﻿SHARP: A Hybrid Adaptive Routing Protocol for
Mobile Ad Hoc Networks ∗
Venugopalan Ramasubramanian Zygmunt J. Haas Emin Gün Sirer
Cornell University,
Ithaca, NY 14853
{ramasv@cs.cornell.edu, haas@ece.cornell.edu, egs@cs.cornell.edu}
ABSTRACT
A central challenge in ad hoc networks is the design of routing
protocols that can adapt their behavior to frequent and
rapid changes in the network. The performance of proactive
and reactive routing protocols varies with network characteristics,
and one protocol may outperform the other in different
network conditions. The optimal routing strategy depends on
the underlying network topology, rate of change, and traffic
pattern, and varies dynamically. This paper introduces the
Sharp Hybrid Adaptive Routing Protocol (SHARP), which
automatically finds the balance point between proactive and
reactive routing by adjusting the degree to which route information
is propagated proactively versus the degree to which it
needs to be discovered reactively. SHARP enables each node
to use a different application-specific performance metric to
control the adaptation of the routing layer. This paper describes
application-specific protocols built on top of SHARP
for minimizing packet overhead, bounding loss rate, and controlling
jitter. Simulation studies show that the resulting
protocols outperform the purely proactive and purely reactive
protocols across a wide range of network characteristics.
Categories and Subject Descriptors
C.2.2 [Computer-Communication Networks]: Network
Protocols—Routing Protocols
General Terms
Design, Performance
Keywords
Hybrid, Adaptive, Ad Hoc Networks
∗
This work was supported in part by ONR grant N00014-
01-0968, DARPA under AFRL grant RADC F30602-99-1-
0532 and the MURI programs administered by ONR grant
N00014-00-1-0564, by AFOSR grant F49620-02-1-0233, and
by NSF grant ANI-0081357. The views and conclusions contained
herein are those of authors and should not be interpreted
as necessarily representing the official policies or
endorsements, either expressed or implied, of these organizations
or of the Government.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
MobiHoc’03, June 1–3, 2003, Annapolis, Maryland, USA.
Copyright 2003 ACM 1-58113-684-6/03/0006 ...$5.00.
1. INTRODUCTION
Ad hoc networks are characterized by frequent change.
Many of the diverse application areas for ad hoc networks,
including emergency relief operations, battlefield applications
and environmental data collection, exhibit a high degree
of temporal and spatial variation. Nodes may join the
network at any time, get disconnected as they run out of
power, or alter the physical network topology by moving
to a new location. Link characteristics, such as bit error
rates and bandwidth, change frequently due to external factors
like interference and radio propagation fading. Traffic
patterns in the network can shift drastically as applications
modify their behavior and redistribute load within the network.
Consequently, a primary challenge in ad hoc networks
is the design of routing protocols that can adapt their behavior
to rapid and frequent changes at the network level.
Ad hoc routing protocols proposed to date fall between
two extremes based on their mode of operation. Proactive
protocols, such as DSDV [22], TBRPF [19], and OLSR [9],
exchange routing information periodically between hosts,
and constantly maintain a set of available routes for all
nodes in the network. In contrast, reactive protocols, such as
AODV [25], DSR [11], and TORA [20], delay route discovery
until a particular route is required, and propagate routing
information only on demand. There are also a few hybrid
protocols, such as ZRP [7], HARP [18], and ZHLS [10], that
combine proactive and reactive routing strategies.
There is a fundamental trade-off between proactive dissemination
and reactive discovery of routing information.
While proactive protocols can provide good reliability and
low latency through frequent dissemination of routing information,
they entail high overhead and scale poorly with
increasing numbers of participating nodes. In contrast, reactive
protocols, can achieve low routing overhead, but may
suffer from increased latency due to on-demand route discovery
and route maintenance. Since the characteristics of
a practical network vary dynamically with time, choosing
an appropriate routing protocol is an important design and
implementation decision. A protocol suited for a given network
size, density, and mobility may behave inefficiently as
the network characteristics and application behavior change.
In this paper, we present the Sharp Hybrid Adaptive Routing
Protocol (SHARP), which utilizes this fundamental tradeoff
between proactive versus reactive routing to find a good
balance between route information propagated proactively
and route information that is left up to on demand discovery.
SHARP utilizes both a proactive and a reactive protocol to
perform routing. Each SHARP node determines the net-work neighborhood, called proactive zone, in which routing
information pertaining to itself is disseminated proactively.
SHARP relies on a novel proactive routing algorithm that is
both efficient and analytically tractable. However, SHARP
can use any reactive routing algorithm whose costs can be
characterized analytically; our current implementation uses
off-the-shelf AODV. SHARP finds the “sweet spot” between
the two routing regimes by dynamically adjusting the extent
of proactive and reactive routing. This boundary is determined
by an analytical model derived herein and guided
by dynamically-performed empirical measurements from the
physical network.
Requirements for network performance vary among applications.
Multi-media applications can tolerate high loss
rates, but are sensitive to variations in delay. TCP traffic is
sensitive to loss in the network, while devices running on battery
power are concerned with the routing overhead. However,
applications have no control over the performance of
traditional routing protocols. In contrast, SHARP enables
each application to pursue different quantitative metrics for
guiding the inherent trade-off between increased overhead
for proactive information dissemination versus reduced latency
and loss rate. Each SHARP node can separately pursue
different application-specific performance guarantees. For
instance, one node may direct SHARP to adjust its route
dissemination to reduce delay jitter, while another node concurrently
uses SHARP to minimize packet overhead. This
paper presents three SHARP-based protocols for pursuing
application-specific goals, namely, minimizing packet overhead,
bounding loss rate, and controlling delay jitter.
An adaptive hybrid routing protocol requires the following
three properties for successful deployment.
• Adaptive: The protocol should be applicable to a
wide range of network characteristics. It should automatically
adjust its behavior to achieve target goals in
the face of changes in traffic patterns, node mobility
and other network characteristics.
• Flexible: The protocol should enable applications to
optimize for different application-specific metrics at
the routing layer. These optimization goals should not
be set by the network designer, but be placed under
the control of the network participants.
• Efficient and Practical: The protocol should achieve
better performance than pure, non-hybrid, strategies
without invoking costly low-level primitives such as
those for distributed agreement or reliable broadcast.
Through a combination of protocol design, model, and
mechanisms, SHARP’s hybridization approach provides all
of these properties.
SHARP achieves its performance through low cost mechanisms
to determine zone sizes and control the extent of
proactive routing. SHARP nodes monitor traffic pattern
and local network characteristics such as link failure rate
and node degree without incurring excessive overhead. The
zone sizes are then determined by each node in isolation,
solely based on local information. An efficient control mechanism
allows SHARP to shrink or grow the region of proactive
routing without excessive control and synchronization
overhead. Where adjacent zones overlap, SHARP takes advantage
of the overlap to disseminate proactive route information
more efficiently. Finally, SHARP does not require
distributed coordination to determine zone boundaries, and
does not necessitate a costly reliable broadcast mechanism.
Overall, this paper makes the following contributions. First,
it presents a novel hybrid and adaptive protocol that provides
application-level control over its performance by integrating
a reactive and a proactive routing algorithm over
temporal and spatial domains. The protocol embodies a lowoverhead
mechanism for adaptation and an analytical model
to guide the trade-off between the proactive and reactive
routing regimes. Second, it describes the application of this
protocol to achieve three separate goals, namely, minimizing
packet overhead, controlling delay jitter, and bounding
loss rate. SHARP enables multiple nodes in the network to
pursue disparate goals at the routing layer. Finally, it motivates
the case for hybrid adaptive routing protocols by showing
that the ideal point of operation for achieving minimal
packet overhead, bounded loss rate, and controlled delay jitter
lies between purely reactive and purely proactive routing
regimes. An extensive, 600-node simulation study demonstrates
that the SHARP hybrid protocol performs better
than its pure components across a wide range of network
conditions.
The rest of this paper is organized as follows. In the
next section, we discuss related work on hybrid and adaptive
routing protocols. Section 3 provides an overview of
SHARP’s operation. Section 4 contains a detailed description
of the routing protocols employed by SHARP and Section
5 outlines the analytical model that drives our adaptation
strategy. Mechanisms and strategies for adaptation
employed by SHARP are described in Section 6. Section 7
demonstrates that SHARP performs well under a wide range
of network conditions. We summarize our contributions and
observations in Section 8.
2. RELATED WORK
While most of the routing protocols proposed to date for
ad hoc networks are either purely reactive or purely proactive,
some hybrid protocols have also been proposed. However,
only a few hybrid protocols embody fine-grained control
over the extent of hybridization, while none allow the
participating node to pursue disparate application-specific
goals. In this Section, we provide a brief overview of hybrid
and adaptive routing protocols and summarize how they differ
from SHARP.
The Zone Routing Protocol [7] (ZRP) was the first hybrid
routing protocol with both a proactive and a reactive
routing component. ZRP defines a zone around each node
consisting of its k-neighborhood. Routing within a zone is
performed using a proactive routing protocol and routing
between nodes in different zones is performed by an ondemand
routing protocol. ZRP performs efficient route discovery
through bordercasting; route requests are spread by
multicasting them directly to the nodes on the border of its
zone. The size of the zone is dynamically determined based
on network load [21, 26].
While SHARP shares with ZRP the insight that hybrid
routing offers a flexible way to adapt between reactive and
proactive routing strategies, the two approaches differ in several
fundamental ways. First, ZRP is optimized for the single
goal of reducing routing overhead, whereas SHARP enables
application-specific adaptation strategies to bound loss
rate and control jitter, in addition to controlling the overhead
of the routing protocols. Second, ZRP defines a proactive
zone around every node in the network irrespective of its
participation in carrying data packets. In contrast, SHARPmaintains proactive routing zones only around those nodes
that have significant incoming data. Finally, the proactive
routing component of ZRP is more expensive than SHARP’s
proactive routing component. ZRP’s use of bordercasting in
order to efficiently propagate route requests involves maintaining
a multicast tree in each zone with the center node as
the root of the tree and the border nodes as the multicast
destinations. In order to maintain this tree, the intra-zone
routing protocol needs to provide routes between every pair
of nodes in the proactive zone. In SHARP, the proactive
routing protocol only maintains routes to the center node.
This leads to a fundamental difference in the determination
of the optimal zone size. ZRP decreases the size of the proactive
zone as mobility and, correspondingly, the frequency of
link-failures increase, whereas SHARP increases the size of
the proactive zone in response to link failures.
ZHLS [10], Zone-based Hierarchical Link State, is another
zone-based hybrid routing protocol, which partitions the
network into non-overlapping zones based on physical location
information. It performs zone assignments once and
zone sizes do not vary dynamically. HARP [18], Hybrid
Ad-hoc Routing Protocol, is a hybrid protocol that combines
proactive and reactive approaches. It relies on the
Distributed Dynamic Routing [17] protocol for decomposing
the network into zones. A set of forwarding nodes in
each zone is responsible for communicating with nodes in
other zones. HARP uses its own custom protocol for interzone
routing, whose main goal is to reduce delays through
early path maintenance. While HARP creates zones of variable
sizes, it has no direct control over the zones and does
not dynamically adjust their sizes.
In addition, there are also several routing protocols that
adapt based on network characteristics to enhance performance.
FSR [6], Fisheye State Routing, is a link-state protocol
that exchanges periodic link-state information. The period
of link state propagation is determined by the distance
to the destination. ADV [1] is Adaptive Distance Vector algorithm
that exhibits on-demand characteristics by varying
the frequency and size of routing updates. Some researchers
have examined supplanting reactive protocols with timerdirected
route discoveries to produce backup routes prior
to losing the primary link [14]. Their protocol uses a fixed
timer value across all nodes, which is determined offline from
a past history of link-failure statistics.
Several researchers have examined the behavior of reactive
and proactive routing protocols through simulations [2,
5]. Our primary goal in this paper is not to perform protocol
comparisons, but to identify and evaluate the trade-off
between the two routing regimes. Our measurements show
that the choice of the optimal routing algorithm is highly
dependent on network and data traffic characteristics, and
that no single point on the continuum that spans reactive
and proactive protocols is well suited for dynamic networks.
3. OVERVIEW OF SHARP
The Sharp Hybrid Adaptive Routing Protocol adapts efficiently
and seamlessly between proactive and reactive routing
strategies. This adaptation is driven by the measured
characteristics of the network and can be directed to optimize
for user-defined performance metrics, such as loss rate,
routing overhead, or delay jitter.
SHARP adapts between reactive and proactive routing
by dynamically varying the amount of routing information
F
E
B
A
Figure 1: The SHARP proactive zones constructed
around destinations A, B, and C. The vertices correspond
to nodes, directed edges represent route table entries.
The size of the proactive zone is adjusted independently
by each destination.
shared proactively. It does so by defining a proactive zone
around some nodes. A node-specific zone radius determines
the number of nodes within a given proactive zone. Each
node at a distance less than or equal to the zone radius is
a member of the proactive zone for that node. All nodes
not in the proactive zone of a given destination use reactive
routing protocols to establish routes to that node.
Node-specific proactive routing is employed within a proactive
zone. Nodes within the proactive zone maintain routes
proactively only to the central node.
SHARP creates proactive zones automatically around hot
destinations that receive data from many sources. The proactive
zones act as collectors of packets, forwarding the packets
efficiently to the destination, once the packets reach any
node at the zone periphery. SHARP amortizes the cost of
maintaining routes to a given destination in a proactive zone
among all the sources that communicate with that destination
node. Hence, it is well suited for applications that
exhibit spatial locality in their network communications.
For example, in sensor networks, selected nodes performing
data aggregation act as hot destinations for data generated
by nearby sensors. In office networks, gateways or
service providers are contacted by several nodes in the network.
In many general-purpose applications, including messaging,
and wireless audio, the popularity of nodes follows a
Zipf-distribution [4], making some nodes more popular than
others.
Figure 1 shows the SHARP topology in a typical deployment.
SHARP maintains proactive routing zones around
popular destinations A, B, and C. It achieves this by dynamically
adapting the zone radius at each destination based
on incoming data traffic and mobility in the network. Consequently,
it will create relatively large zones around popular
destinations, and relatively small proactive zones around
nodes that get little traffic. For instance, nodes with little or
no data traffic, such as D, E, and F , will have no proactive
routing zone and will rely purely on reactive routing.
The zone radius acts as a virtual-knob to control the mix
of proactive and reactive routing for each destination. By
increasing the radius, SHARP can decrease the loss rate and
the variance in delay, but will pay more in packet overhead to
maintain routes in a larger zone. By decreasing the radius,
SHARP can reduce routing overhead, as fewer nodes need to
be proactively updated; however, it may pay more in delay
jitter and experience higher loss rates. Using this trade-off,
SHARP can act as a completely reactive protocol by setting
the zone radius of all the nodes to zero. Conversely, SHARP
C
Dcan emulate a completely proactive protocol by setting the
radii to equal the network diameter.
The primary challenge in the design of a hybrid protocol is
how to determine the optimal trade-off between the components
of the hybrid. Ideally, a hybrid protocol would achieve
fine-grained control over this trade-off, incur low overhead
for adaptation and exploit information locality for maximum
efficiency. In the next sections, we describe the proactive
and the reactive components of SHARP and the analytical
model by which it controls this trade-off efficiently.
4. SHARP ROUTING PROTOCOLS
SHARP is composed of a proactive routing component
and a reactive routing component. This section describes
the two routing components of SHARP in detail.
4.1 Proactive Routing Component
The SHARP Proactive Routing protocol (SPR) is an efficient
protocol engineered with techniques borrowed from
different routing algorithms such as Destination Sequenced
Distance Vector (DSDV) [22] and Temporally Ordered Routing
Algorithm (TORA) [20]. SPR maintains routes to a single
destination in each zone. This enables SPR to employ
low overhead mechanisms to build and maintain routes and
to offer several good properties, including predictable routing
overhead, low loss of data packets, and low variance in
packet inter-arrival time.
SPR performs proactive routing by building and maintaining
a directed acyclic graph (DAG) rooted at the destination.
The DAG is built by employing a construction
protocol initiated periodically by the destination once every
reconstruction interval. The DAG is restricted to nodes
within the proactive-zone of the destination. In between
reconstruction, it is maintained by an update protocol that
relies on periodic exchange of updates and local recovery
mechanisms to handle link-failures.
The construction protocol is used to generate a DAG periodically.
The destination node initiates the construction
process by generating a DAG construction packet. This
packet carries the zone radius and a sequence number to
distinguish the new DAG from the old DAG. Its time-tolive
(TTL) field is set to the zone radius and is propagated
within the proactive zone by a broadcast. It builds a DAG
by assigning a height to each node. The height is used for
routing and corresponds initially to the distance of the node
from the destination. A node B that receives the construction
packet from a node A locally adds the A → B link
to the DAG and rebroadcasts the construction packet after
incrementing its height by one. A node may receive the
construction packet from multiple neighbors along different
links. In this case, the node stores the height information
of all neighbors, sets its own height to the minimum of the
height values it has received.
A naive scheme that uses flooding to disseminate construction
packets suffers from packet loss due to collisions [15].
To avoid broadcast floods, each node waits for a random
interval of time, bounded by T , before retransmitting the
construction packet. However, this might cause imbalance
in the construction of the DAG, since a node might receive
the construction packet from nodes farther away from the
destination before receiving the packet from nodes closer to
the destination. To compensate, each node waits for W seconds
after receiving the first construction packet. It collects
all construction packets received during this time, sets its
height to the minimum of all height values it has observed,
and forwards a single construction packet, if the TTL is positive.
Construction packets arriving after W seconds do not
impact the height chosen by the node. While this algorithm
does not guarantee that all the nodes will detect the shortest
distance to the destination, a conservative choice for W
ensures that the DAG will not deviate significantly from the
optimal. Most importantly, each node forwards a packet
at most once, thereby bounding the number of construction
packets within a zone by the number of nodes in the zone.
An update protocol is used for beaconing and maintenance
of the DAG during link-failures. Once every beacon interval,
each node in the proactive zone broadcasts its current height
in an update packet, whose time-to-live is set to one. Each
node records the height of its neighbors upon receiving update
packets from them. The update packets also serve as
‘hello’ beacons to detect the formation of new links and the
breakage of current links. A new link is formed whenever
an update packet is received from a neighbor for the first
time, while a link is considered broken if at least beacon loss
consecutive update packets are not received from a neighbor.
The update protocol relies on TORA’s failure recovery
mechanisms to restore the DAG in response to link-failures.
Link-failures are detected either when consecutive update
packets are lost, or when the MAC layer is unable to transmit
a data packet to a neighbor. The failure recovery is
invoked only when all of the downstream links at a node
have failed. Following the approach described in [3], the
node reverses the orientation of its upstream links by choosing
a new height greater than the height of its neighbors and
initiating a new update packet. Using a TORA-based failure
recovery provides three properties. First, it allows SPR to
recover from link-failures within the proactive zone without
having to wait until the next reconstruction interval. Second,
it constrains the impact of the link-failure to a small
local region. Finally, as we describe later in Section 5, SPR
incurs low and predictable overhead, which makes the analytical
model more tractable. Note that performing local
repairs distorts the optimality of the DAG, since the correct
distance to the destination is no longer known. This
has been shown to significantly affect the performance of
TORA [2]. SHARP compensates for this problem with the
construction protocol, which periodically rebuilds the DAG
from scratch. Hence, SPR’s DAGs contain relatively short
paths to the destination most of the time.
SPR performs routing of data packets in the following
manner. A data packet arriving at a node is transmitted
along a downstream link to the neighbor with the lowest
height. If the data packet cannot be transmitted to that
neighbor because of a link-failure, it is transmitted to the
next best downstream node. If all the downstream links
have failed, link reversal is performed to repair the DAG
and the data packet is forwarded to the best newly formed
downstream node. Temporary routing loops could be created
(as observed in [2]) if the data packets reach the new
downstream node ahead of the update packet. To prevent
this, the new height of the node is piggybacked with the
data packet, so that the downstream node can update the
new height before forwarding the data packet.
Multi-path routing, local link repairs and the construction
protocol enable SPR to be a robust and efficient protocol,
incurring low loss rate and predictable overhead.4.2 Reactive Routing Component
SHARP’s reactive routing protocol is based on the Ad-hoc
On-demand Distance Vector (AODV) routing protocol [25].
SHARP simply uses standard AODV [23] for reactive routing,
and includes several optimizations such as route caching
and expanding ring search.
SHARP integrates the proactive and the reactive components
seamlessly without incurring additional overhead. If
the source node is within the proactive zone, routing is performed
proactively as described earlier. If the source node is
outside the proactive zone, route requests are broadcast by
AODV. Nodes in the proactive zone of a destination generate
route replies in response to route requests generated by
the source node. The replying node in the proactive zone
sets the distance to the destination as zero. Thus, the proactive
zone effectively acts as a collective destination for the
data packets from the source. Once data packets enter the
proactive zone, they are efficiently routed by the proactive
protocol. Routes between the source node and the edge of
the proactive zone are maintained reactively.
Overall, SHARP is a general framework for hybridization
that can be used with any combination of proactive and reactive
routing protocols whose costs are analytically tractable
with information efficiently available to the route destination.
5. MODEL
An analytical model of proactive and reactive routing
costs enables SHARP to make an informed trade-off between
different routing regimes. This model provides an accurate
estimate of the routing overhead of the proactive component
and an approximate analysis of the overhead of the reactive
routing component.
The notation N A r is used to represent the number of nodes
in a region of radius r around node A. The average lifetime
of a communication link is represented by λ. We assume
that link lifetimes are independent of each other and are
exponentially distributed. Consider a source node S that
transmits data to a destination node D, which is located
at a distance of h hops from S. Let a proactive zone of
radius r hops exist around the destination node D. The
SHARP proactive routing protocol (SPR) is employed for
routing within this proactive zone, while the reactive routing
protocol is used by S for discovering routes to the proactive
zone. This setup is illustrated in Figure 2.
Proactive Routing Overhead
The routing overhead of SPR has two components. A portion
of the overhead is fixed and stems from the periodic
control packets used to generate the DAG. The remaining
overhead stems from event triggered control packets generated
as a result of link-failures. The periodic overhead in
turn consists of control packets sent during periodic DAG
reconstruction, as well as update packets sent every beacon
interval. Let fu be the frequency at which update packets
are generated by each node in the proactive zone and
let fc be the frequency at which construction is started by
the destination node D. The per-node overhead for sending
update packets is fu packets per second. Since during construction,
each node in the proactive zone broadcasts the
construction packet exactly once, the per-node overhead for
construction is fc packets per sec. Thus, the total fixed over-
h
AODV
S
h−r
h
Figure 2: This figure shows a route setup between the
source node S and a destination node D at a distance of h
hops. A SHARP proactive zone of radius r is maintained
around the destination node D, while the source node S
employs AODV to find routes to the proactive zone.
head of SHARP proactive routing component is N D r (fu+fc)
packets per second in a proactive zone of radius r.
The fixed component of the routing overhead of SPR is
independent of both the mobility rate and the number of
sources sending data packets to the same destination. If
there are several destinations, the same node in the network
could belong to the proactive zones of multiple destinations.
In that case, updates for different destinations are aggregated
into a single packet. Thus the number of periodic
update packets sent does not increase even though the size
of update packets increases with the number of destinations.
If a node belongs to the proactive zone of d destinations, the
fixed SHARP overhead at that node can be given by fu +dfc
packets per second.
Link-breaks within the proactive zone trigger control packets
carrying changes in the height of a node in the proactive
zone of a destination D. SPR employs TORA’s repair
mechanism [3] to update the height of each node upon a
link-failure. Accordingly, the height of a node in the proactive
zone changes only if all downstream links in the DAG
at that node fail. Assuming that different links at a node
are formed and broken independently with the lifetime distributed
exponentially with mean λ, the average frequency
at which all the downstream links fail at a node can be es-
1
2βnλ
r
SPR
D
h
n 1
i=1 i
timated to be (see [24]), where βn equals , and
n is the number of downstream links at that node. In addition,
an event triggered by link-failures at a node alters the
status of several of its upstream links. This results in reduction
of downstream links at other nodes. Incorporating
the effect of this in our earlier estimation, we can revise the
approximation of the frequency of event triggered updates
1
2(βn−1)λ
to . Each destination incurs this overhead. Hence,
the overhead scales linearly with the number of destinations.
The contribution of event-triggered overhead to SHARP’s
proactive routing component is very low. The total routing
overhead of SPR in a proactive zone of radius r around a
destination node D is N D r (fu + fc +
1
2(βn−1)λ
) packets per
second. In practice, the frequency of periodic updates, fu, is
chosen to be much higher than the frequency of link-failure,
1
, in order to detect failures quickly. As a result, the pe-
λ
riodic packets dominate the control overhead of SPR. This
trend is also observed in our simulation results presented
in Section 7. Thus, SHARP’s proactive routing component
presents a predictable routing overhead, which is mostly independent
of the mobility rate.
The size of the proactive zone substantially determines
the overhead of SPR. As the zone radius is increased, theoverhead of SPR increases significantly since, the number of
nodes in a proactive zone, N D r , often increases polynomially
with the zone radius.
Reactive Routing Overhead
Accurate estimation of the routing overhead of a reactive
routing protocol such as AODV is very difficult. Here, we
present an approximate analysis of the routing overhead
of AODV, which helps us to characterize the behavior of
AODV and provide heuristics for adapting the radius of the
SHARP’s proactive zones. The evaluation section shows
that these approximations work well in practice.
AODV expends several control packets, such as route requests,
route replies, and route errors during route discovery
and maintenance [23]. Route requests for a destination D
are generated at the source node S when a route to D is
requested for the first time, or when a route is broken due
to link-failures. Route replies are transmitted in response to
route requests, while route errors are generated when a route
to a destination fails. The number of route error packets is
typically low compared to the number of route requests and
route replies, since the route errors are propagated only by
the nodes in the route between S and D, whereas route requests
and route replies could be propagated by any node
in the network.
The overhead for a source node S to reactively discover a
pack-
route of length h is approximately proportional to N S h
ets, where N S h
is the number of nodes at a distance of at
most h from S. In reality, the source node does not know the
correct distance to destination D. Instead, it restricts the
propagation of route requests to within a few hops of h by
using the expanding ring search. AODV routes fail at an average
rate of h
per second under the assumption of indepen-
λ
dent and exponentially distributed link-failures (see [24]).
Thus, the route maintenance overhead of AODV, for active
routes, is approximately proportional to N S h h
. Other fac-
λ
tors such as congestion could also cause route breaks. In
this analysis, we assume that the impact of congestion is
insignificant compared to the effect of mobility on routing
performance. AODV’s routing overhead increases with the
distance between the source and destination, since longer
routes break more often than shorter ones. Also, destinations
farther away require route requests to be propagated
in a larger area. The routing overhead of AODV is proportional
to the number of active routes in the network. Hence,
AODV’s overhead increases with the number of sources and
the number of destinations in the network. If s sources are
transmitting data packets to a single destination, the total
routing overhead of AODV is approximately proportional
s
to
i=1 N Si
hi hi λ , where Si represents ith source and hi its
average distance from D.
While the model presented here provides a quantifiable
metric that can guide how to modify the proactive radius,
it is still an approximation. The values it computes may
diverge from the actual behavior of the deployed routing
protocols. Optimizations such as route caches, local route
repair, and multiple routes would impact the actual cost observed
in the network. However, we show in the evaluation
section that the model captures the overheads of routing protocols
adequately and leads to the construction of adaptive
hybrid protocols that outperform both the purely proactive
and the purely reactive routing components across a wide
range of network conditions.
In the next section, we discuss the application of this analysis
to the construction of specialized adaptation algorithms.
Since each node makes independent decisions, further discussions
in the paper only describe adaptation at a single
node. However, these protocols apply equally well to multiple
nodes, as the adaptation does not require any consensus
or communication between the participating nodes.
6. SHARP ADAPTATION
SHARP controls the performance of the routing protocol
by dynamically adjusting the zone radius around each destination.
Each destination node varies the size of the proactive
zone around itself by taking into account the network characteristics,
such as the mobility rate and the node-degree, as
well as the data traffic characteristics, such as the number
of sources and the distance of the sources from the destination.
The decision to alter the present zone radius is made
by a destination locally and independent of other destinations.
This enables SHARP to control different applicationspecific
performance metrics, such as routing overhead, loss
rate, and delay jitter, and to have different nodes in the
network that optimize for different metrics simultaneously.
This section explains the mechanisms and the strategies used
to perform the dynamic adaptation in SHARP.
6.1 Adaptation Mechanisms
To enable efficient operation and frequent adaptation, we
need low cost mechanisms for monitoring the network and
changing the size of the zone radius accordingly. In particular,
SHARP requires mechanisms for estimating low-level
network characteristics, such as average link lifetime and average
node-degree, as well as traffic characteristics, such as
loss rate and delay jitter. It also requires an efficient mechanism
to propagate the zone radius and modify the proactive
zone based on these measurements and the analytical model.
SHARP nodes continually monitor network characteristics.
In particular, nodes in a proactive zone measure and
maintain the average link lifetime of their immediate links.
In addition to average link lifetime, the nodes in a proactive
zone also measure the average node-degree, that is, the
number of immediate neighbors. They aggregate the average
link lifetime and the average node-degree with the values
received from other upstream nodes as part of the update
protocol and broadcast the aggregated-values periodically
along with the update packet. These values are combined
throughout the update process, taking care not to doublecount
data, until they ultimately reach the destination at
the center of the DAG. The destination then estimates λ
and computes N D r at the end of an update cycle.
The destination node also locally maintains statistics about
the data traffic that it has received. Each destination keeps
track of the identity of the source, the current distance to
the source, and the average performance in terms of loss rate
and delay jitter over a period of time. To facilitate these
measurements, each source node inserts a SHARP header
as an IP option to every data packet. The SHARP header
contains the number of packets transmitted by the source to
this destination and a field marking the distance to the destination,
which is initialized to 1. Every intermediate node
transmitting this packet increments the distance field, enabling
an up-to-date measurement of the hop-count between
the source and the destination. The destination records the
number of data packets it received and estimates the lossrate in the network from the count of the data packets sent.
The destination keeps track of the arrival times of consecutive
packets and directly computes the delay jitter from the
variance of inter-arrival times. Note that computing the delay
jitter does not require clock-synchronization between the
source and the destination, since a local clock is sufficient to
measure the variance of inter-arrival times. Combined with
the low-level network characteristics, traffic metrics enable
each destination to independently compute the optimal radius
for its proactive zone.
Once a new zone radius has been computed, it needs to
be efficiently disseminated to the nodes in the new proactive
zone. The decision to expand or shrink the zone radius
is made periodically, just before a DAG reconstruction is
initiated by the destination. The destination-centered DAG
is reconstructed as described in Section 4, with the new value
of the radius. Thus, expanding the radius from r to s for s >
r is straightforward and imposes no extra routing overhead
on SHARP proactive protocol.
When the radius is shrunk from r to s for r > s, care
must be taken to purge the nodes in the older DAG that
are no longer part of the newly formed DAG. This is performed
by setting the time-to-live field of the construction
packet to r instead of the new radius s. Nodes receiving this
construction packet terminate all proactive activity for that
destination if they do not belong to the newly built DAG.
In addition, the nodes that do not receive this construction
packet due to network congestion stop receiving periodic update
packets from their neighbors if they do not belong to
the new DAG and, thus, perceiving link-failures will start
purging themselves from the DAG. Thus, this scheme exhibits
graceful degradation without the need for costly reliable
multicast services or distributed consensus protocols.
Decisions to alter the radius of the proactive zone are
taken periodically every reconstruction interval. In order to
impart stability to the process of adaptation, SHARP employs
hysteresis in the form of two thresholds. The up thresh
and the down thresh, with values 1 and 1.5, respectively, determine
the high and low water marks in a given metric,
before a change in the zone radius is triggered.
6.2 Adaptation Strategies
SHARP uses the efficient mechanisms described in the
previous section to dynamically manage the trade-off between
proactive and reactive routing. It provides a general
framework that can be used to achieve different applicationspecific
goals. In this section, we describe applications of
this framework to achieve three separate goals.
Minimal Packet Overhead
Routing overhead is a critical consideration when choosing
a routing protocol. In mobile environments, nodes are typically
limited by battery power. Routing algorithms that
require excessive communication will experience greatly diminished
system longevity.
Our strategy for adapting the zone radius in order to reduce
the routing overhead is based on the previously discussed
trade-offs between proactive and reactive routing components.
The cost of the proactive component is independent
of the number of sources, and it can be amortized across
the multiple sources that are sending packets to the same
destination. Since SPR performs local repair, the overhead
of SHARP’s proactive component is largely independent of
the mobility rate. The cost of reactive routing is tightly
coupled to the number of sources as well as the mobility
rate. As the mobility in the network increases, the reactive
routing component is forced to incur higher aggregate costs
for route discovery. Thus, depending on the instantaneous
values of network parameters, there is an opportunity for
optimization by choosing one routing regime over another.
We propose a protocol for minimizing the per-packet overhead
of routing algorithms based on the SHARP adaptation
framework. Called SHARP-PO, this protocol performs
a dynamic adaptation between predictable, high cost SPR
protocol versus the varying costs of reactive AODV protocol
in order to find the optimal mix of routing overhead.
The goal of the SHARP-PO protocol is to dynamically find
values for proactive radii that optimize the total cost. The
model introduced in Section 5 enables SHARP to quantify
the trade-off between different routing regimes in terms of
overhead. The overhead of routing in a proactive zone of radius
r around the destination D is N D r (fu + fc +
1
2(βn−1)λ ).
The overhead of AODV for each source outside the proactive
zone at a distance h hops from D is approximated by
N S h−r h−r
. Note that the source node only needs to find a
λ
route to some node on the edge of the proactive zone. Hence
only h − r nodes in the path perform reactive routing.
The destination node estimates the incremental difference
in the overhead of the two components, when the radius
is increased by 1, using the above expressions. If the reduction
in the overhead of the reactive component is more
than up thresh times the increase in the overhead of the
proactive component, the radius is incremented by 1. Similarly,
the difference in overhead is computed for the case
of radius decrement and the radius is decremented by 1, if
the reduction in the overhead of the proactive component
is more than down thresh times the increase in overhead
of the reactive component. Otherwise, the radius is kept
unchanged. This algorithm is executed periodically, once
every reconstruction interval. During these estimations, locally
observed values of mean link lifetime and node-degree
are substituted for both SPR as well as AODV.
Bounded Loss Rate
Loss rate is a critical parameter for a network-layer routing
protocol. Higher layer protocols such as TCP are quite
sensitive to packet loss in the underlying layers. A routing
protocol that results in a high loss rate will experience
greatly diminished TCP throughput [8, 16]. Other applications
such as compressed video also exhibit relatively low
tolerance to loss rate.
There is a fundamental trade-off in terms of overhead
and reliability between the proactive and reactive routing
components. SPR constantly maintains destination-directed
DAG with multiple redundant paths. Consequently, it incurs
low loss rates, as it can quickly find alternative routes
in response to link-failures. The reactive protocol detects
route-failures lazily, that is, only when attempting to send
data packets, and hence can suffer from packet loss when
links fail. However at low mobility, the reactive routing component
offers small loss rate at very low cost. The proactive
component, though effective, incurs fixed overhead in regions
of low mobility. This motivates an adaptive strategy
to pick the minimal sufficient proactive radii to guarantee a
targeted loss rate without incurring excessive control overhead.We propose an application-specific protocol, named SHARP-
LR, for achieving a target loss rate specified by the user,
while simultaneously minimizing packet overhead. This protocol
operates by adjusting the proactive zone in response to
perceived loss at the destination, so that the protocol does
not experience loss greater than the targeted rate. SHARP-
LR uses the packet loss measured at the destination as the
driving metric for adaptation. The heuristic used here is
to increase the zone radius by 1, if the perceived loss rate
is up thresh times more than the targeted loss rate and to
decrease the zone radius by 1, if the perceived loss rate is
less than down thresh times the target. The proactive zone
is then expanded or shrunk according to the chosen radius.
Controlled Delay Jitter
Delay jitter, measured as the variance of the inter-arrival
time between packets [12], is a critical performance metric
for many real-time applications. For instance, applications
involving periodic transmission of sensor data, such as object
tracking, as well as multi-media applications for transmitting
uncompressed voice streams are highly sensitive to the
variance in packet arrival times [12].
There is a trade-off between the overhead entailed by the
two routing components and the delay performance they
provide at different network conditions. SHARP’s proactive
component exhibits lower variance in latency in highly mobile
networks. The reactive component reconstructs routes
upon route failures resulting in a temporary halt of data
flow. Consequently, frequent route-failures increase jitter.
In low mobility scenarios, the reactive component offers the
desired jitter performance at a very low overhead compared
to the proactive component. This motivates an adaptation
strategy to utilize the trade-off and provide good jitter performance
at low overhead as the network conditions vary.
Note that, in addition to route failures, jitter is also affected
by congestion in the network. Currently, our protocol does
not take into account the effect of congestion on delay jitter.
Our third protocol, called SHARP-DJ, aims to achieve
an application-specified delay jitter with minimal overhead.
SHARP-DJ aims to pick the minimal sufficient proactive
radii to guarantee a targeted jitter without incurring excessive
overhead. SHARP-DJ employs a heuristic very similar
to SHARP-LR. Packet inter-arrival times are tracked by the
destination as packets are received. If the perceived jitter is
more than up thresh times the target value, the zone radius
is increased by 1. Similarly, if the perceived jitter is less
than down thresh time the target value, the zone radius is
decreased by 1.
7. EVALUATION
In this section, we evaluate the three application-specific
protocols that we introduced in the previous section. We
examine the performance of SHARP-PO, SHARP-LR, and
SHARP-DJ, and show that these protocols perform well in
a variety of settings. We also provide the intuition justifying
why they work well by examining a non-adaptive version of
SHARP with statically selected, fixed values for the zone
radius.
SHARP was implemented in GloMoSim [27], a scalable
packet-level simulator with an accurate radio model. The
many parameters necessary for a realistic simulation were
derived from real-world applications. The MAC protocol
and radio characteristics were set to approximately correspond
to the Lucent WaveLAN TM card with a data rate of
11 Mbps at 2.4 GHz radio frequency operating at 250m nominal
transmission range. IEEE 802.11b was used as the MAC
protocol. The internal parameters of the AODV protocol
were set as suggested in [23]. For SPR, we used the value of 1
second for beacon interval and the value of 2 for beacon loss,
which correspond to standard values suggested in hello protocols
[23]. We used 5 seconds for reconstruction interval,
and 10 milliseconds for T and W in the construction protocol.
The values of up thresh and down thresh were set to 1
and 1.5, respectively.
We performed simulations with several topologies. Here,
we present simulations performed by distributing 200 nodes
uniformly at random in an area of 1700m × 1700m and 600
nodes in an area of 3000m × 3000m. Each simulation was
run for a duration of 360 simulated seconds. The mobility
in the environment was simulated using a random-waypoint
mobility model, wherein each node randomly chooses a point
in the field and moves towards it at a randomly chosen velocity.
The node then pauses for a specified period at the
destination before continuing the same pattern of motion.
In our simulations, velocities ranged between 0 m/s and 20
m/s, while the pause time was set to 0 seconds, which corresponds
to constant motion. We controlled the mobility rate
by varying the number of mobile nodes in the network. A
mobility fraction, F, of 0 corresponds to all stationary nodes,
while a mobility fraction of 1 corresponds to all nodes in
continuous motion.
We show data from two scenarios. In the multi-destination
topology, we examined a network where 4 destinations are
sent data packets from 1, 4, 7, and 10 sources respectively.
The source and destination pairs were chosen randomly, and
the relative popularity of the destinations approximated hot
spots in the network. In the single-destination topology,
we examined a single destination in isolation to provide the
intuition behind the performance of the multi-destination
case. A constant bit rate (CBR) generator drove the data
traffic in our simulation. We set the data traffic rate to
correspond to audio traffic at 16 kbps as described in [13].
Accordingly, each source sent packets of 256 bytes at a rate
of 8 packets per second in the 200-nodes topology. However,
in the 600-nodes case, we had to restrict the data rate to
2 packets per second due to scalability limitations of the
simulator. The sources started transmitting from a time
randomly chosen between 50 seconds and 100 seconds of the
simulation, and stopped data transmission after 250 seconds.
Each data point corresponds to a mean of 30 repeated
measurements with different seeds of the random number
generator. We plot the 95% confidence interval as error bars
on the graphs.
Minimal Packet Overhead
We first demonstrate the need for a hybrid algorithm by examining
routing overhead for different zone radii, and show
that the optimal point lies between purely proactive and
purely reactive routing regimes. In the first experiment,
we set the zone radius to a fixed value for the duration of
each simulation. We vary the zone radius from 0 (corresponding
to AODV) to 13 gradually and then increase it
to a very high value to cover the entire network (marked
as SPR in the figure). Figure 3(a) shows the total packet
overhead when 5 sources send packets to a single destination
in a network of 600 nodes. Figure 3(b) shows the total500
2000
packet overhead (1000 pkts)
400
300
200
single destination
F 0.0
F 0.2
F 0.4
F 0.6
F 0.8
F 1.0
packet overhead (1000 pkts)
1600
1200
800
multi-destination
F 1.0
F 0.8
F 0.6
F 0.4
F 0.2
F 0.0
100
400
0
AODV 1 2 3 4 5 6 7 8 9 10 11 12 13 SPR
zone radius
Figure 3(a): Total packet overhead vs. zone radius for a
single destination with 600 nodes. This graph illustrates
that the optimal routing strategy varies with mobility in
the network.
0
AODV 1 2 3 4 5 6 7 8 9 10 11 12 13 SPR
zone radius
Figure 3(b): Total packet overhead vs. zone radius for multiple
destinations with 600 nodes. This graph shows the
trade-off between reactive and proactive components at different
mobility rates.
500
single destination
2000
multi-destination
packet overhead (1000 pkts)
400
300
200
AODV
SHARP-PO
SPR
packet overhead (1000 pkts)
1600
1200
800
AODV
SHARP-PO
SPR
100
400
0
0 0.2 0.4 0.6 0.8 1
mobility fraction
Figure 4(a): The routing protocol overhead vs. mobility
for a single destination with 600 total nodes. This graph
illustrates that SHARP-PO achieves comparable or better
overhead than the proactive and the reactive components.
0
0 0.2 0.4 0.6 0.8 1
mobility fraction
Figure 4(b): The routing protocol overhead vs. mobility
for multiple destinations with 600 total nodes. This graph
illustrates that SHARP-PO achieves comparable or better
overhead than the proactive and the reactive components.
packet overhead for the multi-destination case, where 1, 4,
7, and 10 sources respectively send packets to the 4 different
destinations in a 600-nodes network. The graphs also show
the packet overhead for different values of mobility fraction
between 0 (static network) to 1 (all nodes in motion).
These graphs validate the trade-off between reactive and
proactive routing components inferred from the model presented
in Section 5. The overhead of the reactive component
gradually increases as the network becomes more mobile.
The local-repair mechanism employed by the proactive
routing component entails very little communication, and
therefore, does not contribute much to the total overhead.
Consequently, the reactive component achieves low overhead
when the mobility is low, while the proactive component incurs
lower overhead when the mobility is high. Figure 3(a)
shows that for high mobility, there are intermediate values
of the zone radius where the packet overhead is less than
both, the purely reactive and the purely proactive routing
components. Thus, no single value of zone radius is the best
choice for all levels of mobility. A protocol that dynamically
adapts the zone radius with changing network conditions is
necessary to achieve optimal performance.
Figures 4(a) and 4(b) show the performance of SHARP-
PO, which adapts the zone radius dynamically to minimize
packet overhead, in the single-destination and the multidestination
simulations with 600 nodes. These graphs illustrate
how SHARP-PO adapts as the mobility in the network
changes. In low mobility conditions, SHARP-PO follows
the overhead of the reactive component, whereas in the
high mobility conditions, SHARP-PO follows the overhead
of the proactive routing component. By finding intermediate
values of zone radii, SHARP-PO entails overhead lower
than both the purely reactive and the purely proactive components
(see Figure 4(a)). When the mobility is very low,
the overhead of SHARP-PO is a little higher than AODV’s
overhead, because our current implementation of SHARP-
PO always operates with a minimum zone radius of 1 in
order to measure network characteristics.
Bounded Loss Rate
We next analyze the rate of packet loss in SHARP, as the
zone radius and mobility are varied. Figure 5(a) shows the
packet loss experienced by SHARP in a network of 600 nodes
with multiple destinations receiving packets. At low mobility,
the packet loss experienced is overall quite low. At high
mobility, the loss rate incurred by the reactive component
increases significantly, while the proactive component continues
to provide good reliability. On the basis of loss rate
alone, SPR appears to be a good choice for all scenarios.
However, the overhead incurred by the routing components40%
35%
30%
25%
multi-destination
F 1.0
F 0.8
F 0.6
F 0.4
F 0.2
F 0.0
40%
35%
30%
25%
AODV
SHARP-LR
SPR
multi-destination
loss rate (%)
20%
loss rate (%)
20%
15%
15%
10%
10%
5%
5%
0%
AODV 1 2 3 4 5 6 7 8 9 10 11 12 13 SPR
zone radius
Figure 5(a): Packet loss rate vs. zone radius in the multidestination
topology with 600 nodes. This graph shows that
proactive dissemination of routing information can reduce
loss rate at the expense of overhead.
introduces a trade-off to be considered while picking the optimal
point of operation. Figure 3(b) shows that the reactive
routing component entails very low overhead and provides
good loss rate during low mobility. The proactive routing
component incurs low loss at all levels of mobility, but entails
very high overhead at low mobility. These graphs validate
the trade-off in terms of loss rate and packet overhead, which
we observed earlier. A dynamically adapting protocol is necessary
to capture this variation with network conditions.
Figure 5(b) shows the performance of SHARP-LR, the
SHARP scheme to achieve a target loss rate, compared to
the purely reactive and the purely proactive component in a
network of 600 nodes with multiple destinations. The target
loss rate was set to 5%. The graph shows that SHARP-LR
is successful in attaining the set target for all levels of mobility
considered. Figure 5(c) shows the packet overhead
incurred by SHARP-LR during its operation. At low mobility,
SHARP-LR chooses less proactivity and achieves the
target loss rate with very low overhead. As the mobility increases,
SHARP-LR increases the amount of proactivity to
achieve the target loss rate. The overall cost of this scheme
is comparable to the minimum of the purely proactive and
the purely reactive components at all times. SHARP-LO
experiences slightly more overhead than the purely proactive
SPR, since it takes some time to choose the appropriate
low overhead zone radius. SHARP-LR efficiently adapts the
zone radius to guarantee the set target loss rate of 5%.
Controlled Delay Jitter
0%
0 0.2 0.4 0.6 0.8 1
mobility fraction
Figure 5(b): Packet loss rate vs. mobility for the multidestination
topology, with 600 nodes and a target loss rate
of 5%. This graph illustrates that SHARP-LR bounds the
loss rate by dynamically varying the zone radius.
packet overhead (1000 pkts)
2000
1600
1200
800
400
0
AODV
SHARP-LR
SPR
Figure 6(a) shows the average delay jitter of SHARP as the
zone radius and the mobility fraction are varied for a network
of 200 nodes with the 4 destinations receiving data at
a rate of 8 packets per second from 1, 4, 7 and 10 sources
respectively. The values of the delay jitter have been multiplied
by the square of the data rate in order to normalize
them. The figure shows that as the mobility increases, the
packet inter-arrival times are less predictable for the reactive
component. This behavior is caused by route maintenance
performed to handle link-failures. By utilizing multiple
available paths, the proactive component is able to
provide very low values of delay jitter. Considering the performance
in terms of delay jitter alone, SPR appears to be
the best choice for routing in all scenarios. However, the
overhead incurred by SPR in providing low delay jitter inmulti-destination
0 0.2 0.4 0.6 0.8 1
mobility fraction
Figure 5(c): Packet overhead vs. mobility for the multidestination
topology, with 600 nodes and target loss rate
of 5%. This graph shows that through dynamic adaptation,
SHARP-LR utilizes the trade-off between loss rate and
packet overhead to achieve its goal efficiently.
troduces a trade-off that needs to be considered while picking
the optimal point of operation. The reactive component
provides good performance in terms of delay jitter at a much
lower cost than the proactive component when the mobility
rate is low. This validates the earlier observed trade-offs
between delay jitter and overhead of reactive and proactive
components and justifies the demand for a dynamically
adapting protocol to control delay jitter by operating at an
intermediate point.
SHARP-DJ provides good jitter performance at low cost
for all simulated levels of mobility. Figure 6(b) shows the
performance of SHARP-DJ for the same network topology,
as it tries to achieve a target delay jitter of 0.18 (normalized).
This graph shows that SHARP-DJ closely follows the
performance of the reactive component when the mobility is
low, but slowly increases the zone radius in order to satisfy
the goal at high mobility. SHARP-DJ performs dynamic
adaptation successfully to guarantee a user-defined target
of delay jitter for different network conditions.
All Nodes Destinations
The previous sections showed that SHARP is efficient for
networks in which the traffic is Zipf-like, that is, a few destinations
have several sources sending data to them. Yet0.72
0.54
multi-destination
F 0.0
F 0.2
F 0.4
F 0.6
F 0.8
F 1.0
0.72
0.54
AODV
SHARP-DJ
SPR
multi-destination
delay jitter
0.36
delay jitter
0.36
0.18
0.18
0
AODV 1 2 3 4 5 6 7 8 SPR
zone radius
Figure 6(a): Delay jitter vs. zone radius for the multidestination
topology, with 200 nodes. This graph illustrates
the trade-off in jitter entailed by the reactive and proactive
routing components as the mobility in the network varies.
0
0 0.2 0.4 0.6 0.8 1
mobility fraction
Figure 6(b): Delay jitter vs. mobility for the multidestination
topology, with 200 nodes and a target jitter of
0.18. This graph shows that SHARP-DJ adapts to network
conditions to achieve the target delay jitter efficiently.
10%
all nodes destination
800
all nodes destination
8%
AODV
SHARP-LR
SPR
600
loss rate (%)
5%
3%
packet overhead (1000 pkts)
400
200
SPR
SHARP-LR
AODV
0%
0 0.2 0.4 0.6 0.8 1
mobility fraction
Figure 7(a): Loss rate vs. mobility when all nodes in a
network of 80 nodes are destinations for a target loss rate of
5%. This graph shows that SHARP-LR adapts dynamically
to network conditions to achieve the target loss rate.
0
0 0.2 0.4 0.6 0.8 1
mobility fraction
Figure 7(b): Packet overhead vs. mobility when all nodes
in a network of 80 nodes are destinations for a target loss
rate of 5%. This graph shows that SHARP-LR achieves its
target without incurring excessive overhead.
SHARP could also be invoked in the case where the traffic
distribution is not Zipf-like. We studied the performance of
SHARP in cases where all nodes serve as destinations. We
simulated networks of 80 nodes with data traffic destined to
all nodes from randomly chosen sources at a rate of 2 packets
per second. The limited scalability of GloMoSim prevented
us from simulating larger networks with increased traffic.
Figure 7(a) shows the loss rates achieved by AODV, SPR,
and SHARP-LR for a target loss rate of 5%, when all nodes
are equally popular as route destinations. The reactive component
incurs low loss rates at low mobility, but greater than
5% loss rates at high mobility. The proactive routing component
achieves smaller than 5% loss rates at all levels of
mobility. However, Figure 7(b) shows that it incurs a very
high overhead. SHARP-LR balances this trade-off between
loss rate and overhead and meets the target loss rate at all
levels of mobility, incurring significantly lower overhead than
the purely proactive component. These simulations demonstrate
that SHARP can work well even when the traffic distribution
follows a non-Zipf pattern.
Summary of Results
In this section, we have demonstrated the performance of
three different adaptation schemes under varying network
conditions. Our simulations show that there is a fundamental
trade-off between reactive versus proactive dissemination
of routing info, and that no fixed strategy, reactive or proactive,
is suitable for a wide range of network conditions. This
trade-off mandates the need for protocols to dynamically
adapt based on network conditions. SHARP protocols dynamically
find a good balance between these two routing
strategies and outperform each one under varying network
conditions. Further, they enable multiple nodes in the network
to pursue different application-specific metrics at the
routing layer.
8. CONCLUSIONS
This paper presents SHARP, a hybrid routing protocol
that dynamically adapts to changing network characteristics
and traffic behavior. SHARP is driven by the fundamental
trade-off between proactive dissemination and reactive discovery
of routing information. Our quantitative measurements
demonstrate that no pure strategy is suited for all
network conditions, and that the optimal routing strategy
often lies somewhere in between the two extremes. SHARP
is a hybrid protocol that can automatically find the balance
point between these two strategies through an analytical
model for making an informed trade-off and dynamic
network measurements. It can act as a purely reactive pro-tocol in a quiescent network, or use purely proactive routing
for hosts to which routes are in wide demand. SHARP
uses efficient mechanisms for dynamically manipulating the
zone size and simultaneously performs fine-grained adaptation
with low overhead. Further, SHARP enables applications
with different demands to control the performance
of the routing layer. We described how SHARP could be
used to minimize packet overhead, to bound loss rate, and
to control delay jitter. Our evaluation demonstrates that
SHARP achieves performance that is better than each one
of its concomitant purely reactive and purely proactive protocols
across a wide range of network conditions. Overall,
there is a large spectrum of design points between purely
proactive and reactive protocols. SHARP enables applications
to explore this space in networks, whose patterns of
data traffic, node-degree, and mobility rates are subject to
dynamic change.
9. REFERENCES
[1] R. Boppana and S. Konduru. An Adaptive Distance
Vector Routing Algorithm for Mobile, Ad Hoc
Networks. In IEEE Infocom 2001, Mar 2001.
[2] J. Broch, D. Maltz, D. Johnson, Y. C. Hu, and
J Jetcheva. A Performance Comparison of Multi-Hop
Wireless Ad Hoc Network Routing Protocols. The
Fourth Annual ACM/IEEE International Conference
on Mobile Computing and Networking, Oct 1998.
[3] S. Corson and V. Park. Temporally-Ordered Routing
Algorithm (TORA) Version 1 Functional
Specification. IETF MANET Internet Draft, Jul 2001
(work in progress).
[4] M. Crovella and A. Bestavros. Self-Similarity in World
Wide Web Traffic: Evidence and Possible Causes. In
Proceedings of SIGMETRICS 1996, May 1996.
[5] Samir R. Das, Charles E. Perkins, Elizabeth M. Royer
and Mahesh K. Marina. Performance Comparison of
Two On-demand Routing Protocols for Ad hoc
Networks. in IEEE Personal Communications
Magazine special issue on Ad hoc Networking, Feb
2001, p. 16-28.
[6] Mario Gerla. Fisheye State Routing Protocol (FSR)
for Ad Hoc Networks. IETF MANET Internet Draft,
Jun 2002 (work in progress).
[7] Zygmunt J. Haas and Marc R. Pearlman. The
Performance of Query Control Schemes for the Zone
Routing Protocol. IEEE/ACM Transactions on
Networking, Aug 2001.
[8] G. D. Holland and N. H. Vaidya. Analysis of TCP
Performance over Mobile Ad Hoc Networks. In Fifth
Annual ACM/IEEE Conference on Mobile Computing
and Networking (MOBICOM), Aug 1999.
[9] P. Jacquet, P. Muhlethaler and A. Qayyam.
Optimized Link-State Routing Protocol. IETF
MANET Internet Draft, Mar 2002 (work in progress).
[10] M. Joa-Ng and I-Tai Lu. Peer-to-peer Zone-based
Two-level Link State Routing for Mobile Ad Hoc
Networks. IEEE Journal on Selected Areas in
Communications, 17(8):1415–1425, Aug 1999.
[11] D. Johnson, D. Maltz, Y Hu and J Jetcheva. The
Dynamic Source Routing Protocol for Mobile Ad Hoc
Networks. IETF MANET Internet Draft, Feb 2002.
[12] P. Mehta and S. Udani. VoIP: Sounding Good on the
Internet. IEEE Potentials Magazine, pp. 36-40,
Oct/Nov 2001.
[13] Art Mena and John Heidemann. An Empirical Study
of Real Audio Traffic. In IEEE INFOCOM, Israel,
Mar 2000.
[14] A. Nasipuri, R. Burleson, B. Hughes, and J. Roberts.
Performance of a Hybrid Routing Protocol for Mobile
Ad Hoc Networks. In IEEE International Conference
on Computer Communication and Networks
(ICCCN2001), Phoenix, AZ, Oct 2001.
[15] S. Y. Ni, Y. C. Tseng, Y. S. Chen and J. P. Sheu. The
Broadcast Storm Problem in a Mobile Ad Hoc
Network. In Proceedings of ACM/IEEE MobiCom,
Aug 1999.
[16] Mathis, Semke, Mahdawi and Ott. Macroscopic
Behavior of the TCP Congestion Avoidance
Algorithm. In Computer Communications Review,
27(3), 1997.
[17] Nikaein Navid, Labiod Houda, and Christian Bonnet.
DDR: Distributed Dynamic Routing Algorithm for
Mobile Ad hoc Networks. First Annual Workshop on
Mobile Ad Hoc Networking and Computing,2000,
Boston, 2000.
[18] Nikaein Navid, Wu Shiyi, and Christian Bonnet.
HARP: Hybrid Ad hoc Routing Protocol.
International Symposium on Telecommunications, IST
2001, 2001.
[19] M. Lewis, F. Templin, B. Bellur and R. Ogier.
Topology Broadcast based on Reverse-Path
Forwarding. IETF MANET Internet Draft, Nov 2002
(work in progress).
[20] V. D. Park and M. S. Corson. A Highly Adaptive
Distributed Routing Algorithm for Mobile Wireless
Networks. In IEEE INFOCOM, Japan, Apr 1997.
[21] Marc R. Pearlman and Zygmunt J. Haas. Determining
the Optimal Configuration of the ZRP. IEEE JSAC,
17(6), Aug 1999.
[22] C. Perkins and P. Bhagwat. Highly dynamic
destination sequenced distance vector routing for
mobile computers. In Proceedings of ACM
SIGCOMM, 24(4), Oct 1994.
[23] C. Perkins, E. M. Royer, and S. Das. Ad-Hoc On
Demand Distance Vector (AODV) Routing. IETF
MANET Internet Draft, Nov 2002 (work in progress).
[24] V. Ramasubramanian, Z. J. Haas, and E. G. Sirer.
SHARP: A Hybrid Adaptive Routing Protocol for
Mobile Ad Hoc Networks. Technical Report, Cornell
University, Computer Science Department, Apr 2003.
[25] Elizabeth M. Royer and Charles E. Perkins. Ad-hoc
On-Demand Distance Vector Routing. In 2nd IEEE
Workshop on Mobile Computing Systems and
Applications, pages 90–100, Feb 1999.
[26] P. Samar, M. R. Pearlman, and Z. J. Haas. Hybrid
Routing: The Pursuit of an Adaptable and Scalable
Routing Framework for Ad Hoc Networks. The
Handbook of Ad hoc Wireless Networks, CRC Press,
2002.
[27] Xiang Zeng, Rajive Bagrodia, and Mario Gerla.
GloMoSim: A Library for Parallel Simulation of
Large-scale Wireless Networks. In 12th workshop on
Parallel and distributed simulation, pages 154–161,
Banff, Canada, May 1998.