with the actual measured data. To determine the rest of the parameters in the model, i.e., the transition prob abilities within the uniform activity level
(a
and
b),
and the two basic data rates
(A,
and
Ah),
the first and second order statistics are matched with the actual measured data.
As
in Model
B,
the process described in Fig.
7
can be decom posed into a superposition of simpler processes. This process can be thought of as a superposition of
M,
independent identical
ON
OFF
minisources of the type shown in Fig. 8(a) and
M2
of the type
a
r
b
(a)
Fig.
8.
Miniprocess models shown in Fig. 8(b). The state of the aggregated amval process can thus be described as the number
of
each type minisource which is in the
ON
state.
b)
Model
D: Discretestate continuoustime Markov Pro cess with batch arrivals
15,
161:
In this model, the cell arrival process from a single video source
with
scene changes is modeled as a discretestate continuoustime Markov process with batch arrivals. The uniform activity level is represented by a discrete state continuoustime Markov process as in Model
B.
This Mstate
BAE AND
SUDA
SCHEMES
AND
PROTOCOLS
IN
ATM
NETWORKS
~~
~___~
_
173