Download as pdf or txt
Download as pdf or txt
You are on page 1of 21

Modeling and Performance Analysis of

Multihop Packet Radio Networks


FOUAD A. TOBAGI,

FELLOW, IEEE

Invited Paper

The design of packet radio networks involves alarge number of


issues which interact in a very complex fashion. Many of these
pertain to the RF channel and its use, others pertain to theoperational protocols. Clearly, no single modelcan be formulated which
incorporates all the necessary parameters and leads to the optimum solution. The oneessential element which complicatesmatters is that, contrary to point-to-point networks in which each
channel is utilized by a single pair of
nodes, the radio channel in
packet radio networks is a multiarcess broadcast resource: i) in a
given locality determined by radio connectivity, the channel
is
shared by many contendingusers, hence the needfor channel access protocols; ii) radiois a broadcast medium and thus the action
taken by a node
has an effect on theactions taken by neighboring
nodes and their outcome.
Despite the complexity of the problem, there has been significant progress worth reporting on. The work accomplished so far
has been either theanalysis of specificexamples of networks or an
attempt to create models that would be useful in the design of
general networks. The purpose of this paper is to survey the various modeling techniques that
have been used for the performance analysis of packet radio networks, and to discuss the assumptions underlying these models, their scope of applicability,
and someof theresults obtained.

I. INTRODUCTION
Packet Radio is a communications technologywhich applies packet switching to the domain of broadcast radio. It
is an attractive concept because it brings together the advantages of both broadcast communications and packet
switching. The broadcast radio medium is a readily available resource, easily accessible, and particularly suitable
for communication among mobile
users; packet switching
offers the fair and efficient sharing of communications
resources by manycontending users with unpredictableand
bursty demands. A packet radio networkis typically a collection of packet switchingnodes that communicate with
each other via broadcast radio. Each node consists of a
broadcast radio and a digital controller.The broadcast radio acts like the modem andline in wire-based
a
terrestrial
Manuscript received March I O , 1986; revised August 14, 1986.
This work was supportedby the Defense AdvancedResearch Projects Agency (DARPA) under Contract MDA-903-84-K-0249.
The author is with the Department of Electrical Engineering,
Computer Systems Laboratory, Stanford University, Stanford, C A
94305, USA.

PROCEEDINGS OF THE IEEE, VOL. 75,

NO.

1, JANUARY 1987

network, providing the connectivity


between neighboring
radios. Thedigital controller provides the
packet switching
functions, receiving packetsfrom neighboring
radios, making routing decisions, and forwarding packets to the next
radio [26]-[28], [76].
The design of apacket radiosystem involves alarge number of issues which interact in a very complex fashion[43].
Many ofthese pertain to the RF channel and its use, others
pertain to the operationalprotocols. Clearly, no single
model can be formulated which incorporates all the necessary parameters and leads to the optimum solution. It is
often the case that certain design choices are dictated by
physical or technological constraints; others
are driven by
past experience orthe desire for simplicityin system manease
agement. For example, for rapid development and for
of communication among mobile users, it may be advantageous to have all users employ omnidirectionalantennas
and share a single high-speed channel.
(In fact, by providing the available communication bandwidth as a single
channel to be dynamically accessed by the users, the additional benefit ofstatistical load averaging is gained, leading to higher channel efficiency.)
With such considerations,
the design space is somewhat reduced. Additionally, one
may select many of the remaining design variables on intuitive grounds,with the hope
that the design
will work;
however, with a system as complex as this, the end result
is very likely to be uncertain at best.
More important thanever before, performanceanalysis
via mathematical modeling and simulation should be an
integral step of the design process. It is a sensible and inexpensive way bywhich one is able to compare variousdesign alternatives, gain insight into the behavior of packet
radio systems, predicttheirperformance, etc. Unfortunately, even when the system functionality has been simplified into its most essential elements, the modeling of
packet radio networkshas proven to be a formidable task.
What makes packet radio networks
so different from pointto-point networks which we seem to handle fairly adequately? Why
are the existing techniques
devised for pointto-point networksnot readilyapplicabletopacket radiosystems?Theoneessential difference that complicates matters
is that, contrary to point-to-point networks in which each

135

channel is utilized by a single pair of nodes, the channel


in packet radio networks is a multi-access broadcast resource: i) in a given locality determined by radio connectivity,theradiochannel issharedby manycontending users,
hence the need for a channel
access protocol; ii)the channel is a broadcast medium, and thus the action takenby a
device has an effect on the actions taken by neighboring
devices and their outcome.
Despite the complexityof the problem, there has been
significant progress worth reporting upon. The work accomplished so far has been either an attempt to create
models that would be useful in the design of general networks, or theanalysis of specific examples of networks. In
some cases, the effort has permitted a comparison of alternatives pertainingto some specific aspect of thedesign
problem, such as, for example, channel access protocols.
But overall, the efforthas not yet ledto a true networkdesign methodology, and agreat deal of work remains to be
done. The purpose of this paper is to survey the various
modeling techniques that have been used for the performance analysisof packet radio networks.
We shall discuss
the modeling assumptions underlying
these models, their
scope of applicability, and the results obtained. Throughout the paper, particular attention will be givento the appropriateness of the modeling assumptionsmade in each
case in view ofthe intended objective.
This is important in
order to clearly identify theusefulness of the modelingeffort accomplished. Indeed, we recall that': " A theory has
only the alternativeof being right or
wrong. A model has a
third possibility: i t might be rightbut irrelevant." (No discussion, however, will be provided of network design and
a
of studies
optimization; in [39]the reader will find survey
related to spatial reuseof channel bandwidthin packet radio networks and to topological design.)
The content of the paperis as follows. In Section I I , we
provide some background to assist the reader in understanding the problems underlying the
analysis of multihop
packet-switched networks. We first review the work accomplished in point-to-point wire
networks, and then discuss briefly the characteristics of packet radio networks
which renderthese different and more complex.In Section
I l l , we address in more detail two main aspects of packet
radio operation which are particularly relevant to performance modeling. The first pertains to channel signaling
(narrow-band and spread-spectrum)
and the resulting conditions for the correct receptionpackets.
of
The second refers to channel access. Wedescribe a number ofaccess protocolssuitablefornarrow-bandandspread-spectrum
systems. Sections IV-VI deal with performance modeling
and analysis. The performance measures of interest are i)
networkcapacity(defined asthe maximum throughputthat
the networkcan support fora given traffic requirement pattern) and ii) the network's throughput-delay characteristics. As the models used to derive these two performance
measures tend to be of differenttypes (and more complex
for thelatter than the former),we discuss thesetwo issues
separately. We addressthe networkcapacity in Section IV.
We begin there by examining special
the
case of single-hop
fully connected networks;these aresimpler to analyze and
have thus been the subjectof many studies; their models

are important as they constitute the groundwork for the


analysis of multiple-access protocols, and multihop packet
radio networkswith general topologies.We then consider
the problem of deriving network
capacity for networkswith
general topologies. We present a general Markovian model
which can accommodatethe many different combinations
of channel signaling characteristics and channel access
protocols, and which has been used to compare various
such alternatives. In Section V, we consider the problem
of deriving the throughput-delay performance aofpacket
radio network.We review various methods
used in solving
this problem. We present
some approacheswhere models
are analyzed exactly and aretypically usefulto investigate
small or specially constructed networks.We then discuss
some approximate analyses suitable for more general topologies. In Section VI, we discuss the role of computer
simulation in relaxing many of the simplifying assumptions
introduced in the various mathematical models, and discuss their own limitations. Section Vlll concludes the paper.

II.

BACKGROUND

To best understand the task of modeling the performance of amultihop packet-switching networkin general,
and to appreciatethe complexity underlying thatof packet
radio networksin particular, we begin by reviewing briefly
the work done on point-to-point packet switching networks, such as ARPANET, and then discuss those features
of packet radio networkswhich are responsible forthe difficulty in analyzing them.
A. Modeling Point-to-PointPacket Switching Networks

A packet switching network


is a complex network of
queues towhich there is no tractable solution. Accordingly,
many simplifications are introduced, reducing the view of
the network to merely a simple abstraction.In thecase of
ARPANET, for example, the network was stripped of all its
operational protocols(link level, routing, flow control,
etc.),
so that, with theassumption ofa known fixed routing, the
only network aspect that is modeled is that of queueing
packets in intermediate nodesfor the appropriate
links, and
then transmitting themon these links when thelatter beis no
come free. (Note thatin point-to-point networks there
contention on the links in the network, as only one node
pair may transmit and receive on any given directed link;
furthermore, each node is equipped withas many modems/
transceivers as it has outgoing links so that transmissions
on separate links are totally independent). With the addition of other simplifyingassumptions, suchas infinite buffers, exponential packet length,Poisson packet generation
at the sources,and independencebetweeninterarrival
times and service times (Kleinrock's celebrated independence assumption [31], [36]), the model is a network of
queues of the Jacksonian type.
The network is considered to be consisting of N nodes
and M channels numbered 1 , 2 , . . . , M. Let yjk denote the
average number ofpackets per unit time to
be transmitted
from source j to destination k
N

'From A Selection of Scientific Quotations, collected


Mackay (M. Ebison, Ed.).

136

by A. L.

e j = l k = l yjyir;

PROCEEDINGS OF THE IEEE, VOL. 75, NO. 1, JANUARY 1987

and
A

ff/k

= *fjk/Y.

The matrix [ f f j k ] represents thetraffic pattern, and y, the network throughput. Let Ci denote the capacity of channel i
in bits per second and let X; denote the mean traffic rate
in packets per second that flows on channel i.Given [ y j k ]
and the routing scheme, the vector { A,}y=, is easily computed. Let I/p denote the mean packet length, and nj, the
i.For a Jacknumber of packetswaitingfor or usingchannel
sonian network, the vector (nl, . * * , nM)is a complete state
description; moreover, the state probability is given by Rn,)
P(n2) . . . P(nM),where P(nj) is the probability of finding nj
in an M/M/l queue with arrival rate Xi and service rate l/pCi.
From this resuIt,theexpected end-to-end packetdelayaveraged over all user traffic is easily derived as [36]
M

T =

,x[-].
y pcj 1- hj

1=1

Equation (1) is important in that it provides the network


[ f f j k ] .It also
throughput-delay tradeoff for the traffic pattern
provides the network capacity for the pattern [ f f / k ] ;this is
simply the value of y at which the delay becomes infinite;
or thevalue of y for which thechannel with themaximum
Xi/pCi, say channel io,
becomes saturated (i.e., when X;, =
pCi,). Furthermore, (1)proved to be extremely useful
in networkoptimization. Indeed, it represents anobjective function which has the important property of beingseparable
inthedesignvariables {Ci,hj)E1.Thiswastakenadvantage
of immenselyin the design of point-to-point networks
[36],
1241.

B. Packet Radio Network Features


A major designissue in packet radio networksis the managment of the RF bandwidth. As pointed out in the Introduction, to satisfy communication among mobile users, and
to gain higher bandwidth efficiency, it is advantageous to
consider the entire network be
to operatingon a single radio channel t o be shared by all nodes. The sharing canbe
done either in the time
domain, or in thecode domain; or
a combination of both. For simplicity in presentation and
discussion, we limitourselves in this paper only to the
task
of modeling such networks.
Even whenmodelingsimplifications
similar to those
made for point-to-point networksare made,the analysis of
packet radio networksremains complex. This is due to the
nature of wave propagation in radio channels, and to the
multi-access/broadcastnatureof these networks. First,connectivity betweentwo nodes is not a Boolean function. Depending on many factors, such as the radio frequency in
use, power of the transmitters, terrain, antenna type and
orientation, and the distance separating the two nodes,
there are cases when two nodes are, for all practical purposes, considered disconnected; but when radio connectivity exists, the quality of the radio linksamong nodes i s
not thesame for all links andat all times, and depends on
the factors listed above as well as on thedynamic activity
in the network.Thus, the topology of the network
is not
as clearlydefined as with point-to-point wire
networks, and
this fact is rendered even more complex when nodes are
mobile.

TOBACI: MULTIHOP PACKET RADIO NETWORKS

Secondly, in a given locality (as determined by radio connectivity), the radio channelis a server which is contended
upon and shared by the many nodes in that locality. The
success of a transmission undertaken by a node depends
not only on the quality of radio
the link between the node
and its intended neighbor, but also on the particular topology ofthe network in the
locality, transmission activity
of other nodes, state of the receiving node, the signaling
method used, etc. Consequently, essential
to the
operation
of a radio networkis a policy accordingto whichnodes determine their rightto access the channel. The objective of
such a policy is to improve the likelihood of success of a
transmission to beundertaken, and thus maximize the
overall networkthroughput. Thus in packet radionetworks, the decision as to whether a node should transmit
or not may be a function of the state of the network in its
neighborhood (as dictated by the access protocol); and once
the transmission has been initiated, the outcome is dependent on the state of the neighboring nodes and their
evolution during thetransmission time in question. Thus
we are in presence of a network of queues in which a packsysets service time is dependent on thestate of the entire
tem and its evolution in time. This i s indeed a complex
queueing problem to which nosimple solution exists.
The conditions underwhich atransmitted
packet may be
correctly received and the channel access protocols used
(also called link activation protocols)are so essential to the
performance of packet radio networks and so central t o
every modeling effort reported upon, thatwe devote Section Ill to a clear description of these issues.
In thesequel, unless otherwise stated, we shall consider
a packet radio network to be consisting of N nodes distributed over some geographical area. Theconnectivity of
the networkis considered to be a known Boolean function
and specified by an N X N hearing matrix H = [h,,], where
h , = 1 if j can hear i,and 0 otherwise. Thus each nonzero
to a directed raentry h, in thehearing matrix corresponds
dio link in the network fromito
node
nodej , and viceversa.
We denote such a link by (i,j ) , and call node i i t s source
and node j its destination. Alternatively, a network can be
represented graphically by a directed graph in which vertices represent nodes in the network, and directed edges
represent one-way connectivity. A link is said to be active
whenever a transmission is taking place over that link; i.e.,
whenever the source node i s transmitting a message destined to thedestination node on that link.
Ill. PACKET RADIO NETWORK
OPERATION

The operation and hence performance of a packet radio


network differ accordingto thechannel signaling method
used:narrow-band or spread-spectrum. Indeed, the two
methods result in different capture effects, whereby capture refers to the ability ofa receiver to receive correctly
a packet despite the presence of other time-overlapping
packets, and thus different conditions for the correct reception of a transmitted packet. Another aspect by which
the operation and performance ofa packet radio network
differ i s the choice of channelaccess protocol. Inthis section, we first define the two above mentioned signaling
methods andthe corresponding captureeffects, and then
describe the various channel access protocols that can be
used with each method.

137

A. Channel
Signaling
-

and Capture

Two types of channel signaling exist: narrow-band signaling and spread-spectrumsignaling. In narrow-band signaling, databits are modulated directlyon the carrier.
Thus
the overlap at some receiver of two packets with roughly
the same signal power typically results in theloss of both
packets. If the two overlapping
packets areof distinctly different signal power, thensome form of powercapturemay
come into effect, whereby the strongerpacket is received
correctly, and the weaker one is lost. (This, however, sometimes depends on timing as a weak signal may tie up the
synchronization circuit at the receiver.) Considering a networkwithafixedtopologyinwhichalllinksareofthesame
quality, (thus removing the possibility of power capture),
we then say that the network operates in a zero-capture
mode: the overlap of
two ormore packets at some receiver
results in the destruction of all. (Note that, on the other
to detect
hand, in narrow-band systems, it is easy for a node
activitydue to other nodes within radio connectivity simply
by sensing carrier.)
Spread-spectrum refers to signaling schemes which are
based on some form of coding
and whichuse a much wider
bandwidth than do narrow-band schemesforthe
same data
rate [63].One way of achieving spread-spectrum operation
consists of using directsequence pseudo-noise (PN) modulation. Another is frequency hopping (FH). Spread-spectrum techniques have been introduced to combat multipath effectsand jamming; but they also offerspecific
advantages in multiuser interference which center around
two properties: code-division and time-capture. Codedivision refers to the fact that transmissionswith orthogonal
codes may overlap in time with little or no
effect on each
other. Time-capture refers
to the ability ofan idle receiver
to successfully receive a packet with agiven code despite
the presence of other time-overlapping transmissions
with
the same code (or othercodes) [21], [28],
[49]. Packet radio
networks may take advantage of these two properties in
different ways as described in the following.
Since the transmission ofpackets is asynchronous, each
packet transmission mustbe preceded by the transmission
of a preamblewhich the receivers use to acquire bit synchronization. Often, the preamble
consists of aknown code
is used
with strongautocorrelationpropertieswhich
throughout the network, and which idle receivers constantly search for. Furthermore, this code is chosen to be
orthogonal to those used in encoding thepackets. In order
to be able to receive a packet, an idle receiver must first
process successfully the preamble. Failures are primarily
due to overlaps in time with another packets preamble;
they may also be due to the presence of other ongoing
transmissions, which act as background noise and thus
reduce the effective signal-to-noise
ratio. Errors due to overlapping preambles can be reduced by the use of receiverdirected codes. In this mode of operation,each receiver is
assigned a distinct preamble code waveform, and transmitters must use the waveformassigned to the receiver they
are trying tocommunicate with in thetransmission of the
preambles. The result is that a receiver will process the
preamble of only those packets that are specifically destined to it, all other transmissions by neighboring transmitters, includingtheir preambles, appearing as background noise. We shall refer to the two above-mentioned

138

preamble
code
assignment methods as space-homogeneous and receiver-directed, respectively. Havingsuccessfully processed the preamble of an incoming packet, the
receiver at a nodewill thus lockonto the
data portion of the
packet until its transmission is completed.
is locked onto byareceiver, it is important
Once a packet
that future overlapping packets do not interfere with the
completion of its successful reception. The conditions for
such a success are, here too, dependent on the ways the
PN waveform patterns are selected and used, in addition
to theset of active linksand their signal strengths. Several
cases can be identified. Consider first thecase in which a
fixed PN spread-spectrum pattern identical for
each bit
transmitted throughout the network is used. An overlap
ping packet would not interfere with
a packet lockedonto
earlier as long as its autocorrelation peaks do not coincide
with those of the earlier packet
(see Fig.1).This meansthat

DesiredSignal

Inteferlng Slgnal

Sum

Fig. 1. Effect of interference with no multipath in a PN


spread-spectrum network. (Tis the bit duration; n is a measure of the degreeof spread and is given by 2TW, where
W is the RF bandwidth [20].)

there is a vulnerable period of a few chip times foreach bit


during which thearrival of a new packetwould cause destruction. (Note, however, that the cumulative vulnerable
period for a packet remains small compared to its entire
transmission time, hence the capture effect. Clearly this
represents the ideal picture when there is no multipath;
multipath has the effect of smearing the correlation
peaks,
spreading them over time, and causing intersymbol interferences. Such effects are ignored in this paper.) We shall
refer to this case as the space-and-bit-homogeneous code
assignment. To reduce the effect of overlapping packets,
one could assign distinct orthogonal code waveforms to
different receivers, and have the transmittersuse the waveform assigned to the intendedreceiver in coding their
bits.
This mode reduces the amount of
traffic that interfereswith
a packet transmitted to a receiverto only thattraffic which
is destined forthat particular receiver. We shall
refer to this
case asthe receiver-directed bit-homogeneous code
assignment.Theelimination of interference from
any overlapping
packets canbe approached by using a pattern
which varies
on a bit-by-bit
basis, and equipping thereceiver with a programmable matched filter which follows the pattern as it
varies from bit to bit. If the
pattern is long enough so that
it does not repeat itself during the
transmission ofthe longest packet, then anyoverlapping packetwould not interfere
with thepacket locked ontoearlier, since it would notproduce any autocorrelation peak during the entire
reception

PROCEEDINGS OF THE IEEE, VOL. 75, NO. 1, JANUARY 1987

time of the earlier packet. (The problem of packet acquisition incurred with arrival times too close to each other
would persist, since one would require the PN sequence
to start from thesame point, known to thereceivers. Note
that the startingpoint can be thesame for all receivers, or
different for each receiver.) Werefer to this scheme as the
bit-by-bit code changingscheme or bit-nonhomogeneous
code assignment. Assuming that backgroundnoise is negligible andall codes aretruly orthogonal, this mode ofo p
eration achieves perfect capture, whereby correct recep
tion is guaranteed oncethe packet is locked onto. Perfect
capture can also be achieved by assigning distinct orthogonal codes to nodes which the latter use to encode their
packets when transmitting.In this case, the preamble contains information on the spreading waveform used, thus
allowing the receiver to program
its matched filter accordingly. Note that, in this case, perfect capture does not necessitate bit-by-bit code changing as long as transmitters
are assignedorthogonal codes. We refer to this scheme as
the transmitter-directed codeassignment.
As true orthogonality ofcodes is not always guaranteed
on a bit-wise basis, and the background noise cannot always be negligible, some form of encoding (e.g., convolutional coding) is performed on the information
to be
transmitted prior tospread-spectrum coding,so that, with
the use of a decoder (e.g., sequential decoder), the correction of bit or symbol errors
can bedone at the receiving
end. Consequently, the correct reception
of apacket is considered to be achieved if the decoder does not make any
error in decoding the entirepacket. Perfect capture refers
then to the ideal case whereby, once locked on, a packet
is always successfully received.

any node i, let X ( i )denote the set of all nodes connected


to it including itself. Let X * ( / )4 X(;) - { i } . The elements
nodes
of X *(i)are called neighbors ofi. For a collection of
A, we let X(A) 2 U,,,X(i) and X2(A) P X(X(A)).A transmission over link ( i , / ) is denoted by 3 ( ( i , j ) ) . Given 3 ( ( i ,
j ) ) , a nodek is said to be hidden withrespect to the transmitting node i if k is a neighbor of the destination node j
but not of the source node i ; i.e., if k E X ( / ) - X(;).
For all the protocols described below, it is considered
that nodes attempt transmission of their
packets at random
points in time defined bysome point process. In the event
that a packet scheduled for transmission
is inhibited by the
or in theevent that a transmitted
operation of the protocol,
packet is not received correctly, then that packet is again
considered for transmission at some future point in time
determined by the scheduling point process.
1) PureALOHA: In this scheme, a node is allowed to
transmit only if it is not already transmitting, regardless of
any other activityin the network.It was first introducedby
Abramson for the single-hop narrow-band
ALOHA System
which uses separate channels for inbound and outbound
traffic [I]. When appliedto multihoppacket radio networks
with a single broadcast channel, thedefinition implies that
the reception of a packet by a node is aborted if a packet
transmission is scheduled at that node during the timeof
reception, i.e., transmission has priority over reception. In
systems where nodes receive packets indiscriminantly,
(e.g., narrow-band systems or spread-spectrumsystems in
which the codeassignment usedallows an idle receiver to
lock onto any incoming packet), then aborting the reception of a packet may be beneficial since the packet may be
destined to another node. If, on the other hand, the only
packets that can be received are destined to the node, (as
B. Channel Access Protocols
would possibly be thecase in spread-spectrum systemswith
some code assignments), then abortingan ongoing recepA channelaccess protocol defines the conditions under
tion results in thrashing. The following scheme overcomes
which a node may accessthe channel.For someprootcols,
this problem.
theseconditionsareexpressed intermsof notonlythestate
2) Disciplined ALOHA: A node may transmit as long as
of the nodein question, but also the state of other nodes
in the network.For practically realizable protocols, the rules it is not already transmitting nor locked onto any packet.
This schemeis certainly ideal for spread-spectrum
systems
embodied in theaccess protocol are constrained to be dein
which
receivers
lock
onlyonto
packets
destined
to
them.
fined only in terms of information that
can be acquired loFor
other
systems,
this
scheme
achieves
some
limited
form
cally at the node, such as the state of the receiver at that
of activity sensing which might lead to improved perfornode, and the state of the transmitters in some neighbormance; indeed, if a node is locked onto a packet not deson the
hood ofit. Thus the feasibility of a protocol depends
tined
to it, then it is likely that there exists a neighboring
ability of a node to dynamically acquire knowledge renode
to which the packet is destined, and it may be adgarding thestate of othernodes, which in turn
depends on
visable to inhibit transmission, thus increasing the probthe particular channel signaling scheme and code-assignability of correct reception of the earlier
packet.
ment scheme in use. Furthermore, the choiceof an access
3)
Slotted
ALOHA
[2],
[3],
[32],
[Sl],
1521:
The time axis is
in
protocol is heavily influenced by the capture properties
considered
to
be
universal
for
all
nodes
and
divided into
effect, since certain actions may be desirable under one
equally sized slots. A node with a packet scheduled for
capture mode but undesirable under others.2
transmission in a particular slot transmits thatpacket, synFor simplicity in presentation anddiscussion, we assume
chronizing the start of transmission to coincide with the
atthisstagethatthehearingmatrixHissymmetric,andthus
beginning of the slot.3 It is assumed that the packet transall edges in thegraph representationare bidirectional. For
mission fits entirely within the slot. Since the slot size is

'When different codes are assigned to nodes (e.g., receiverdirected, transmitterdirected, etc.), the termCode-Division Multiple
Access (CDMA) is generally employed. However, the termCDMA
in itself is not sufficient to describe completely the operation of
a network as it does not specify any particular channel access protocol. To be precise, one must indicate, in addition to the codeassignment scheme in use, a channel access protocol whichnodes
follow in transmitting their packets.

TOBAGI: M U L T I H O P PACKET RADIO NETWORKS

'Note that a synchronous scheme such as slotted ALOHA makes


good sense in a satellite network where synchronization to a universal time axis is easily accomplished; it does, however, present
implementation difficulties in a multihop packet radio network,
especially if nodes are mobile. It is nevertheless an interesting
scheme to consider because it often lends itself to tractable performance models.

139

channel traffic is desirable in view of improving anti-jam


fixed, slotted ALOHA is best suited for fixed-sizepackets.
capabilities [28].
Clearly, in narrow-band systems, thearrival ofseveral pack5) Busy Tone Multiple Access (BTMA) [62], [67l,VI:The
ets in thesame slot at a receiver results in the destruction
problem of collisions caused by hidden nodes can be alof all. The same effect results in spread-spectrum systems
leviated by the use of a busy tone on a separate channel,
with certain code-assignment policies. This is the case, for
which is emitted by a nodeto indicate that it is currently
example,when all preamblesuse the same code, sincethen
receiving apacket. Thebusytone is then used to inhibit the
slot synchronization would always cause overlap among
receiving node's neighbors from transmitting and thereby
preambles in a given slot. Some improvement can be obinterfering withit. Consider firstits use in narrow-band systained with receiver-directed preamble code assignment,
tems. Several variants of BTMA exist dependingon which
but transmissions destined to the same receiver still interset of nodes transmit abusy tone in any given situation.In
fere with each other. Additional improvement can be obConservative BTMA (C-BTMA),
whenever a node
senses cartained by considering a slot size slightly larger than the
it emits abusy
rier (due to the presence of a transmission),
packet transmission time, and randomizing the start of
tone. Then any node that wishes to transmit is allowed to
transmissions within aslot so as to achieve some degreeof
do so only ifit is not already transmitting and no busy tone
time-capture [14]. In such a case, it is conceivable that the
is sensed. Accordingly, given that a transmission 3((i, j ) )
firstpackettoarrivetoareceiverinagivenslotwiIIbelocked
has been undertaken, a one-hop propagation delay cononto if later packets happened to be delayed by a suffistitutes its vulnerable period as far as transmissions from
ciently long periods of time. Beyond that, the conditions
nodes in 'X*(;) f l X ( / )are concerned, anda two-hop p r o p
for correctly receiving
the entire packet
will depend on the
agation delay as far as transmissions from nodes in X ( j )
code assignment used in thetransmission of the packet.
- X(;) are concerned. Basically in C-BTMA, for a short-pe4) Carrier Sense Multiple Access (CSMA) [33], 1351: In this
riod of time following the
start of a transmission,
all nodes
beto sense the presprotocol, it is required that a nodeable
withinatwo-hopradiusaroundthetransmittergetblocked.
ence of transmissions byits neighbors. (Recall that in nar(Note that if the propagation delay between nodes were
row-band systems this is simply accomplished by sensing
zero then C-BTMA would be a collision-free protocol.)
The
carrier.) A packet will be transmitted by a node onlyif that
remaining variants are similar to C-BTMA exceptthat only
node is not 'already transmitting and no ongoing transthe destinationsof active links transmit the busytone.
They
missions are sensed.Consider first its use in narrow-band
form the class referred to as destination-basedBTMA (Dsystems. In spite of carrier sensing, two factors remain
BTMA). In idealisticdestination-basedBTMA (ID-BTMA),the
whichcontributetocol1isions.Thefirstisthenonzeroprop
agation delaybetween neighbors: given that atransmission destination of every active link emits the busy tone (even
if that node could not receive the packet destined to it;
3 ( ( i , j ) ) has been initiated, all nodes in the set 3t *(i) are not
hencethe name idealistic). It is introduced for performance
blocked from transmitting
until thetransmissionfrom node
comparison purposes. In receiving destinationBTMA (RDi has been sensed by them all; thus the propagation time
BTMA), the destination node emitsthe busy tone only if it
from node ito its neighboring nodes constitutes a vulneris able to synchronize on and start receiving thepacket. In
able period for 3((i, j ) ) as far as transmissions from 3t *(i)
narrow-band systems, this implies that thechannel around
n 3t(j ) are concerned. The secondfactor is hidden nodes:
the receiving node is idle prior to the arrival time of the
given that a transmission 3 ( ( i , j ) ) has been undertaken,
3((i,j)) has
packet. In this protocol, given that transmission
nodes in 3t(j ) - %(i) cannot sense the presence of 3 ( ( i , j ) ) ,
been undertaken, then after a vulnerable period equalto
andarethusneverblocked bythattransmission;inthiscase,
atwo-hop propagationdelay, all nodes in theset ' X * ( j ) are
the entire packet transmission time constitutes a vulnerblocked from tran~rnitting.~
able period for3((i,j)) as far as transmissionsfrom hidden
Since a node does not generally knowif a particular transnodes areconcerned. In spread-spectrumsystems, the caBTMA
pability of sensing the activity of neighboring transmitters mission is destined to itor not, receiving-destination
is also idealistic. It too i s considered for comparison purdepends on the mode of operation of the network[q.In
poses. In practice, this information is obtained from the
systems using space-and-bit homogeneous codes, activity
packet header; assuming that the packet header is prothe output of
sensing is possible, andis done by observing
cessed as soon as it is received and before the entire packet
the matched filters corresponding to the desired waveis received, the time at which a node can first determine
forms. In systems which employ receiver-directed bit-howhether ornot it is the intended immediate destination for
mogeneousortransmitter-directed bit-homogeneouscode
a particular packet reception
is at the endof the processing
assignment, activity sensing requires morehardware: if diof the packet header. In Hybrid destination-based BTMA
rect sequencepseudonoise modulationis used, a nodewill
(HD-BTMA) a nodeoperates as in C-BTMA until theheader
have to possess a bank of filters matched to all possible
is processed, upon which time it operates as in receiving
codes usedby the neighbors;
if frequency hopping is used,
destination BTMA.' For spread-spectrumsystems, the parsensing is done by maintaininga bank of filters forall frequencies used. In systems which employ bit-by-bit code
changing, activity sensing is difficult toachieve, sincethere
'Alternatively, one may additionally makeuse of carrier sensing.
i s no way to know whichcode to try to
detect at anygiven
Thisway,all nodes in X*(i) (which includes 3t*(f
i)3l
t ( j ) ) are
time.Itisnotdefinitelyc1earthattheuseofCSMAinspreadblocked after a vulnerable period of only a one-hop propagation
time, while nodes in the set X( j ) - %(i)
are blocked after a vulspectrum systems improves overall network performance
nerable period equal to a two-hop propagation time. This is presincespread spectrum already exhibitsstrongcapture
cisely the RD-BTMA scheme considered in the simulation study
properties, and inhibiting transmissions may actually dereported upon in Section IV below [VI.
crease network throughput. It may, however, be useful if
'Alternatively, one may conceiveof a scheme in whichthe node
increased probability of correct receptioni s desirable, esoperates as in CSMA until theheader is processed prior to switching to RD-BTMA.
pecially for systems with nonperfect capture, or if lower

140

PROCEEDINGS OF

THE IEEE, VOL. 75, NO. 1, JANUARY 1987

ticular variant of BTMA which makes most sense implementation-wise is receiving-destination BTMAand is referred to in thiscontext as locked-onto Destination BTMA
(LD-BTMAI [8]: the destination node of a link emits a busy
tone whenever the link is active and the destination node
is locked onto the packet (receiver-directed code assignment i s assumed in thiscase).
The performance of any of the access protocols cannot
be easily predicted in a multihop packet radio network, as
many conflicting effects actually take place; furthermore,
the performance of
an accessprotocol is highly dependent
on the topology of the network
and the imposed trafficrequirements. consider, for example, ALOHA and CSMA in
narrow-band systems. If the network is fully connected,
then CSMAwouldsignificantlyoutperformALOHA[35]. But
in a general multihop topology, the presence of hidden
nodes may have such a degrading effect that CSMA may
only be slightly better than
ALOHA. (See numerical results
below.) The situation becomes more complexwith BTMA.
In conservative BTMA,a transmission maybe blocked, while
in destination-based BTMA it may be allowedto take place
and may result in a successful transmission. On theother
hand, in destination-based BTMA (and CSMAfor that matter) it i s possible to allow a transmission to take place that
will be unsuccessful, and whosepresencemay block a
number of other potentially
successful transmissions,while
in conservative BTMA, the former transmission would be
inhibited bytheprotocol,allowingthe later potentiallysuccessful transmissions to then take place.6 It is not clear
whichof theeffectswouldpredominate in agiven situation,
and hence one must resort to analysis and simulation to
compare the different schemes.
IV. NETWORK
CAPACITY

As with point-to-pointpacket-switched networks, we define the network input traffic requirement by the matrix
r = [?,&I,where y i k , j # k, is the average number ofpackets
per unit timeoffered at n o d e j and ultimately destined for
k. (We let yii = 0, vi.) We let y again denote the totaluser
traffic, i.e.,
N

A
and define a,&= yikly: The matrix [alk] represents the traffic
pattern. We also deftne network capacity to be the maximum value y that the network can support, keeping the
matrix {aik] fixed. In this section, we focus on analytical
models appropriate for the determination of network capacity.
As with point-to-point networks, the analysis of packet
radio networks is simplified by assuming that the routing
algorithm is of thefixed type (i.e., all packets from a given
source to a given destination flow through thesame set of
nodes), all nodes haveinfinite buffers, and no link nor flow
control protocols are in effect. While under such assumptions the determination of network capacity for point-topoint networksbecomes a trivial task, for packet radio networksitremainscomplexduetothemanyoperationalcharacteristics that may be in effect, as discussed in the previous
section. Models differentfrom those used for point-to-point
networks are needed. Beforewe proceed with a discussion
6For more details on such a discussion and an example, see [77].

TOBACI: MULTIHOP PACKET RADIO NETWORKS

of themodels introduced andused for multihoppacket radio networks, weshall first devotethe following
subsection
to special cases of single-hop networks and
to a brief review
of relevant models used in theiranalysis; these models constitute a precursor to multihop network
modeling, andwill
thus assist the reader in the understanding of the latter.
A. Single-Hop Networks
Extensive analysisof single-hop networks operating under a variety of channel access methods have appeared in
the literature. Most analyses related to narrow-band systems with zero capture, although some did address the case
of spread-spectrum and time-capture. In [I],Abramson introduced and analyzed pure ALOHA. Slotted ALOHA was
then introduced byRoberts [51], [52] and later analyzed by
Abramson [2], [3] and Kleinrock and
Lam [32], [34].Kleinrock
and Tobagi introduced and analyzed CSMA [33], [35] and
BTMA[67]. Raychaudhuriand analyzed slotted ALOHA with
codedivision [50]; Cronemeyer andDavis
considered
spread-spectrum slotted ALOHA with capture due to time
of arrival [14]. Musser and Daigle derived the throughput
ofpureALOHAwithcodedivision[46].Finally,Pursleystudied the throughput frequency-hopped
of
spread-spectrum
communications [MI. While these represent the key references, additional publications on the subject have appeared in the literature. Basically,two models haveemerged
from this work: an infinite-population model and a finitepopulation model. We
discuss thesehere briefly inspecific
representative cases.
In theanalysis of single-hop networks, the environment
i s assumed to consist of a population of nodes communicating with a single station. For the ALOHA schemes, radio connectivity need not exist among the nodes themselves. For CSMA,it is first assumed that allnodes areradioconnected, and thus every terminal can sense the activity
due toany other terminal.This assumption is then relaxed
allowing hidden nodes to be present. Note that, if the system under consideration
is fullyconnected and of thezerocapture type,then all packets need not bedestined to a single node for theresults to apply.
I) An Infinite-Population Model for the Determinationo f
ChannelCapacity Under Various Channel AccessProtocols: In the infinite-population
model, the numberofusers
is assumed to be very large (infinite).Users collectivelygenerate packets at some aggregate rate,
sayy packets per second. Assuming fixed-size packetswith transmission timeT
seconds, and normalizing timeto T, the packet generation
rate becomesS = yTpackets perpacket transmissiontime.
The assumption of a large population allows us to consider
thateach nodewill beabletosuccessfullytransmit its packet
in a period of time small compared to the timeit takes it
to generate a new one. Thus no queueing of packets ever
takes place, andall outstanding packets (i.e., generatedbut
not yet successfully transmitted) belongto different nodes.
From the point of view of multiuser interference, this is
clearly the worst case. Due to packet collisions and blocking by the access protocol, the rate at which packets get
submitted to ortransmitted on the channel i s larger than
S, and is here denoted byC.In this model, it is furthermore
assumed that the mean rescheduling delay X incurred by
a packet that was blocked orunsuccessfully transmitted is
very large compared to thepacket transmission time (theoretically infinite).(Recall that this modelis aimed towards

141

derivingthe capacity ofthenetworkunder


anaccess
method, and not packet delay.) Accordingly, little correlation exists between newpacket arrivals and their rescheduling. Assuming the process of new packet generation to be Poisson with rate S, the channel traffic process
can also beconsidered to be Poisson with rate G. (We note
that this has been confirmed for slotted ALOHA by conducting a more exact analysis which takes into account X,
and then lettingX + 00 [32].) The analysis then reduces to
considering a channel traffic which is Poisson, rate G, and
deriving the correspondingrate of successful packetsS as
S = GP,, where P, is the probability thatan arbitrary scheduled packet is successful. P, varies with theaccess scheme
and capture assumptions used, and may involve various
relevant system parameters. Since
the packetsize has been
assumed fixed, S represents also the throughput. In systems with zerocapturefor pureALOHA we have
S = Ge-2C;
for slotted ALOHA we get S = Ge-; for nonpersistent
CSMA we get S = Ge-aC/[G(l 2a) e-a], where a i s the
ratio of propagation time 7 (assumed pessimistically to be
the same for all pairs ofusers) to packet transmissiontime
T. Similar expressions can be derived for other protocols
and other captureassumptions. The network capacity Cis
simply obtained by maximizingS with respect to G. It was
shown, for example, that C = 0.18,0.36, and 0.85 for pure
ALOHA, slotted ALOHA, and CSMA(a = 0.01), respectively.
2) A Finite-Population Model for the Determination o f
Channel Capacity Under Slotted ALOHA: Abramson also
considered a finite-population model for the analysis of
slotted ALOHA [2], [3], [36]. There are N nodes numbered
1,2, . * . , N. Node i,i = 1,
* , N, follows a Bernouilli process rate G; to determine the slots during which transmit
to
packets. Assuming that at each such slot node i does actually have a packet to transmit, Gi is then therate of channel traffic due to node i.The total channel traffic is

G=

,E G;.

,=1

Let {C,};=,denote that point. Let


N

c = r,x
c;.
=l
We note that C = maxcc,);-,(S) such that Si = ais, vi, and
is achieved by
a
unique set of
transmission
rates
{G:}Y=l. Note that the set of equations (2) is valid only so
long as each node hasalways packetsto betransmitted, and
thus is valid at the boundary. We therefore deduce that the
approachisvalidforthedeterminationof networkcapacity.
3) The Hidden Terminal Problem in CSMA:As pointed
out in Section I l l , CSMA suffers performance degradation
in the presenceof hidden terminals.
This problem has been
studied for single-hop networks by Tobagi and Kleinrock
in [67]. Anenvironment consistingof a large number ofterminals communicatingwith asingle stationis considered.
Packets are of fixed size. All terminals are in line-of-sight
and within range of the station
but not with
respect to each
other. All assumptions introduced above for the infinite
population model are considered to hold true. From the
hearingmatrixdescribingconnectivityfortheenvironment, the population can then be partitioned into say N
groups, suchthat all nodes within a grouphear eachother
and hear the same subset of nodes in the rest of the population. Aswith the infinite populationmodel, it is assumed
that each group consists of a large number of nodes who
collectivelyform an independent Poisson sourcewith mean
ratesi packets per unit time. Heretoo, given atraffic pattern

where

s=

csi

i=l

the problem is to find {Ci}Y=l such that


N

cn
= C
i=l

C; = max

{s)

Let Si denote the probability that node


itransmits a packet
and Ci = ajC.Note that forthe ALOHA schemes, the result
and is successful. Si also representsthe throughput for node is the same whether or notall nodes are radioconnected.
i.The total throughput is
For CSMA, major degradation is incurred, rendering such
N
an analysis essential.Considering Gi to be the rate of chans = i = l s;.
nel traffic offered by group i,a set of relationships of the
form
Under the assumption that nodes always have packets to
S; = C; fi(G1, Gz, . , G N ) ,
i = 1, 2,
, N (3)
be transmitted, the following set of equations is derived:
is derived [67]. This set
of equationsis similar to(2) obtained
N
with the finite-population model
that Abramson devised for
Si = G; II (1 - C,),
i = 1, 2, . * * , N.
(2)
/=1
slotted ALOHA; the difference is simply the complexityof
j#i
(3) as compared to (2) requiring numerical procedures for
the
determination of network capacity. This was done by
This set of equationsis used to determine the boundaryto
writing
(3) in the form
the
region
for
This is done by computingthe Ja-

cobian 1 = )(Sl,Sz, * * , SN; G1, Gz, * * , GN),and setting


it to zero. This leadsto the conditionCYsl Ci = 1, which, in
turn, leads to the determination of asurface in the(S1, S2,
*
,SN) plane representingthe boundary.For agiven traffic
pattern {ai}Y=l= {Si/S}Y=l,the network capacity is given
by the point on the surface which corresponds to its intersection with the line

142

and by solving iteratively the set of equations for {Gi},=l


given {Si}:=, starting with the initialvalues

If the iterative procedure results in a finite traffic vector


{G,},then the inputvector i s feasible; otherwise it is not.
Thus the networkcapacity can bedetermined numerically
by gradually increasing{Si} until the procedure diverges.

PROCEEDINGS OF THEIEEE,

VOL. 75, NO. 1, JANUARY 1987

B. Multihop Networks

observed that it is applicable to


other
two
schemes;
namely,
ALOHA with zero and perfect capture, and C-BTMA wiih
While there is an abundance of papers in the literature
perfect capture. (C-BTMA with perfect capture was also
on single-hop networks,until 1980, veryfewdealtwith multreated independently by
W. C.-N. Chen [lO].)Theyalso obtihop systems; and even then, those which did examined
served that not all schemes could be modeled simply by
specific topologies.The first paper of this type
was perhaps
tracking the set of transmitting nodes, nor could they all
Gitmans [25] in which network capacitywas derived for a
lead to a product form solution.In their subsequentpaper
two-hop centralized network consisting of user terminals
161, they extended the model by considering thestate decommunicating with a single station via repeaters located
scription to consist of the set of active links (i.e., the set of
around the station, and operating under slotted ALOHA.
transmitting nodes along with their respective intended
The model used is acombination of the infinite-population
neighbors).This newmodel could thus
accommodate other
model representing the terminals,and Abramsons finiteaccess schemes(suchas ID-BTMA), andnetworks with nonpopulation model representing the
repeaters. Lateron, Tosymmetric hearingmatrices. They alsoprovided necessary
bagi considered the same topology in [70]-[73]; he derived
and sufficient conditions on the network (topology,
traffic
the network capacity
as well as the throughput-delaycharrequirements, andaccess protocol) for theanalysis to lead
acteristics for both slotted ALOHA and CSMA, and disto a steady-state distribution with a product form. In the
cussed their dependence on such key parameters as resame paper, they also solved for the throughput in netpeaters connectivity, transmission protocols, and repeatworks operating under the zero-capture mode. these
With
ers storage capacity (see Section V). Yemini analyzed the
results, the networkcapacity for narrow-bandsystems with
capacity of one-waytandem slotted
ALOHA networks in [79];
arbitrary topologies under pureALOHA, CSMA, C-BTMA,
the slotted nature of the system and the consideration of
and ID-BTMA could be determined.
unidirectional networks in which one or more
queues feed
The aforementioned models are not appropriatefor
a third resulted in a problem that could be analyzed (see
spread-spectrum systems. Indeed, they accommodatesysalso[70]). Network capacitycanalso
bededuced from
tems in which the nodes decision to transmit, as well as
throughput-delay analyses by considering the asymptotic
thecaptureof a packet at itsintended receiver, depend only
value of throughput as delay goesto infinity. Discussionof
on theset of activelinks just before the transmission
starts.
such analysis is deferred to the next section.
A spread-spectrum system exhibits a different type of beIn this sectionwe areconcerned with the determination
havior. Whenever the transmitterat a given nodeis not acof network capacity for systems with a general topology,
tive, the receiver will alternate between being locked
onto
and operating under
any possible combination of channel
a packet and being idle (searching for newpackets); and,
access protocol and capture
mode. We describe thevarious
a packet, one
in order to determine thesuccess or failure of
models that have been devised and discuss the major remust not only knowthe
set of active links, but also the state
sults obtained. To that end, we begin with the description
of the receivers (locked or free).Brazio and Tobagiadof a network abstraction adequate for modeling
multihop
dressed this problem by augmenting the
state description
networks and present some basic definitions. Nextwe adso as to include thestate of receivers.In [7], they restricted
dress slotted ALOHA and show that the finite-population
themselves to access protocols for which the decision
made
model which Abramson introduced for slotted ALOHA in
on whether ornot totransmit can be considereda function
single-hop networks can be easily extended to study netof only the set of active links, independent of the state of
works with general topologies.
the receivers (this is precisely the case for pure ALOHA,
In order to study networks operating under other (unCSMA, and C-BTMA). For theseprotocols, the modelleads
slotted) access protocolsandvariouscapture
modes,
to an analysis which takes advantage of results obtained
models based on continuous-time processes are needed.
previously. For the remainingaccess protocols (disciplined
The first work in that direction was by Boorstyn and KerALOHAand locked-onto destination
BTMA) the modeldoes
shenbaum [4]. They formulated a Markovian model for the
not enjoy any special structure, and the solution is thus
analysis of a network with general but symmetric hearing
more tedious to reach [8]. In all cases, errors due to noise
matrix, andzero-propagationdelay, operating underCSMA,
and interfering users canbe properly incorporatedso that
with the assumption that the first packetto arrive to a remostspread-spectrum
code assignments andcapture
ceiver after the channelhas been idle is captured perfectly
modescanbe
effectively addressed. Notethatforthe
(regardless of later events), and all other packets are lost.
models referredto inthis paragraph, the analysis tends to
The zero propagation delay assumption suggests that the
be exactin nature. Unfortunately, thesize of networks that
ratio a is very small, andthat interference is mostly caused
can be effectively handled is rather small.
by hidden nodes. Under these conditions, it was observed
In [Ill, [12], M.-S. Chen and Boorstyn considered the
disis the set
that a sufficient state description for the network
ciplined ALOHA receiverdirectedCDMAprotocol, and
of transmitting nodes (irrespective of their intended neighprovided an approximate analysis which can accommodate
bors.) Analysis then led to a steady-state distribution with
networks of general topology withseveral hundred nodes.
a product form, and relatively simple throughput
equaA noise model is introducedand analyzed wherebya
tions, allowing the design of efficient algorithms for the
threshold i s set on the number of transmitting neighbors
analysis of large networks, of the order of 100 nodes [30].
that would cause packet loss. Based on this approximate
Tobagi and Brazio considered the same model in [75], and
analysis, the performance of several protocols was cornNotethat it mayprovedifficultto haveasignalingschemewhich
allows both CSMA and the
special capture effect assumed (see Section 111).

TOBACI: MULTIHOP PACKET RDlO NETWORKS

Note that for C-BTMA, given the zero-propagation delay assumption, the assumption of perfect captureis a natural one,
even
in narrow-band systems.

143

paredand the effects ofnodeconnectivitystudied in [15].


node and
In [55], Sen considered a more detailed model aof
link and derived better approximations.
A completely
different
approach to modeling packet
radio networks has been taken by Yemini
[80].He used techniques of statistical mechanics which,for some topologies,
led to similar results as those obtained with the above
models. (For more detail, consult also [39].)
7) Network Abstraction and GeneralModel: Given a
static routing function, the throughput requirement for
each link (i, j ) , denoted by Slj, can be computed in terms
of theend-to-end traffic requirementsr. For agiven sourcedestination traffic pattern, there is a corresponding link
traffic pattern. It may happen that for some links the requiredthroughputiszero.Werefertothese1inksasunused
links, and all other links as used links.
Since the entire packet radio network operates using a
single radio channel, each node in the network has one
transmitter, but can in general have more than one outgoing link. We consider that each outgoing linkat a node
has a separate queue for thepackets to be transmittedon
it, and that the transmitter is shared among all queues at
that node. To avoid repeated interference between transmissions in the network,transmission requests for thevarious queues ata node are scheduledaccording to randompoint processes, one for each queue. Consider a point in
time defined by the point process for some link i.If the
queue is empty, this scheduling point is ignored. If the
queue is nonempty then a packet in thequeue is considered fortransmission.Thetransmission mayor may not take
place depending on the status of the transmitter at the
source node (busy or idle), the priority structure (if any)
among the queues at the source node, the channel access
protocol in use,<andthe currentactivity on thenetwork. I f
the transmission is inhibited, or if thetransmission is undertaken but unsuccessfully (due to a collision at the intended destination or to a preemption by another transmission at the source), then thepacket in question (or any
other packet in thequeue, for that matter) is reconsidered
at the next point intime. Otherwise(i.e., the transmission
is successful), the packet is removed from thequeue, and
the same process is repeated at the next scheduling point
for that link.
It is assumed here that neither preemption nor priority
functions are supported at the nodes. We also assume infinite buffer space for each link, zero propagationdelay between neighboring nodes, and instantaneous and perfect
acknowledgments, providing immediate feedback regarding thesuccess or failureof each transmission. As the objective here is network capacity, it is also assumed that at
each scheduling point of the point
process there is a packet
in the queue for consideration.
2) SlottedALOHA: In modeling slottedALOHA systems,
we consideras in Abramsons finite-population model that
the scheduling process according to whichtransmissions
over link (i,j ) take place i s Bernouilli with parameter G,
independent ofall other such processesin thenetwork. Let
Gj = CiGi. Since there is only onetransmitter per node, we
require G j I1. Due to thememoryless and independence
propertyof the Bernouilli
process,and havingassumedthat
no queuei s ever empty,the systems behavior canbe studied byexaminingonearbitraryslot.Inanyslot,wecanwrite
throughput equations similar to (2) taking intoaccount the
specific network topology. For systemswith zero capture,

PROCEEDINGS
144

we have

s- = Gl k = 1 , 2n;..,

(1 - Gk).

(5)

ht, = 1

An iterative numerical procedurecan be used to determine


network capacity for a given traffic pattern. Slotted ALOHA
in spread-spectrum systemswith code-division or time-capture features can be analyzed the same way. Equation (5)
is, in fact, the basis for Gitmans analysis of two-hop centralized networks [25]. This approach hasalso been successfully applied by Silvester andKleinrock to study the capacity of slotted ALOHA networks with regular structure
[61]. One-dimensional networks such as loops and two-dimensional networks such as the square lattice, for which
(5) takes on a simpler form, were considered.
This model for slotted ALOHA, and those discussed below for unslotted
systems, do notattempt to accurately r e p
resent the multihopoperation of a network, since they do
not follow theevolution of packets as they are forwarded
from source to destination. Instead, they only attempt to
represent the interactions that exist among nodes and their
effect on throughput. For the determination of networkcapacity, this representation is believed to be acceptable.
3) Transmitter Activity Model for Unslotted Networks:
a) CSMA with perfect capture o f first packet following
idle channel in networks with symmetric hearing matrix
[4]: In order to lead to a tractable Markovian model, Boorstyn and Kershenbaumconsider the scheduling point process for each used link ( i , j ) to be Poisson with mean Aii,
independent of all other
such processesin the network. In
addition, the message lengths are considered to be exponentially distributed, with mean transmission time l/pj
for messages transmitted by node i.Furthermore, it is assumedthateachtimeanodetransmits(orattemptstotransmit) a message, the message length is redrawn from thecorits past
respondingdistribution,independentlyfrom
history. This implies thateach node has a large mix ofmessages waiting to be transmitted, and that
at eachscheduling
point a different message is considered. (This is a form of
Kleinrocks independence assumption applied to the
traffic at each node or link in the network.) The fact that
different nodes havedifferent mean message transmission
times suggests that nodes may be transmittingat different
data rates, but otherwise the length inbits of all messages
should bethought ofas drawn from the
same distribution.
Consider a network with a zero propagationdelay. Given
CSMAand the capture assumption made, a scheduling
j ) results in a successful transmission if
point for link (i,
and onlyif all nodes in theset X ( { i , j } )are idle (nottransmitting) at the scheduling point. The averagetime ofsuch
a successful transmission is simply l/pi. Since the scheduling process is Poisson, each scheduling point is random
look in time. Accordingly, the throughput S, is expressed
as
Sij =

x,

- Pr {X({i,j } )
Pi

idle}.

To compute Pr {X({i, j } ) idle} forall used links, given that


the hearing matrix is assumed to be symmetric, it is sufficient to consider the process {X@),t 2 0}, where X ( t ) i s the
set of nodes busy transmittinga time t. Due to blocking in
CSMA, not all subsets of nodes can be transmitting simultaneously; the state-space S can be determined fromH.

OF THE IEEE, VOL. 75, NO. 1, JANUARY 1987

Let X i C,X,. Due to the


assumptions of exponential message lengths and Poisson scheduling point process, X ( t ) is
a continuous time Markovchain. Letting Q(D) denote the
steady-state probability of state D, the following balance
equations hold:
r

U(D)to be the set of nodes that are not blockedby any node
in the set D, given the access protocol. The balance equations then become:
r

,x

+ reD A , Q O

+ iC
X,Q(D cD

{i}).

(7)

Let Gi = Ailpi. It is easy to check that the expression

- {i})

(11)

for which the solution


given in (8) also holds. (Note that for
CSMA U(D) = { i ; i @ Z(D)}while for C-BTMA U(D) = { i ;
i@ 3Z2(D)}.) The throughput equation for link ( i , j ) in CBTMA is simply
Si, =

hi

- Pr {X2(;)idle}.

(12)

Pi

satisfy the balance equations and therefore constitute the


solution. The constant Q(4) is obtained by normalizing the
distribution. Given (8), (6) can then be written as

Given the result in (8), (12) can be expressed in terms of a


sum of products similar to (9).
Consider now pureALOHA. Here S = 2 { 1 v 2 c - . , Nand
) , all
nodes actindependently. Moreexplicit expressions maybe
written. for every node i,we simply have

DnX({r.j))=6

Pr { i idle} =

Here again, in general, an iterative numerical procedure is


used to derive the network capacity for a given traffic pattern. For each link ( i , j ) with nonzero desired throughput,
(9) requires the computationof the sum of products forall
subsetA of nodes,
states D containing nodes only in agiven
X({;,j } ) . Let SP(A) denote such
namely, the complement of
sum of products corresponding to subset A. The computation ofSP(A) is aided by recognizing the existence of the
following recursion:
SP(A) = SP(A - { k } )

+ GkSP(A - Wk)),
for k

E A.

(IO)

for large networks with a general topology this task re[4],


[301
mains formidable. KershenbaumandBoorstyn
found iterative relations for these sums of products, and
developed a set of algorithms to efficiently generate these
expressions and perform the iterations of (4). Other relationsallowthedecompositionof
larger networks into
smaller, more tractable, segments. The complexity of their
algorithm, althoughexponential in general,growsquadratically or cubicallywith the number of links
most
fornetworks on the orderof 100 nodes. Running time is minutes
on an IBM-PC and ten times faster on a VAX 780.
Maglaris eta/. [44]generalized the above analysisto allow
nonexponential packet length distributions. The method
of stages developed by Cox to approximate any given distributionfunction byanErlang-branchingconfigurationwas
used to model packet length distributions. It was shown
that the above analysis i s still valid and the form of there-

+ Cj'

Thus Q(D) i s given by the following closed form expressions:

In writing the throughput


equations, thetwo cases of perfect capture and zerocapture are distinguished. In either
case, a necessary condition for a scheduling point corresponding to link< i , j ) to result in a successful transmission
is that 32( j ) be idle at that time. Furthermore, the average
length of a successful transmission is no longer 1Ip;. Indeed,

thesuccessorfailureinthereceptionofapacketisnotsolely
dependent on thestate of the networkat the beginning of
reception, as eventsmay occur during its reception affecting its successful completion. for the perfect-capture
case,for example,receiving nodejmayabortthe reception,
giving priorityto i t s transmitter (see SectionI l l ) . In the zerocapture case, in addition to thepossibility of nodej aborting its reception, transmission by any neighbor of j causes
a collision. Let T ( ( i ,j ) ) denote theaverage period of time
that a transmission over ( i , j ) contributes to throughput,
given that X ( j ) i s idle at the beginningof transmission. The
general throughput equation for ALOHA is given by

which results in the following closed-form expressions:


perfectcapture

suItsarepreserved.Inaddition,itwasshownthatthemodel
could accommodate different packet length distributions
for each link emanating from a node. In [5], this modelwas
also extended to include theeffects of acknowledgments.
b) PureALOHA and C-BTMA in networks with symmetric hearingmatrix: Tobagi and Brazio [75] showed that
the state description required for the analysis of ALOHA
and C-BTMA in symmetric topologies is, as above, the set
of nodesbusy transmitting.Independently, C.-N. Chen
treated C-BTMA the same way in [IO]. To write thebalance
equations for these systems,for each stateDESwe define

TOBAGI: MULTIHOP PACKET RADIO NETWORKS

zerocapture

It,

4) Link Activity Model[6]: In order to accommodate networkswith nonsymmetric hearingmatrices,aswelI asother


access schemes such as ID-BTMA, Brazio and Tobagi in-

145

troduced the linkactivity model [6]. The focus here is still


on access protocols in which thedecision to transmit can
only be basedon thestate of thetransmitters, and not the
receivers (namely, pure ALOHA, CSMA, and BTMA).
Recall that a used link is said to be active whenever a
transmission i s taking place over that link. Clearly, the activityofthelinksofthenetworkisconditioned bytheaccess
protocol in use. For the access protocols considered here,
an accessprotocol is a set of rules which, given
the current
set of active links in the network, determines whether or
not a given inactive link can become active. Consider all
A
used links to be numberedg 1, 2, . . . , L, and let 6: = { I ,
2,
* , L}. The point process for link i, i E C, is again considered to be an independent Poisson processwith rate XI
( X I > O).Thetransmissiontimeof themessagestransmitted
over link i i s assumed to be exponentially distributedwith
mean I/pi ( p i > 0), and to be redrawn independently from
thisdistribution each time themessage is transmitted. (Note
that different links emanating from a nodemay have different means.) Let X ( t ) denote the set of all active links at
time t. X ( t ) is a continuous time Markov chain.
Given an access protocol, link i E C is said to block link
j if, whenever link i is active, the protocol used does not
allowaschedulingpointforIinkjtoresuItinanactualtransmission. (Note that, in general, if linki blocks link j, it does
not necessarily follow that link j blocks link i.) Let D be a
C.D blocks link j E C - D if there exists some
set of links in
link ;E Dwhich blocks j . (The definition of blockingcan be
extended to include the case where links i and k do not
block j individually, but dotogether.) DefineU(D)to be the
set of all links inC - D which are not blocked by D.It was
shown in [6] that the Markov chainX ( t ) is irreducible, positive recurrent, and ergodic. Denote by {Q(D), D E S } its
stationary probability distribution.The equilibrium equations are shown to take the form

I C M(D)

Q(D U { ; } ) p i ,

DES
(18)

D such that D U { i } E
S,andI(D)isthesetofallIinksj~Dsuchthatjisnotblocked
by D - {j}.

where M(D) is the set of all links i

In general, the solution to (18) does not have a product


form. (Note the difference between (18) and (II).) In [6] it
was shown that the existence of a product form is equivalent to X ( t ) being reversible[29], and that X ( t ) is reversible
if and only ifD = /(D)
for all D E S, or equivalently, U(D)=
M(D) for all D E S. (Note that under these conditions, (18)
and (11) become similar.) This result, in turn, was shown to
lead to the followingsimple characterization:A necessary
and sufficient condition for a channel access protocol, together with a given network topology and traffic requirements, to have a productform solutionis that, for all pairs
of used links i and j, link j blocks link i whenever link i
blocks link j.
As an application ofthe above characterization, onecan
easily prove that, with asymmetric hearing matrix,CSMA
always leads to a product form solution. If, however, the
'Note that in the link activity model
i refers to a link, as opposed
to a node.

146

hearing matrixis not symmetric, onedoes not get a product


form solution, unless all pairs of nodes n, and n, for which
h,
= 1 and ,,h
,,
= 0 are such that at least one element
of the pair is not the source of any used links. As another
application, it iseasyto prove that, general,
in
ID-BTMAwill
not lead to a product form solution, since symmetry in
blocking does not usually prevail, even when the hearing
matrix is symmetric. For somespecifictopologiesand traffic
patterns, however, ID-BTMA will have a product form solution. Examples of these are a star network with arms of
length 1 and arbitrary traffic pattern, or a four-node chain
in which the outernodes generate no traffic.
When the condition stated above is satisfied, the stationary distribution is simply given by

When it does not hold, onehas to resort to a numericalsolution of (18).


Thethroughputoflinki,Si,isthelong-runfractionoftime
that link i is engaged in successful transmissions.Using results from the theoryof regenerativeprocesses, it was formally shown in [6] that SI is given by

(20)
where Us(;)
is the collection of
states D E S that donot block
link i and do not contain an active link which causes link
i to be unsuccessful from its start, and T(D, i ) , D E U5(i),
is
the average time that a transmission
over link icontributes
to throughput given that thestate just prior to thestart of
the transmission is D. For CSMA and BTMA with perfect
capture, we haveT(D,i ) = l/pI.Equation (20) then takes on
a simple form; applied
to CSMA, it is similar to (9);the same
is true forC-BTMA. Under zero-capture, however,similarly
to what was encountered above in the analysis of pure
ALOHA, the average transmission time ofsuccessful
a
mesexists between
sage is not l/pi,due to the dependency that
the message length andits success. In order to derive T(D,
i ) , an auxiliary Markov chainis created as follows. Let a,(;)
be the collection of states in which i is active and no interfering link is active, let &,(;) be the collection of states
in which i is active and some interfering linkis also active,
and let 9 ( i )be the set of states obtained from a,(;)by deactivating linki. The start of atransmission over link iwhich
does not suffer a collision at its very start corresponds to
a transition ofX ( t ) from astate D E 'u,(i) into state D U { i }
E e&). Link i remains free of collisions as long as X ( t ) remains in a,(/]while i is active. The successful completion
of link i's transmission corresponds to a transition from
some state in a,(;)into a state in 9 ( i ) (without having ever
visited any state in @,(i)).The derivation of T(D, i ) is then
simply obtained by considering the auxiliary Markov chain
consisting of the states in a,(;),and two absorbing states
representing a,(i)and 9(i). This approach is used to derive
the capacity of any narrow-band systems operating under
pure ALOHA,CSMA,C-BTMA,andID-BTMA.
When a p
plied to pure ALOHA, it provides results similar
to (16) and
(17). It is also usedto model errors due
to noise and capture
effects other than zero and perfect
capture, as encountered
in spread-spectrum systems [8], [MI. Numerical examples
illustrating theuse of these analyses are
given belowin the
following subsection.
5) Models forSpread SpectrumSystems: The models ex-

PROCEEDINGS OF THEIEEE,

VOL. 75, NO. 1, JANUARY 1987

amined so far accommodate systems in which theaccess


protocol and the outcome of transmissions are independent ofthe receivers state at each node.
While these models
are adequate for some narrow-band systems described in
this paper, for the analysis of other narrow-band systems
(such as receiving destinationBTMA) and spread-spectrum
systems, models which incorporate the operation of the
receivers must be considered. At any given point in time,
each node may be in one of threebasic states: idle, transmitting, or receiving (i.e., locked onto a packet).
a) A general Markovian model 17, [ti]: Recall that for
anodetocorrectlyreceiveapacket,i)thenodemustbeidle,
ii) it must process successfully the preamble so as to lock
onto thepacket, and iii) it must complete the reception
of
the packet free of error. Given that the node is idle, the
probability thatit locks onto the incoming
packet depends
on theset of links which
are activeat the start of thepreamble, the evolution of the
system during thereception ofthe
preamble (which,in turn, depends on theset of parameters
{X,} and { p i } defined previously), the topology, the propagation characteristics, the power levels of neighboring
transmitters, the receivers threshold, etc. An accurate representation of most of these aspects is possible. The main
difficulty arises from having to track the evolution of the
system during thereception ofthe preamble, which would
lead to non-Markovian models. Modelsof probabilisticcapture have first been considered by
Ephremides [161, [ I q . I n
[ 7 l , Brazio and Tobagi gave an approximate model of the
operation of receivers so as to maintain the tractable Markovian nature of the model. It is assumed that, given the
set of links Dwhichare active just before thestart of transmission of some link j , the receiver at node n, if idle, successfully locks onto j s transmission with probability 6 J D ;
j ) , independently from trial to
trial. In essence, the model
assumes that preambles are of zerolength, and that6 J D ;
j ) equals the average probability of success for thesystem
with preambles of nonzero length, given
D and j , and averaged over all possible evolutions of the system during the
preambles transmission time.As long as the lengthof the
preamble is much smaller than min {l/Xi, l/pi}, one can
consider the use of 6,(D; j ) which is independent of { X i }
and { p i } to be a good approximation of the real behavior
of the receiver. By appropriately selecting {6,(D; i ) } one
can model the preamble capture behavior of both spacehomogeneous andreceiver-directedpreamblecode
assignments as well as the effect of background noise. If the
preamble is not successfully received, the receiver will remain idle, waiting for a new packet to lock onto, and the
whole process is repeated. If the preamble is successfully
received, the receiver remains locked onto thatpacket until
the end of i t s transmission (unless, in the meanwhile, the
node switches to transmit mode, as for example in pure
ALOHA). The successful
reception ofthe entirepacket then
depends on the activity ofthe transmitters in the network
during thepacket reception time and the capture modein
effect.
With this receiver model, the state of a node at anypoint
in timet, z,(t), can be definedto be (o if the nodeis idle, +k
if link k (with source node n) is active, and - P if node n is
locked onto a packet transmitted over link 0. The state of
the system is then Z(t) = (z,(t), * * ,zN(t)).Furthermore, Z(t)
i s a Markov process. Writing thebalance equations for Z(t)
and solving for the steady-state distribution numerically,
the throughputover eachlink may be derivedin away sim-

TOBACI:MULTIHOP

PACKET RADIO NETWORKS

ilar to (20). This is the approach used in [8] to analyze disciplined ALOHA and D-BTMA.
For systems in which the
access protocol is such that the
decisiontotransmit isfunction onlyof
linkactivity,Brazio
and Tobagi[ilhave shown that a simplification in theanalysis manifests itself in the form of some degree of separation in therepresentation of transmitters activity
and receiversactivity,whichallowstheresuItsobtainedintheIinkactivity model to be useful. Consider a generic noden. Let
rJt) representthe linkwhich the
receiver at noden is locked
onto at time t; (r,(t) = (o if the receiver is not locked onto
any packet). Let,as before, X ( t ) denote theset of active links
at time t. Define
Y,(t)

4 (XW;r,,(t)).

With thereceivers model considered, giventhe Markovian


nature of X(t), YJt) is also Markovian, for each n. The state
space S,* of YJt) is obtained by finding the
possible values
of rJt) for each X ( t ) = D. In [q,it is shown that the Markov
chain Y,,(t) possessesa unique steady-state distribution. Let
{QJD; j ) : (D; j ) E SX}

denote this distribution. Let Q(D)denote, as before, the


steady-state probability for state D. Define

P,,(D; i ) being the probabilitythat node


n is locked onto link
i given that the network is in thestate D. The throughput
equation for link i is then shown to take the form

of
that do not block
i, and
where u( i )is the collection states
where given D, p , ( D ; p) is the probability that the destination of linki, di, is not busy receiving,6 , ( D ;i ) is the probabilitythatthe preamble is processed correctly(i.e., d,locks
onto i), and Ti@) i s the average time that a transmission
over link i contributes to throughput. Pd,(D; p) is determined froma set of linear equations derivedfrom thebalance equations ofprocess YJt). T,(D)depends on theconditions under which the
packet will be successfullyreceived
after the preamble is successfully locked onto, and is derived in a way similar to thesame in (20).
Numerical Results: To illustrate the use of all the models
described so far, and compare the performance of some
access schemes, we consider here a ring of N nodes (N 1
3) [q,[8],[751,in whichall nodes behave identically, and all
links carry equal traffic. Figs. 2-5 show the linkcapacity as
a function of N. In Fig. 2, narrow-band systems with zerocapture are considered. (In this case, disciplined ALOHA
assumes that a receiving node, which actually cansynchronize on and receive a packet only if the
packet reaches
the receiver when the channel is idle, knows the length of
the packet being received.) Theeffect of noise is also taken
into account by consideringa constant bit errorrate Pb, for
packets that are not interfered with. (In Fig. 2, Pbe is 1.9 x
It is clear that BTMA offers a performance thatis superior t o all other schemes. Note thatCSMA performs well
when N = 3 and thus the networkis fully connected; for
N > 3, however, the performance of CSMA is heavily dePure ALOHA, CSMA, and C-BTMA are such protocols.

147

In [8], [75], topologies other than


the ring, with higher degrees of connectivity, have also been considered, leading
to a ranking ofaccess protocols similar to that ofthe above
figures.
b) An approximate analysis of disciplined ALOHA receiver-directed CDMA: M.-S. Chen and R. Boorstyn [Ill,
[I21 considered the disciplined ALOHA receiver-directed
spread-spectrum case with the same model assumptions
regarding link traffic and message length distribution as
before. For any state D and link ( i , j ) , it is assumed that,

aslongasjisidle,6j(D,(i,j))=1.Furthermore,ifareceiver
locks onto a packet, then all future transmissions by neigh-

.ooo'

6
7
8
Number of nodes

Fig. 2. Link capacityof narrow-band systemswith a ring topology operating under various channel
access modes.

-1 . 2 5 0 -

.-

A V packet length = t W 0 bils

Narrowband zero Capluce

u
Y

Topcuwes

P&

= 19'10 Io

Bottom curves P k = 6 1.10

2 ,200

,000
2

Number of nodes
Fig. 4. Linkcapacityin spread-spectrum systemswitha ring
topology and bit-homogeneous code assignment.

\,!

I..-.+---..a____.-A

A v packet length = t o o 0 bds

Top Curves Spread spectrum brt homogeneous. m l o l m codes

7
8
Number of nodes

Bottom Curves Narrowband zero capture


P k

1.9'10~0

Fig. 3. Effectof P
o,n
linkcapacity in narrow-band systems
with a ring toplogy.

graded owing to the


presence of hiddennodes. Fig.3 shows
the effect of Pbe by comparing the results for Pk = 1.9 X
IO-'' and Pbe = 6.1 x
In Fig. 4, spread-spectrum systems with bit-homogeneous codes operating
under
ALOHA, disciplined ALOHA, and CSMA are considered.In
these systems, CSMA does
offer an advantage over ALOHA.
The effect of code assignment on capacity is shown. Forall
schemesconsidered, transmitter-directedcode assignment performs best, uniform code assignment performs
worst. (Recall however, that transmitter-directed code assignment requiresa certain degree of coordinationamong
the nodes rendering it more complext o implement.) In Fig.
5, narrow-band systems and spread-spectrum systems are
compared. Spread spectrum achieves higher link capacity
than does narrow band. (Note that for spread spectrum,
bandwidth expansion is not factored into the results.)

148

$?-

ALOHA

.ooo
2

Number of nodes
Fig. 5. Comparison of narrow-band and spread-spectrum

systems with a ring toplogy.

PROCEEDINGS OF THE IEEE, VOL. 75, NO. 1, JANUARY 1987

bors cause no interference (perfect-capture). Underthese


conditions
S, =

;,

- Pr {i

and j idle).

Pij

While the marginal probabilities

a
pi = Pr {i is idle}

the same transmission range, maximum capacity is reached


when theaverage nodal degree is 15 to 25 (for networkswith
60 nodes). Network capacity is proportional to the square
root of the number of nodes in the network. At levels of
connectivity where a maximum capacity is achieved, that
capacity increases linearly with noise threshold until the
threshold reaches approximately half the
average nodal degree (see Fig. 7). The capacity of CSMA and C-BTMA lies

are found to be expressed as

t
I

10.0

the joint probability Pr {i and j idle} could notbe derived


given thelimited state description embodied in the model.
Indeed, disciplined ALOHA does not lend itself to a reversible Markovprocess and thusdoes not possess a product form solution.Pr { i and j idle} is then approximatedby
pipjl(l - Si,- SjJ, based on the assumption that when i
and jare not successfully communicating witheach other,
then i and j are idle independently. The resulting set of
equations is solved iteratively in the same way as before.
is taken
The effect of overlapping orthogonal transmissions
into account in the following
manner. They consider asimple noise model whereby an intended receiver is not able
to receive a packet if the number of neighboring transmissions at the timeof arrival of the packet exceeds a certain threshold n. Thus
< i , j > )is either Oor 1depending
on the number of ongoing transmissions.They make the
approximationthat,conditioned
on thetransmitter-receiver pair being idle, any other neighbor of the receiver
is transmitting independently. The accuracy of these a p
proximations is shown to be satisfactory.
The performance of a number of different protocols
was
compared in [15]. Theyconsidered CSMA with perfect c a p
ture (as in SectionIV-B3),C-BTMA, and the disciplined
ALOHA receiver-directed CDMAwith perfect capture discussed above (see Fig.
6). Networks with arbitrary topology,

+ + +

+ CDMA/2
A C-BTMA
A CSMA-Perlect Capture
CDMM

0
0 CDMAIaJ
x Approxlmatlon

CDMW
CDMNJ

6 CDMNP

A CDMAll

40

A
00
0

A
I

20
40
Average Nodal Degree

60

Fig. 7. Network capacity versus connectivity for a 60-node


ALOHA receiverdirandomnetworkusingdisciplined
rected CDMA with a noise thresholdn (CDMNn).

between the capacities of CDMA with thresholds of one


andtwo.Thelastresu1tdidnottakeintoaccountthegreater
bandwidth required for spread-spectrum signaling
used in
CDMA. More recent results with these models show that
when the transmission range of each node is adjusted to
allow it to reach the same number of neighbors, the o p
timum connectivityis reduced to approximately eleven for
random networks with 60 nodes.
In [55] a different approach to find an approximate solution is used. Detailed models for a node and a link are
is summarized by average
used, but therest of the network
properties. Each node andlink is modeled andvarious parameters areobtained by iteration. They found that thisresults in a much better approximation than the previous
work.
V. THROUGHPUT-DELAYANALYSIS

0.0
0

20

40

60

Average Nodal Degree

Fig. 6. Network capacityversus connectivity for a60-node


random network using various protocols.

uniform end-toend requirements, varying levels of connectivity, andup to60 nodes were considered.
The following results were obtained for CDMA. When
all nodes have

TOBACI: MULTIHOP PACKET RADIO NETWORKS

In capacity
analysis,
the heavy traffic assumption,
whereby at each scheduling point a nodeis guaranteed to
need to
have a packet readyfor transmission, removed the
include in thestate description any information regarding
thenumberofpacketsqueuedateachlink.Also,atcapacity
packet delay is infinite. For loads below capacity, queue
lengths are not infinite and must be included in the state
description. As a result, the state space grows dramatically
in size rendering analysis either complex, or limitedto smallsize networks. Another major problem
in thedelay analysis
of multihop packetradionetworks(and
any multihop
packet-switched network for that matter), is the need to
keep track ofpackets as they are relayed through the network. This too requires additionalstate information which

149

senting very poor performance (lowthroughput and high


delay), in whichcase the average performance is also poor.
For high enoughu, then, thesystem is saturated exhibiting
again a singleequilibrium point but witha very poor performance. (In thiscase,with avery high probabilityall
nodes
are backlogged.) Thesame arguments hold ifM and u are
fixed, and p is varied. One key observation made in these
system is monostudies is the fact that,whenthe
stable, the steady-state average systemperformance is accurately represented bythe performanceof thesystem were
it to be held fixed inits equilibrium state. This also means
that by applying simple arguments based on a fluid apwork
proximation, one can lead to a determination of the conditions forstability as well as the average performance when
the system is stable. This approach has been extensively
used by Tasakaet a/. [23],[66]in studying many systemsof
varying
complexity;it is referred to there as the equilibriumA. Single-Hop Systems
point analysis (EPA) technique. Other delay and stability
In most of these studies it was assumedthat all nodes are
analyses of ALOHA and CSMA have also appeared; seefor
identical and possess a single packet buffer. A nodeis in
example [13], [19], [45].
one of two states: idle (when its buffer is empty), or backSingle-hop packet radio networks with larger numbers
logged (whenits buffer is nonempty). A backlogged node
of buffers per node havealso been
addressed buttoa much
does not generate any new packets; instead it follows the
lesser extent. The first paper to address this issue was by
channel-accessprotocol under considerationin attempting
Tobagi and Kleinrock [Mi] i n which two nodes with infinite
to transmit its packet. Once successful (andthus becoming
buffers each communicate with a station over a slotted
thinkidle), the node generates a new packet after a random
ALOHA channel. It is indicated that, even for a system as
ing time. In essence, this model represents the operation
simple as this, the throughput-delayanalysis is not an easy
of interactive userswhich undergoonetransaction atatime. task and simulation was then employed. The mathematFurthermore, it is a good approximation for asystem with
icaltreatmentoftwointerferingqueueswasfurtherstudied
a large population of noninteractive users, since then the
by Kleinrockand Yemini [38],and later on by Sidiand SegalI
likelihood of having to queue more than one packet at a
[56]-(581for some specific cases for which a solution
exists.
node is low. Assuming the schedulingprocess and distriBut the true solution to this problem (two
nodes with inbution of thinking timeat each node to be memoryless, a
finite buffers) was not reached until recently by Nain [47l
complete description of thestate of the network consists
following the workby Fayolle and lasnogorodski[18].The
simply of the number of backloggednodes. Furthermore,
difficulty encountered i n the analysis of the two-nodecase
the process representing the system state is Markovian.
has not permitted exact solutions to the problem of netStraightforward analysis of the Markov
process leadsto the
works with nodes possessing infinite buffers to progress
desired throughput-delay tradeoffsin terms of thesystem
any further.
parameters.This is basically theapproachfollowedby
Saadawi and Ephremides [!SI have derived an approxiKleinrock andLam in the analysis of slotted ALOHA [34]and
mate solution for thedelay in a single-hop, slotted ALOHA
later by Tobagi and Kleinrockin theanalysis of CSMA [69].
system, with a finite number M of identical nodes. Each
Using the same model, Tobagi derived the distribution of
node has an infinite buffer in whichtransmission
a
unit capacket delay and interdeparture times for both slotted
pable of holding one packet representingthe head of the
ALOHA and CSMA i n P4].
queue is identified. Packets arrive accordingto a Bernouilli
In addition to throughput-delay tradeoffs, the above
process with probabilityu in a slot. A packet is immediately
analyses also led to a definition and understanding of the
transmitted upon entering the transmission unit. If sucstability of systems employing multiple access [MI.(This
cessful, it leaves the node. If unsuccessful, due to a colliissue has also been studied by Carleial and Hellman
in [9].)
sion, it stays in the transmission unit, andis retransmitted
It was shown in particular that, depending on the system
in the nextslot with probabilityp.An approximate Markov
parameters (number of nodes, rescheduling delay, thinkchain is developed for a node in terms of average probaing time), the system exhibits either a monostable or abibilitiesof successful and unsuccessful transmissionson the
stable behavior; in the former case, there is a single equichannel. Expressionsforthe steady-state behavior of a node
librium point(i.e., a networkstate for which themean rate
in terms of these average
probabilities arethen found.
These
of new packet generationis equal to the mean throughput
average probabilities are, in turn, found from the solution
rate); in the latter case, there are two such equilibrium
of asecond Markovchain for the
channel, anddepend upon
points. Consider, for example, a slotted
ALOHA systemwith
the node parametersp,
u, and M and the nodes steady-state
M users, a Bernouilli schedulingprocess rate p, and a geprobabilities previously found. An expression is also deometrically distributed thinking time withmean l /For
~
a . rived for theaverage delay. The resultant set of nonlinear
given M and p, as long as u does not exceed a certainvalue
urnax,the system will have a singleequilibrium point; at urnax
Clearly, if one limits the number ofbuffers per node to some
the system performsverywell providing large
a throughput
maximum, then a numerical solution to the problem is undoubtand a lowdelay. Beyondurnax,
the system enters its bistable
edlyfeasible; this is theexactana1ysisapproach referred toabove
and discussed later for some multihop topologies.
behavior with one of the two equilibrium points repreneeds to be included in the state description. In the literature we
find twoapproaches taken: oneconsists of exact
analysisof specificsystems,generallysmall in sizeor simple
in their topological structures
and their operation; the other
consists of approximate analysis which requires a reduced
is aimed atmore general
state-space description and which
and larger topologies.
As with capacity analysis,a large number ofstudies have
appeared in the literature which
addressed single-hop networks and their throughputdelay analysis. These studies
have greatly helped in the understanding of the behavior
of multi-access channels and in modeling multihop networks. We begin this section by a brief account of the
done on single-hop systems, and then proceed withsura
vey of work done to-dateon multihop networks.

150

PROCEEDINGS

OF THEIEEE,

VOL. 75, NO. 1, JANUARY 1987

coupled simultaneous equations are solved numerically


and the delay computed. In that paper, issues regarding
stability were also investigated.

B. Multihop Networks
Studies of throughput-delay in multihop packet radio
networks have appearedfor slottedALOHA as well as some
unslotted schemes. All assume zerocapture.
Slotted
ALOHA is somewhat easier to analyze since questions regarding generation of new packets, transmissions, recep
tions, andinterference are all resolved in each slot. The state
of all nodes in a givenslot determines the activityand outcome in the next.For unslotted systems, there are no such
convenient times as slot boundaries, and various timings
(e.g., packet length, propagation times, etc.) must be accounted for. We begin by examining thosestudies which
provide exact analyses of specific instances, and then discuss approximate methods aimed at a more general a p
plicability.
7) Exact Analysis: The earliest work on the analysis of
throughput-delay performance in multihop packet radio
networks can be found in [70]-[73] where a specifictopology is considered, namely, a centralized two-hop system.
Terminals communicate with a central stationvia repeaters
with finitebuffers located around the station.
The first hop
isfromgroupsofterminalstotherepeatersontowhichthey
are homed;the second hop is from therepeaters to the central station. Only inbound traffic is considered, and both
slotted ALOHA and CSMA are studied. The analysis assumesadecoupling between
thetwo hops; thisdecoupling
is natural for slotted ALOHA; for CSMA it is based on the
observation that the distribution of time separating two
conflict-free transmissions from a group of terminals can
be approximated by an exponential one, an observation
which is validated by simulation. The effects on performance of the number repeaters,
of
the connectivity among
them,variations in thetransmission protocolstheyuse,and
the number of buffers they possess, have all been examined. Perhaps one of the most interesting results is that,
under the assumptions of zero-processing times
at the
nodes and instantaneous acknowledgments,packet radio
networks are channel bound rather than storage bound: a
slight improvement i s gained by increasing the buffer size
at the repeaters from 1to 2, but nosignificant improvement
is obtained beyond that.Fuduka and Tasaka [22] later analyzedthe same two-hopnetworksusingthe
EPA techniques, addressing similar issues.
Other studiesexist which employed exact analysisof specific instances, but with more
arbitrarytopologies [53], [65].
The models used lead to a large number ofstates and are
thus applicable to small networks. They are used to study
some specific protocol features in small networks.
Takagi and Kleinrock consider two arbitrary networks
with 8 and 12 nodes, operating under slottedALOHA [65].
A distinction is made between terminal nodes which are
sources and sinks of traffic but cannot serve as relays, and
repeaters which serve only as store-and-forward relays.
Otherwise, both types of nodes operateusingslotted
ALOHA under the same conditions. Each node has a finite
number of buffers, initially equal to 1. It is assumed that
there is afixed number of source-sink terminal
pairs which
have nonzero throughput requirements. A terminal
can be

TOBACI: MULTIHOP PACKET RADIO NETWORKS

the source for only oneend-to-end requirement. For each


source-sink pair, a single fixed route
exists. Repeaters may
serve as relays for more than one source-sink requirement.
A terminal node generates and transmits packets in the
same way as described above for single-hop models: a terminal is either idle or backlogged; if idle, it generates and
transmits a new packet in a slot with probability u; if it is
backlogged, it transmits its packet in aslot with probability
p. A repeater has several states:
it is idle ifits buffer is empty;
it is backlogged if its buffer has a packet; in this case the
stateof therepeater corresponds to thesource-destination
pair to which the packet ' ?longs. When backlogged, the
repeatertransmits its packet in aslotwith probabilityp.
The
parameters u and p may be different for differentsourcesink requirements. The resulting model is clearly Markovian, and the throughput-delay performance can be derived in away similarto the analysis of singlehop networks.
Furthermore, it can readily be seen that the behavior of
these networks with respect to throughput-delay tradeoffs
and stability is similar to that of single-hop networks.The
authors then propose differentways of improving performance: i) transmission suppression, in which itis assumed
that some mechanism exists which prevents the neighbors
of a repeater from transmitting to itwhen its buffer i s full;
a
transmits
and ii) transmission acceleration, in which node
a packet in a slot with probability1 if itknows that thedestination and all its neighbors are empty. Furthermore, the
study addressed the multiple buffer case for repeaters,
showing similar results as in [71], [73].
The number of states required to describe the problem
grows rapidly with network size. For the &node network
144statesare
with 3 requirementsandsinglebuffers,
needed. The12-nodenetworkwith 4 requirements and single buffers needs 3456 states. With double and triple buffers, the numbers of states needed for the &node network
are 1176 and 7200, respectively.
Roy andSaadawi[53] developed anexactanalysisof
throughput and delay in a multihop packet radio network
employing a form of CSMA with busy tone and collision
detection. A node transmits a scheduled packet itifis idle
and does not detect carrier ora busy tone generated by a
neighbor. Otherwise,it reschedules the packet. ltwill abort
an ongoing transmission if it detects carrier orbusy tone.
(It is to be noted that implementation such
of a scheme may
prove to be difficult.) A packet will be received successfully
if theneighbors of thereceiver, and all their neighbors,remain idle for the duration of the packet. Packet lengths,
retransmission (scheduling) intervals, transmission detection time, busytone detectiontime, collision detectiontime,
collision termination time,
and nonzero propagationdelay
are all considered in the model and assumed to be exponentially distributed.They also assumethat packets arrive
to the network as Poissonprocesses and are either attempted to be transmitted immediately or are first scheduled for transmission. Finite-size buffers are assumed at
each node. Thus they formulatea Markov process and solve
for the steady-state probabilities numerically. They consider a 3-node chain network which requires 52 states to
describe,andobtain throughput-delaycurvesfor manyvalues of the parameters.
2) Approximate Analysis: The large state space required
in exact analysesand the difficulty in keeping
track ofpackets as they are relayedthrough the networkhave led some

151

to consider approximateanalyses. These analyses assume


that packets areindependently generatedat each node for
transmission to aneighbor, whereby the amountof traffic
generated is a function of the topology, routing, and endtoend traffic requirements (similarlyto the approach taken
in the previous section for the capacity
analysis). They also
consist of developing a model of the neighborhood of a
it with a number ofparameters repnode and characterize
resenting average behavior. Parameters for all nodes are
then found throughan iterative process. Theperformance
of the nodeis then analyzed for the particular protocolin
terms of these parameters.
One of thefirst approximations for analysis of throughput delay in multihopnetworkswas done by
Leiner [42]. His
model considersan arbitrary network with fixed topology
and connectivity, a set of end-to-end requirements, fixed
routing, and a particular protocol. The approach consists
of searching iteratively for the channel
traffic rate (i.e.,
average total number of packettransmissions, successful
or not, per unit time), in the neighborhood of each node,
using the following process.The amount of successful
traffic that a link and a node must support are calculated
as afirst step.Assuming no retransmissions, thetotal traffic
transmitted on the channel in the neighborhood of any
the neighnode is computed. Using single-hop models for
borhood ofeach node,the factor representing the number
of retransmissions required by a particular node can be
found as a function of the totaltraffic in that neighorhood.
The traffic of each link is then increased to reflect the number of retransmissions required by the receiving node of
that link. Node and neighborhood
traffic are recomputed,
and the process is repeated until itconverges. It is assumed
that, if the iteration diverges, the input traffic is not supportable by the network. Leiner uses the one-hop simplified analytic expressions derived in [371 evaluated for the
traffic just found, and then determines the delay experienced by packets at each node; the network delay for the
assumed input requirements (throughput) is obtained by
summing nodaldelays. Queue lengths and other
measures
produced by the one-hop analysis can also befound. The
approach is suitableforarbitrarytopologies,
traffic rea one-hop analysis
quirements, andany protocol for which
exists, fairly large networks can be analyzed. Results obtained using this approach were compared
with simulation
I and a 20-node network withthree levels of
results for aO
inputtrafficandaslottedAL0HAprotocol;thedelayresuIts
of the analysis were within a factor of 2 of the simulation
results.
Lee and Silvester ([41]and [60])developed an approximate
analysis based on a more detailed model of the behavior
of a node. In their paper, they considered slottedALOHA.
They define four parameters that govern the node's operation. In each slot, external packets arrive with probability r; internal packets, from other nodes in thenetwork,
arrive with probabilitya.(The valuesof r f o r each node are
obtained from the input traffic requirements.) The values
of a depend upon thebehavior of neighbors.A node may
be in one of two states-idle when its buffers are empty, and
active when there is a nonempty buffer.In a slot, an active
node will transmit a packet
from thehead of its queue with
probability p. (The value of p for each node serves as an
input control parameter.) A transmitting node cannot accept an internal arrival. A transmission will be successful
if the intended
receiver and all of its neighbors are not trans-

152

mitting. The fourth parameter, @,


is the probability that the
transmission from a node
is successful, and reflects thebehavior of neighboring nodes. The parameters a and @ for
each node are expressed in terms of the same and other
parameters for neighboring nodes. An iterative process is
used to reach consistent values for these parameters. They
then obtainaverage queue lengths and
delays as functions
ofthesefourparameters. Inexpressingaandpasfunctions
of theparameters of neighboringnodes, two complications
arise. First, one must find the proportionof packets transmitted to a particular neighbor.An approximation is to assume that this can be found by considering the ratio derived from the traffic requirements and the routing. (This
is an approximation since some nodes may be more congested, andthus require more retransmissions
to them for
a reception to be successful, leading to different proportions.) The authors also solved for the exact results and
found the approximation to be sufficiently accurate. The
second problem is that a neighboring node will transmit
only if its buffer is not empty. Their modelassumes an independent process with rate equal to the mean transmission rate averaged over all states of the buffer.
In [55] disciplined ALOHA receiver-directed CDMA is
considered. Detailed models afor
node and a link are used,
but therest of the networkis as above represented byaverage properties and treatedparametrically. These local
models are then solved iteratively to obtain consistentvalues for the parameters.
VI.

SIMULATION MODELS

Mathematicalmodelsintroducesimplifyingassurnp
tions which limit their
usefulness in addressing certainaspects of packet radio network performance. Simulationmodels maybeused
to overcome these limitations.
Furthermore, in the absence of real networks operating under the various modes of interest, simulation models are
used to validate mathematicalmodels. Attempts have been
made to write simulation programs thatwould accurately
encompass a / / operational aspects of packet radio networks. Unfortunately, such effort has not been successful
due to the size and complexity of the resulting program.
As
a result, it seems more prudentto create simulation models
which incorporate the essential elements of packet radio
networks, and which can be used to study first-order effects. One such simulation model has been created and
successfully used by Silvester
to study spatial reusein multihop packet radio networks
[59], [39]. Another is by Tobagi
and Shur[771, [78].The latter incorporates a variety of channel access protocols, flow control protocols,andnode
buffer management strategies, and can accommodate arbitrary toplogies with nonzero propagation delays. It can
also be used to study other network aspects, such as the
handling of hop-by-hop acknowledgments and routing,
which have not yet been addressed in any form.
In [77l,[78], the focus has been on the performance of
channel-access protocols.Bothnetwork
capacity and
throughput-delay performance are evaluated, and the effects on performance of the nodal scheduling
rate, the
propagation delay among nodes, and the nodal buffersize
and management strategy are
studied. With regard to nodal
buffer management, structured bufferpool strategies have
been consideredto prevent deadlocksto which networks
with finitebuffers are prone. So far, numerical resultshave

PROCEEDINGS OF THE IEEE, VOL. 75, NO. 1, JANUARY1987

been obtained for networks


in which boththe topologyand
the traffic pattern are symmetric. In such cases, all nodes
in the networkare statistically identical, and thus the complexity of the task is reduced considerably without jeopardizing the objective.
As an illustration of the results obtainedso far, consider
6 nodes. Consider first infinite
again the ring topologywith
buffers at all nodes. Theeffect of nodes' schedulingrate G
on networkcapacity, C(G),
i s seen in Fig. 8 for the ring operating underC-BTMA; (the ratioa of propagationdelay to

with that obtained by analysis. Throughput-delay performance is shown in Fig. IO. The effect of buffer size on the
performance proved to be more complex. The nodes' optimum scheduling rate is found to vary with the nodes'
buffer size, increasing with increasing buffer size. Interesting resultshave been obtained showing
thatwhile some
schemes, such as ALOHA and CSMA, are indeed channelbound, others, such as C-BTMA, are storage bound. This
can be seen in Fig. 11 in which network capacity is plotted
as a function of buffersize. Future steps to be taken in this
simulation study will deal with more general (asymmetric)
topologies, and the optimization of network performance
with respect to system parameters.

ALOHA'

[ 1

CSMA - -

,
.o

IG'
'

'

" " '

'

"

"

'

1100.0
.o
10.0
Scheduling Rate G

'

'

I
'

""".

Fig. 8. Network capacity versus scheduling rate in a narrow-band &node ring network, underC-BTMA
the
protocol,
with a uniform traffic matrix and a = 0.01.

.O

.2 .3 .4 .5 .6 .7 .8 .9 1.0 1.1 1.2 1.3 1.4


Network Throughput

Fig. 10. Network delay versus network throughput for the


&node ring topology with uniform traffic and a = 0.01.

packet transmission time, assumed constant, is 0.01). Define nodal capacity to be the sum of link capacities over
all links emanating from a node. The nodal capacity optimized over G for various schemes is plotted in Fig. 9 as a
function ofa. The ranking among the
schemes is consistent

1 .o,

C-BTMA

.9
.8.

.7
50~

451

RING
N=6 d=2

401

Slotted ALOHA

CSMA

ALOHA

10

100

Translt Buffer Slze per PRU

Fig. 11. Network capacityversusnodes' buffer size for various access schemes.

. .o o w 1

,0010

.01m

.loo0

1.03oO

Fig. 9. Optimum nodalcapacityversusaforthehodering


topology with a uniform traffix matrix.

TOBACI:MULTIHOP

PACKET RADIO NETWORKS

VII. CONCLUSION

Performanceevaluationandoptimizationof packet radio


networks has proven to be a complex task. A great deal of
progress has been made in the past decade.Single-hop net-

153

networks in the presence of noise, i n Proc. lNFOCOM85


(Washington, DC, Apr. 1985).
131 E. J.Coyle and B. Liu, A matrix representation of CSMNCD
networks, /E Trans. Commun., vol. COM-33, no. 1, pp. 536 4 , Jan. 1985.
141 D. H. Davis and S. A. Cronemeyer, Performance of slotted
ALOHA random access with delay capture and randomized
time ofarrival, /E Trans. Commun., vol. COM-28, no. 5, pp.
703-710, May 1980.
151 0. DeSouza, M.-S. Chen, and R. Boorstyn, A comparison of
the performance of protocolsin packet radio networks, i n
Proc. /E Mikom85, Oct. 1985.
[16] A. Ephremides, Thecaptureeffect i n spread spectrum multiaccesscommunications, i n Proc. Princeton Conf., Mar. 1982.
[17] -, Chapter 5 of lectures in Secure digital communications, ClSM Summer School, Udine, 1982.
[I81 G . Fayolle and R. lasnogorodsky, Two coupled processorsThe reduction to a Riemman-Hilbertproblem, Z. Wahrscheinlichkeitstheorie, vol. 47, pp. 325-351, 1979.
[19] M. J. Ferguson, An approximate analysis of delay for fixed
and variable length packets
i n an unslotted ALOHAchannel,
/ E Trans. Commun., vol. COM-25, no. 7, July 1977.
[20] S. C. Fralick and J. K. Leung, Study of throughput and delay
ofspread-spectrum multiple accessmodes, SRI International Packet Radio Note, Menlo Park, CA, Apr. 1974.
[21] S. Fralick and J.C. Garrett, Technological considerations for
packet radio networks, in AflPS Conf. Proc., pp. 233-243,
1975.
[22] A.FukudaandY.Muramatsu,Performanceanalysisofpacket
radio networks with repeaters, Trans. lCE japan (in Japanese), vol. JWB, Dec. 1981.
[23] A. Fukuda and S. Tasaka, The equilibrium point analysisAunifiedanalytictoolforpacketbroadcastnetworks,inProc.
ACKNOWLEDGMENT
/E Clobecom83 (San Diego, CA), pp. 33.4.1-33.4.8, Nov.
1983.
The author would like tothank Prof. R. Boorstyn of the
[24] M. Cerla andL. Kleinrock, On the topological design ofdisPolytechnic Institute of New York for numerous discustributedcomputer networks, I Trans. Commun., vol.
sions which helped tremendouslyin the writing of this
paCOM-25, no. 1, pp. 48-60, Jan. 1977.
per.
(251 I. Gitman, On the capacity of slottedALOHA networks and
some design problems, /E Trans. Commun., vol. COM-23,
no. 3, pp. 305-317, Mar. 1975.
REFERENCES
1261 Proc. /E (Special Issue on Packet Radio Networks), this issue, pp. 3-167.
[ l ] N. Abramson, The ALOHA system-Another alternative for
computer communications, in A W S Conf. Proc., vol. 37,
[27 R. E. Kahn, The organization of computer resources into a
FJCC, pp. 281-285,1970.
packet radio network, i n Nat. Comput. Conf., A W S Conf.
[2] -, Packet switching with satellites, i n Proc. AHPS Conf.,
Proc., vol. 44. Montvale, NJ: AFIPS Press, 1975, pp. 177-186;
vol. 42, June 1973.
also in / E Trans. Commun., vol. COM-25, Jan. 1977.
[3] -,
The throughput of packet broadcast channels, /E
[28] R. E. Kahn, S. A. Cronemeyer, J. Burchfiel, and R. C. KunTrans. Commun., vol. COM-25, no. 1, pp. 117-128, Jan. 1977.
zelman, Advances in packet radio technology, Proc. E ,
[4] R. R. Boorstyn andA. Kershenbaum,Throughput analysisof
voi. 66, no. 11, pp. 1468-1496, 1978.
Proc. ICCIIO(Seattle, WA,
multihoppacket radio networks,in
[29] F. Kelly, Revenibility and Stochastic Networks. New York,
NY: Wiley, 1979.
June 1980).
[30] A. Kershenbaum and R. Boorstyn,Evaluation of throughput
[5] R. R. Boorstyn, A. Kershenbaum, and V. Sahin, A new acknowledgment protocol foranalysisof multihoppacket radio
in multihop packet radio networks with complex topolonetworks, i n Proc.COMPCON82 (Washington, DC, Sept.
gies, in Proc. lNfOCOM84 (San Francisco, CA), pp. 330-335,
Apr. 1984.
1982).
[6] J. M. B r b i o and F. A. Tobagi, Theoretical results on the
[31] L. Kleinrock, Communication Nets: Stochastic Message Flow
throughput analysis of multihop packet radio networks, i n
and Delay. New York, NY: McGraw-Hill, 1964 (out of print).
Proc. ICC84 (Amsterdam, The Netherlands, May 1984).
Reprinted by Dover Publications, 1972.
[A -, Throughput analysis of spread-spectrum multihop
[32] L. Kleinrock and S. Lam, Packet-swtiching i n a slotted satellite channel, i n Nat. Computer Conf. AHPS Conf. Proc.,
packet radio networks, i n Proc. INFOCOM85 (Washington,
DC, Mar. 1985).
voi. 42. Montvale, NJ: AFIPSPress,1973, pp. 703-710.
[8] J. M. Brbio, Throughput analysis of multihop packet radio
[33] L. Kleinrock and F. A. Tobagi, Carrier sense multiple access
networks under a general class of channel access protocols
for packet switched radio channels, i n Proc. /E lnt. Comand capture modes, Ph.D. dissertation, Dept. Electrical Enputer Conf. (Minneapolis, MN), pp. 218-1-218-7, June 1974.
gineering, Stanford Univ., Stanford, CA, June 1986.
[ M I L. Kleinrock, S. S. Lam, Packet switching i n a multiaccess
broadcastchannel:Performance
evaluation, /E Trans.
191 A. B. Carleial and M. E. Hellman, Bistable behavior ofALOHAtype systems, / Trans. Commun., vol. COM-23, pp. 401Commun., vol. COM-23, no. 4, pp. 410-423, Apr. 1975.
[35] L. Kleinrock, F. A. Tobagi, Packet switching i n radio chan410, Apr. 1975.
nels: Part I-Carrier sense multiple-access modes and their
[lo] W.C.-N Chen, Analysis of busy-tone multiple access protocols for multi-hop packet radio networks, Ph.D. disserthroughputdelay characteristics, / E Trans. Commun., vol.
COM-23, no. 12, pp. 1400-1416, Dec. 1975.
tation, Polytechnic Instituteof NewYork, NewYork,
NY, June
1983.
[36] L. Kleinrock, Queueing Systems, Volume 2: Computer Appli[ l l ] M.4. Chen and R. Boorstyn, Analysis of multihop packet
cations. New York, NY: Wiley, 1976.
Performance of distributedmultiaccesscomputer
radio networks withreceiver directed CDMA, presented at
[37 -,
communication systems, in Proc. lFlP Cong., 1977.
the 1984 Allerton Conf., Oct. 1984.
[38] L. Kleinrock andY. Yemini, Interfering queueing processes
[12] -, Throughput analysis of CDMA multihop packet radio

works have been well studied. The models created and results obtained forsuch networks have helped considerably
the later effort of modeling and analysis of multihop network. Unfortunately, the modelsand techniques required
for multihop networks provedto be much more complex.
There are basically two approaches: i) create precise and
accurate models with the results that onlysmall networks
can be handled; properties of large networks have to be
inferred from those of
small ones; ii)provide approximate
analyses based on major simplifying system assumptions
so as to handle general cases, with the risk of producing
inaccurate results. Neither of the two is yet fully satisfactory. Several general observations can be made. Network
capacity appearsto have beenbetter covered than throughput-delay performance.In general, results obtainedin single-hop networks do not always apply to multihop networks (e.g., performance ofCSMA and i t s relation to other
schemes). Verification of the modeling techniques
and validation of the results obtained
have also been limited.Despite the considerable
progress madein the
last decade and
reported upon in thispaper, it is clear that the problemof
analyzing the performance of packet radio networks remains a challenging one, and the desire to reach a methodologyfor thedesign, control, and optimization of packet
radio networks remains unfulfilled.

154

PROCEEDINGS OF THE IEEE, VOL. 75, NO. 1, JANUARY 1987

in packet-switched broadcast communication, presentedat


the IFlP Cong., Tokyo, Japan, 1980.
[39] L. Kleinrockand J.Silvester,SpatiaI reusein multihoppacket
radio networks, this issue, pp. 156-167.
[40] S.S. Lam and L. Kleinrock, Packet switching i n a multiaccess
broadcast channel: Dyanmic controlprocedures, /E Trans.
Commun., vol. COM-23, pp. 891-904, Sept. 1975.
[41] I. Lee and J. A. Silvester, An iterative scheme for performance modeling of slotted ALOHA packet radio
networks,
in Proc. ICC82, pp. 6G.1.1-6G.1.5, June 1982.
[42] B. M. Leiner, A simple model for computation of packetradio network communication performance,
ITrans. Commum, vol. COM-28, no. 12, pp. 2020-2023, Dec. 1980.
[43] B. M. Leiner, D. L. Nielson, and F. A. Tobagi, Issues in packet
radio network design, this issue, pp. 6-20.
[#I B. Maglaris, R. Boorstyn, and A. Kershenbaum, Extensions
to theanalysis of multihop packet radio networks, i n Proc.
INFOCOM83 (San Diego, CA, May 1980).
[45] J. S. Meditch and C.-T. A. Lea, Stabilityof optimization of the
CSMA and CSMNCD channels, / E Trans. Commun., vol.
COM-31, no. 6, pp. 763-774, June 1983.
[46] M. Musser and J. Daigle, Throughput analysis of an asynchronous code division multipleaccess (CDMA) system, in
Proc. K C 8 2 (Philadelphia, PA, June 1982).
[47l P. Nain, Analysis of a two-node ALOHA-network with infinite capacity buffers, i n Proc. Int. Seminar Comp. Networking and Perf Eva/., pp. 2.2.1-2.2.15 (Tokyo, Japan, 1985).
[a] M. Pursley, Throughput of frequency-hopped spread-spectrum communications for packet radio networks, in Proc.
7983 ClSS (Johns Hopkins Univ., Baltimore, MD, Mar. 1983).
[49] -, The roleof spread spectrum i n packet radionetworks,
this issue, pp. 116-134.
[50]D.
Raychauduri, Performance analysis ofrandom access
packet-switchedcodedivision multipleaccess systems,/
Trans. Commun., vol. COM-29, no. 6, pp. 895-901, June 1981.
[51] L. G. Roberts, ALOHA packet system with and withoutslots
and capture,ARPA Network InformationCenter, Standford
Res.Inst.MenloPark,CA,ASSNote8(NIC112390),June1972.
[52] -, ALOHA packets system with and without
slots and capture, Comput. Commun. Rev., vol. 5, pp. 28-42, Apr. 1975.
[53] R. Roy and T. Saadawi, Analysis of multihop packet radio
networks-Carrier sense multiple access with busy tone and
collisiondetection (CSMNBT-CD), i n Proc. INFOCOM84
(San Francisco, CA, Apr. 1984).
[54] T. N. Saadawi and A. Ephremides, Analysis, stability, and optimization of slotted ALOHA with
a finite number of buffered
users, / E Trans. Automat. Contr., vol. AC-26, no. 3, pp. 680689, June 1981.
[55] P. Sen, Local Markov models for multihop packet radio networks: CDMA, in Proc. / K C 8 6 (Toronto, Ont., Canada,
June 1986).
packet-radio
[56] M. Sidi and A. Segall, Two interferingqueues in
networks, /E Trans. Commun., vol. COM-31, no. 1, Jan.
1983.
[57l -, An acknowledged-based scheme in a two-node packet
radio network, /E Trans. Commun., vol. COM-32, no. 6, pp.
741-744, June 1984.
[58] -,
A three-node
packet-radio
network,
i n Proc.
INFOCOM83, pp. 222-228, Apr. 1983.
[591 J. A. Silvester, On the spatial capacity of packet radio networks, Computer Science Dep., Computer Systems Modeling and Analysis Group, UCLA-ENC-8021, UCLA, Los AnA, 1980.
geles, C
1601 1. A. Silvester and I. Lee, Performance modeling of buffered
CSMA-An iterative approach, in Conf. Rec. CLOBECOM82,
pp. F1.6.1-F1.6.5 (Miami, FL, Nov. 1982).
1611 J. Silvester and L. Kleinrock, On the capacity of multihop
/E Trans.
slotted ALOHA networks with regular structure,
Comm~n.,VOI.COM-31, pp. 974-982, Aug. 1983.

TOBACI: MULTIHOP PACKET RADIO NETWORKS

[62] P. Spilling and F. Tobagi, Activity signaling and improved


hop acknowledgements i n packet radio systems, i n SRI-International, Packet Radio Temporary Note No. 283, Menlo
Park, CA, Jan. 1980.
[631 /E Trans. Commun. (Special Issue on Spread-Spectrum
Communications), vol. COM-30, no. 5, May 1982.
[64] J. S. Storey and F. A. Tobagi, Throughput performance of a
direct sequence CDMA packet radio network,
SELTech. Rep.
85-277, Computer Systems Lab., Stanford Univ., Stanford, CA,
June 1985.
[65]H. Takagi and L. Kleinrock, Throughputdelay characteristics of some slotted-ALOHA multihop packetradionetworks, / Trans. Commun., vol. COM-33, no. 11, pp. 12001207, Nov. 1985.
[66] S. Tasaka, Performance Analysis of MultipleAccess Protocols.
Cambridge, MA: MITPress, May 1986.
[67l F. A. Tobagiand L. Kleinrock, Packet switching in radio
broadcast channels: Part Il-The hidden terminal problem i n
carrier sense multiple-access and the busy tone solution,
/Trans.Commun.,vol. COM-23, no. 12, pp. 1417-1433, Dec.
1975.
[68] -,On
theanalysis and simulationof buffered packet radio
systems, Proc. 9th Int. Conf. on System Science (University
of Hawaii, Honolulu, Jan. 1976).
[69] -,
Packet switching i n radio channels: Part IV-Stability
considerationsand dynamic controlin carrier sense multiple
access, /E Trans. Commun., vol. COM-25, pp. 1103-1120,
Oct. 1977.
[70] F. A. Tobagi, Performance analysis of packetradiocommunication systems, in Proc. / E Nat. Telecommunication
Conf. (Los Angeles, CA, Dec. 1977), pp. 12:6-1-12:6-7.
[71] -,
Analysis of slotted ALOHA in a centralized two-hop
packet radio network, in Proc. 7th /E Computer Society
Int. Conf (CompCon,Fall 1978, Washington, DC, Sept. 1978).
[72] -,
Analysis of a two-hop centralized packet radio network-Part I: Slotted ALOHA, / Trans. Commun., vol.
COM-28, no. 2, pp. 1%-207, Feb. 1980.
[73] -, Analysis of a two-hop centralized packet radio network-Part 11: Carrier sense multiple access, / Trans.
Commun., vol. COM-28, no. 2, pp. 208-216, Feb. 1980.
[74] -, Distribution of packet delay and interdeparture time
in slotted ALOHA and CSMA channels, J.Assoc. Comput.
Mach., vol. 29, no. 3, pp. 907-927, Oct. 1982.
[75] F. A. Tobagi and J. M. Brkio, Throughput analysis of multihop packet radio network under various channel
access
schemes, in Proc. INFOCOM83 (San Diego, CA, Apr. 1983).
[76] F. A. Tobagi, R. Binder, and B. Leiner, Packet radio and satellite networks, / Commun. Mag. (Progress i n Computer
Communications), vol. 22, no. 11, pp. 24-40, Nov. 1984.
[ 7 7 l F.A.TobagiandD.H.Shur,Performanceevaluationofchanne1 access schemes in multihop packet radio networks with
regular structure bysimulation,CSLTech. Rep. 85278, Stanford Univ., Stanford, CA, June 1985; to appear in Computer
Networks and /SDN Systems.
[78] -, Throughputdelay performance of multihop packetradio networks with finite buffers, Stanford Electronics Lab.
Tech. Rep., Mar. 1986.
[79] Y.Yemini,Onchannel sharing indiscrete-time, multi-access
broadcast communication, Computer Science Dep., School
of Engineering and Applied
Science, Univ. of California, Los
Angeles, UCLA Tech. Rep. UCLA-ENGS061, Sept. 1980.
[80] -, A statistical mechanics of distributed resource mechanism sharing, i n Proc. IFNOCOM83 (San Diego, CA, Apr.
1983).

Fouad A. Tobagi (Fellow, IEEE), for a photograph and biography,


please see page 5 of this issue.

155

You might also like