1488 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 8, AUGUST 1999
Distributed Quality-of-Service
Routing in Ad Hoc Networks
Shigang Chen and Klara Nahrstedt, Member, IEEE
Abstract—In an ad hoc network, all communication is done
over wireless media, typically by radio through the air, without
the help of wired base stations. Since direct communication is
allowed only between adjacent nodes, distant nodes communicate
over multiple hops. The quality-of-service (QoS) routing in an
ad hoc network is difficult because the network topology may
change constantly, and the available state information for routing
is inherently imprecise. In this paper, we propose a distributed
QoS routing scheme that selects a network path with sufficient
resources to satisfy a certain delay (or bandwidth) requirement
in a dynamic multihop mobile environment. The proposed al-
gorithms work with imprecise state information. Multiple paths
are searched in parallel to find the most qualified one. Fault-
tolerance techniques are brought in for the maintenance of the
routing paths when the nodes move, join, or leave the network.
Our algorithms consider not only the QoS requirement, but also
the cost optimality of the routing path to improve the overall
network performance. Extensive simulations show that high call-
admission ratio and low-cost paths are achieved with modest
routing overhead. The algorithms can tolerate a high degree of
information imprecision.
Index Terms—Ad hoc quality-of-service (QoS) routing, impre-
cise state information, ticket-based probing.
I. INTRODUCTION
A
N ad hoc network consists of a set of mobile nodes
(hosts) that are equipped with wireless transmitters and
receivers which allow them to communicate without the help
of wired base stations. Since each transmitter has a limited
effective range, distant nodes communicate through multihop
paths with other nodes in the middle as routers. There are
numerous applications for this type of network. A group of
moving soldiers in a battlefield communicates and coordinates
with each other. A group of islands and ships communicates
with the help of floating balloons and passing airplanes. A
group of people with portable computers share their data in a
conference room without laying cables between them.
Much work has been done on routing in ad hoc networks:
the destination-sequenced distance vector (DSDV) protocol
[1], the wireless routing protocol [2], Gafni–Bertsekas algo-
rithms [3], the lightweight mobile routing protocol [4], the
temporally-ordered routing algorithms [5], the dynamic source
routing protocol [6], the associativity-based routing protocol
Manuscript received June 16, 1998; revised February 1, 1999. This work
was supported in part by the Airforce Grant under Contract F30602-97-2-0121
and the National Science Foundation CAREER Grant under Contract NSF
CCR 96-23867.
The authors are with the Department of Computer Science, University
of Illinois at Urbana-Champaign, Urbana, IL 61801 USA (e-mail: s-
chen5@cs.uiuc.edu; klara@cs.uiuc.edu).
Publisher Item Identifier S 0733-8716(99)04802-7.
[7], the spine-based routing algorithms [8], and the zone-
routing protocol [9], etc. Emphasis has been on providing the
shortest-path routing and achieving a high degree of avail-
ability in a dynamic environment where the network topology
changes quickly. However, all the previous routing solutions
only deal with the best-effort data traffic. Connections with
quality-of-service (QoS) requirements, such as voice channels
with delay and bandwidth constraints, are not supported.
The provision of QoS relies on resource reservation. Hence,
the data packets of a QoS connection
1
are likely to flow along
the same network path on which the required resources are
reserved. The goal of QoS routing is twofold: a) selecting
network paths that have sufficient resources to meet the QoS
requirements of all admitted connections and b) achieving
global efficiency in resource utilization.
QoS routing has been receiving increasingly intensive at-
tention in the wireline network domain [10]. The recent work
can be divided into three broad categories: source routing,
distributed routing, and hierarchical routing. In source routing
[11]–[13], each node maintains an image of the global network
state, which is based on a routing path that is centrally
computed at the source node. The global network state is
typically updated periodically by a link-state algorithm [14].
In distributed routing [15]–[18], the path is computed by a
distributed computation during which control messages are
exchanged among the nodes, and the state information kept
at each node is collectively used in order to find a path. In hi-
erarchical routing [19], nodes are clustered into groups, which
are recursively clustered into higher-level groups, creating a
multilevel hierarchy. In every level of the hierarchy, source or
distributed routing algorithms are used.
The QoS routing algorithms for wireline networks cannot
be applied directly to ad hoc networks. First, the performance
of most wireline routing algorithms relies on the availability
of precise state information. However, the dynamic nature
of an ad hoc network makes the available state information
inherently imprecise. Though some recent algorithms [13],
[20] were proposed to work with imprecise information (e.g.,
the probability distribution of link delay), they require the
precise information about the network topology, which is not
available in an ad hoc network. Second, nodes may join, leave,
and rejoin an ad hoc network at any time and any location;
1
A connection (call) is a transport-layer concept. It means: 1) the logical
association between the end users and 2) the correct, ordered delivery of data
[5]. A QoS connection is a connection that has an end-to-end performance
requirement such as a delay or bandwidth constraint. A connection is im-
plemented at the network layer by a network path (routing channel) through
which data packets are delivered.
0733–8716/99$10.00 1999 IEEE
CHEN AND NAHRSTEDT: DISTRIBUTED QOS ROUTING IN AD HOC NETWORKS 1489
existing links may disappear, and new links may be formed as
the nodes move. Hence, the established paths can be broken
at any time, which raises new problems of maintaining and
dynamically reestablishing the routing paths in the course of
data transmission.
It is difficult to provide QoS in an ad hoc network due to
its dynamic nature. The overhead of QoS routing in an ad hoc
network is likely to be higher than that in a wireline network
because the available state information is less precise, and the
topology changes in an unpredicted way. If the topology of an
ad hoc network changes too fast, the provision of the QoS can
be even impossible. However, QoS is feasible in many other
cases where the network topology changes less frequently. For
example, in a conference room, most portable computers may
be stationary most of the time, while some computers move,
join, or leave the room. In this paper, we shall only study the
type of ad hoc networks whose topologies are not changing
that fast to make the QoS routing meaningless. Since we do
not make any specific assumptions ensuring reliable routing
paths in ad hoc networks, we want to emphasize that this paper
only supports soft QoS without hard guarantees. The soft QoS
means that there may exist transient time periods when the
required QoS is not guaranteed due to path breaking or net-
work partition. However, the required QoS should be ensured
when the established routing path(s) remain unbroken. Many
multimedia applications accept soft QoS and use adaptation
techniques to reduce the level of QoS disruption [21]–[23].
For instance, the QoS disruption caused by rerouting in an
ad hoc network can be mitigated by using the rate-adaptive,
layer-based encoded voice/video compression schemes [24].
We propose a distributed QoS routing scheme for ad hoc
networks. Two routing problems are studied. They are delay-
constrained least-cost routing and bandwidth-constrained least-
cost routing. The first one is NP-complete [25]. The second one
is solvable in polynomial time, given precise state information.
A path that satisfies the delay (or bandwidth) constraint is
called a feasible path. We designed routing algorithms for
these problems. The algorithms have the following distinct
properties.
1) The algorithms can tolerate the imprecision of the avail-
able state information. Good routing performance in
terms of success ratio, message overhead, and average
path cost is achieved even when the degree of infor-
mation imprecision is high. Note that the problem of
information imprecision exists only for QoS routing; all
best-effort routing algorithms, such as DSR [6] and ABR
[7], do not consider this problem because they do not
need QoS state in the first place.
2) Multipath parallel routing is used to increase the prob-
ability of finding a feasible path. In contrast to the
flooding-based path discovery algorithms used in [6]
and [7], we search only a small number of paths, which
limits the routing overhead. In order to maximize the
chance of finding a feasible path, the state information
at the intermediate nodes is collectively utilized to make
intelligent hop-by-hop path selection.
3) The algorithms consider not only the QoS requirements,
but also the optimality of the routing path. Low-cost
paths are given preference in order to improve the
overall network performance.
4) In order to reduce the level of QoS disruption, fault-
tolerance techniques are brought in for the maintenance
of the established paths. Different levels of redundancy
provide tradeoff between the reliability and the over-
head. The dynamic path repairing algorithm repairs
the path at the breaking point, shifts the traffic to a
neighbor node, and reconfigures the path around the
breaking point without rerouting the connection along a
completely new path. Rerouting is needed in two cases.
One case is when the primary path and all secondary
paths are broken. The other case is when the cost of
the path grows large and hence it becomes beneficial to
route the traffic to another path with a lower cost.
The rest of the paper is organized as follows. The system
models are given in Section II. The idea of ticket-based
probing is briefly described in Section III. Delay-constrained
routing and bandwidth-constrained routing are studied in
Sections IV and V, respectively. Dynamic path maintenance is
discussed in Section VI. The simulation results are presented
in Section VII. The related work is studied in Section VIII.
Finally, Section IX concludes the paper.
II. S
YSTEM MODELS
A. Ad Hoc Network Model
A network is modeled as a set
of nodes that are inter-
connected by a set
of full-duplex directed communication
links.
and are changing over time when nodes move,
join, and leave. Each node has a unique identifier and has
at least one transmitter and one receiver. Assume that the
effective transmission distance of every node is equal. Two
nodes are neighbors and have a link between them if they
are in the transmission range of each other. We assume there
exists a neighbor discovering protocol. Each node periodically
transmits a BEACON packet identifying itself [7], [26], so
that any node
knows the set of its neighbors. Neighboring
nodes share the same wireless media, and each message is
transmitted by a local broadcast. We assume the existence of a
MAC protocol, which resolves the media contention, supports
resource reservation, and ensures that, among the neighbors
in the local broadcast range, only the intended receiver keeps
the message, and the other neighbors discard the message.
An example of such a MAC protocol, one which supports
bandwidth reservation, was proposed by Gerand and Tsai [27].
B. Stationary and Transient Links
The links between the stationary or slowly moving nodes
are likely to exist continuously. Such links are called stationary
links. The links between the fast moving nodes are likely to
exist only for a short period of time. Such links are called
transient links. A routing path should use stationary links
whenever possible in order to reduce the probability of a path
breaking when the network topology changes [7].
Given the unpredictable nature of an ad hoc network, it
is impractical to determine exactly which links are stationary
1490 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 8, AUGUST 1999
and which are transient. However, various approximation ap-
proaches exist. One simple approach is based on an empirical
observation that moving nodes are more likely to move at
the next moment, and stationary nodes are more likely to
be stationary at the next moment. An immediate implication
is that the links that are just formed are more likely to be
broken than the links that have already existed for some time.
Therefore, whenever a new link is formed, it is set as a
transient link. After the link remains unbroken for a time
period, it is changed to be a stationary link. This approach
is similar to the one proposed in [7]. Let
be is a
stationary link,
. A node in is called a stationary
neighbor
2
of , and a node in is called a transient
neighbor.
C. QoS State Metrics
A node
is assumed to keep the up-to-date local state
about all outgoing links.
3
The state information of link
includes: 1) delay , the channel delay of the link, in-
cluding the radio propagation delay, the queuing delay, and
the protocol-processing time; 2) bandwidth
, the residual
(unused) bandwidth of the link; and 3) cost
, which can be
simply one as a hop count or a function of the link utilization.
In order to make a preference of stationary links over transient
links, the cost of a transient link should be set much higher
than that of a stationary link. The delay, bandwidth, and cost
of a path
are defined as follows:
delay
delay delay
bandwidth bandwidth bandwidth
cost cost cost
D. Routing Problems
Given a source node
, a destination node , and a delay
requirement
, the problem of delay-constrained routing is to
find a feasible path
from to such that delay .
When there are multiple feasible paths, we want to select the
one with the least cost. Finding the delay-constrained least-cost
path is an NP-complete problem [25].
Another problem is bandwidth-constrained routing, i.e.,
finding a path
such that bandwidth , where
is the bandwidth requirement. When there are multiple such
paths, the one with the least cost is selected.
Finding a feasible path is actually the first part of the
problem. The second part is to maintain the path when the
network topology changes [28].
E. Imprecise State Model
The following end-to-end state information is required to
be maintained at every node
for every possible destination .
The information is updated periodically by a distance-vector
2
It should be noted that stationary neighbors (links) are only relatively
stationary by definition.
3
We assume every node has the precise information about its local state.
The problem of information imprecision only exists for the global state.
protocol for mobile computers. Readers are referred to [1] for
a detailed description of such a protocol.
4
1) Delay:
keeps the minimum end-to-end delay from
to , i.e., the delay of the least-delay path.
2) Bandwidth:
keeps the maximum end-to-end band-
width from
to , i.e., the bandwidth of the largest-
bandwidth path.
3) Cost:
keeps the least end-to-end cost from to ,
i.e., the cost of the least-cost path.
The previous information is inherently imprecise in an ad
hoc network because the network state and topology may
change at any time.
The imprecision model proposed by Guerin and Orda [13]
for wireline networks is based on probability distribution
functions. For instance, every node maintains, for every link
,
the probability
of link having a delay of units. This
model is not suitable for an ad hoc network where links may
be short-lived [7] and do not give enough time for collecting
the probability distribution. In contrast, we propose a simple
imprecision model that does not rely on the topology and
can be easily implemented. Two additional state variables are
required.
a) Delay variation:
keeps the estimated maximum
change of
before the next update. That is, based
on the recent state history, the actual minimum end-
to-end delay from
to is expected to be between
and in the next update
period.
b) Bandwidth variation:
keeps the estimated maxi-
mum change of
before the next update. The actual
maximum bandwidth from
to is expected to be
between
and in the next
period.
For the purpose of simplicity, we do not apply the imprecise
model to
. The cost metric ( ) is used for optimiza-
tion, in contrast to the delay and bandwidth metrics used in
QoS constraints. Since a strict cost bound requirement does not
exist, a certain degree of imprecision for
is tolerable.
In the following, we describe a possible way to calculate
. can be computed similarly. is up-
dated periodically together with
. Consider an arbitrary
update of
and . Let and be
the values of
before and after the update, respectively.
Similarly, let
and be the values of
before and after the update, respectively. is provided
by a distance-vector protocol.
is calculated as
follows:
The previous formula is similar to the one used by transmission
control protocol (TCP) to estimate the round-trip delay. The
factor
( ) determines how fast the history information
(
) is forgotten, and determines how fast
converges to .
4
Because our algorithms can tolerate high degree of information impreci-
sion, a relatively low updating frequency is allowed, which leads to better
scalability.
CHEN AND NAHRSTEDT: DISTRIBUTED QOS ROUTING IN AD HOC NETWORKS 1491
By the previous formula, it is still possible for the actual
delay to be out of the range
.
One way to make such probability sufficiently small is to
enlarge
. Hence, we shall modify the formula and
introduce another factor
( )
converges to at a speed
determined by
.
It should also be noted that our imprecision model and
routing algorithms do not intend to cover every possible
situation, which is impractical in an ad hoc network. Our
objective is to improve the average performance, and the
proposed algorithms based on the previous model may fail
in finding a feasible path in the extreme cases where the state
and the topology change very rapidly.
III. A
N OVERVIEW
We propose a multipath
5
distributed routing scheme, called
ticket-based probing.
6
Our design is based on the following
observations.
• The QoS routing is done on a per-connection basis.
Hence, the routing overhead is one of the major concerns.
We shall not use any flooding path-discovery approaches,
which may send routing messages to the entire network.
Instead, we want to localize the routing activity in a
portion of the network between the source
and the
destination
. More specifically, we want to search only
a small number of paths from
to , instead of making
an expensive exhaustive search.
• There are numerous paths from
to . We shall not
randomly pick several paths to search. Instead, we want
to make an intelligent hop-by-hop path selection to guide
the search along the best candidate paths.
The basic idea of ticket-based probing is outlined below.
A ticket is the permission to search one path. The source
node issues a number of tickets based on the available state
information. One guideline is that more tickets are issued
for the connections with tighter requirements. Probes (routing
messages) are sent from the source toward the destination to
search for a low-cost path that satisfies the QoS requirement.
Each probe is required to carry at least one ticket. At an
intermediate node, a probe with more than one ticket is allowed
to be split into multiple ones, each searching a different
downstream subpath. The maximum number of probes at any
time is bounded by the total number of tickets. Since each
probe searches a path, the maximum number of paths searched
is also bounded by the number of tickets. See Fig. 1 for an
example.
Upon receipt of a probe, an intermediate node decides, based
on its state: 1) whether the received probe should be split and
5
Search multiple paths for a feasible one.
6
A preliminary version of the ticket-based probing algorithm for wire-
line networks was published in Proc. IEEE 7th Int. Conf. on Computer,
Communications and Networks (IC3N’98).
Fig. 1. Two probes,
p
1
and
p
2
, are sent from
s
. The number in the
parentheses following a probe is the number of tickets carried in the probe.
At node
j; p
2
is split into
p
3
and
p
4
, each of which has one ticket. There
are at most three probes at any time. Three paths are searched, and they are
s
!
i
!
t
,
s
!
j
!
t
, and
s
!
j
!
k
!
t
.
2) to which neighbor nodes the probe(s) should be forwarded.
The goal is to collectively utilize the state information at the
intermediate nodes to guide the limited tickets (the probes
carrying them) along the best paths to the destination, so
that the probability of finding a low-cost feasible path is
maximized. A number of advantageous properties of the ticket-
based probing are outlined below.
1) The routing overhead is controlled by the number of
tickets, which allows the dynamic tradeoff between the
overhead and the routing performance. Issuing more
tickets means searching more paths, which results in a
better chance of finding a feasible path at the cost of
higher overhead.
2) The proposed scheme is designed to work with imprecise
state information. The level of imprecision (information
uncertainty) has a direct impact on the number of tickets
issued. Multipath parallel search increases the chance
of finding a feasible path and thus helps to tolerate
information imprecision.
3) A distributed routing process is used to avoid any cen-
tralized path computation that could be very expensive
for QoS routing in large networks. It is not necessary
for any node to maintain the topology information. The
most current topology is used during the hop-by-hop
path selection process.
4) The information at the intermediate nodes, both local
and end-to-end states, are collectively used to direct
the probes along the low-cost feasible paths toward the
destination. This approach not only increases the chance
of success but also improves the ability to tolerate the
information imprecision because the intermediate nodes
may gradually correct a wrong decision made by the
source. Stationary links are used whenever possible,
which makes the routing path more stable.
Our ticket-based probing approach is proposed as a gen-
eral QoS routing scheme, which can handle different QoS
constraints. Delay-constrained routing and the bandwidth-
constrained routing are the most studied QoS routing problems
[12], [13], [29], [30]. In the sequel, we shall use them as
examples to explain the operation details.
1492 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 17, NO. 8, AUGUST 1999
IV. DELAY-CONSTRAINED ROUTING
Based on the idea of ticket-based probing, we propose a
heuristic algorithm for the NP-complete, delay-constrained,
least-cost routing problem. When a connection request arrives
at the source node, a certain number
of tickets are
generated, and probes are sent toward the destination
. Each
probe carries one or more tickets. Since no new tickets are
allowed to be created by the intermediate nodes, the total
number of tickets is always
, and the number of probes is
at most
at any time. When a node receives a probe with
tickets, it makes at most copies of , distributes
the received tickets among the new probes, and then forwards
them along to selected outgoing links toward
. Each probe
accumulates the delay of the path it has traversed so far. A
probe can proceed only when the accumulated delay does not
violate the delay requirement. Hence, any probe arriving at
the destination detects a feasible path, which is the one it has
traversed.
There are two problems: 1) how to determine
and 2)
how to distribute the tickets in a received probe among the new
probes. We will solve these problems based on the following
two basic guidelines.
1) We want to assign different numbers of tickets to differ-
ent connections based on their “needs.” For a connection
whose delay requirement is large and can be easily
satisfied, one ticket is issued to search a single path; for
a connection whose delay requirement is smaller, more
tickets are issued to increase the chance of finding a
feasible path; for a connection whose delay requirement
is too small to be satisfied, no tickets are issued, and the
connection is immediately rejected.
2) When a node
forwards the received tickets to its
neighbors, the tickets are distributed unevenly among
the neighbors, depending on their chances of leading
to reliable low-cost feasible paths. A neighbor having
a smaller end-to-end delay to the destination should
receive more tickets than a neighbor having a larger
delay; a neighbor having a smaller end-to-end cost to the
destination should receive more tickets than a neighbor
having a larger cost; a neighbor having a stationary link
to
should be given preference over a neighbor having
a transient link to
. Note that some neighbors may not
receive any tickets because
may have only a few or
just one ticket to forward.
A. Determining the Number of Tickets
1) Yellow Tickets and Green Tickets: The
tickets are
colored either yellow or green. The two types of tickets have
different purposes.
1) The purpose of yellow tickets is to maximize the proba-
bility of finding a feasible path. Hence, yellow tickets (or
more precisely, the probes carrying them) prefer paths
with smaller delays, so that the chance of satisfying a
given delay requirement is higher.
2) The purpose of green tickets is to maximize the prob-
ability of finding a low-cost path. Green tickets prefer
the paths with smaller costs, which may, however, have
Fig. 2. Curves of
Y
0
and
G
0
with respect to
D
.
larger delays and hence have less chance to satisfy the
delay requirement
.
The overall strategy is to use the more aggressive green
tickets to find a low-cost feasible path with relatively low
success probability and to use the yellow tickets as a backup to
guarantee a high success probability of finding a feasible path.
, where is the number of yellow tickets, and
is the number of green tickets. We show how to determine
and in the following section.
2) Number of Yellow Tickets:
is determined based on
the delay requirement
.If is very large and can be surely
satisfied, a single yellow ticket will be sufficient to find a
feasible path. If
is too small to be possibly satisfied, no
yellow ticket is necessary, and the connection is rejected.
Otherwise, more than one yellow ticket is issued to search
multiple paths for a feasible one. Based on the previous
guideline, we choose a linear ticket curve in Fig. 2 (upper
curve) for simplicity and efficient computation. The curve is
explained in the following and the three system parameters
are defined in Table I.
Let
and be the source and the destination, respectively.
is a function of , , and .
1) If
, then . Because is
equal to or greater than the largest possible end-to-end
delay (
),
7
a single yellow ticket will be
sufficient to find a feasible path.
2) If
, then
,
where
is a system parameter specifying the maximum
allowable number of yellow tickets. It shows that more
yellow tickets are assigned for smaller
.
3) If
, then . Because
is even less than the best expected end-to-end delay
(
), such a tight delay requirement will
7
By our imprecise state model, the actual end-to-end delay is expected to
be in
[
D
s
(
t
)
0
1
D
s
(
t
)
;D
s
(
t
)+1
D
s
(
t
)]
. The probability for the delay to
be out of the range is assumed to be negligibly small.