Scalable Information Dissemination for Pervasive
Systems: Implementation and Evaluation
Graham Williamson, Graeme Stevenson, Steve Neely, Lorcan Coyle, Paddy Nixon
Systems Research Group
School of Computer Science & Informatics
University College Dublin
Ireland
graham.williamson@ucd.ie
ABSTRACT
Pervasive computing systems require large amounts of infor-
mation to be available to devices in order to support context-
aware applications. Information must be routed from the
sensors that provide it to the applications that consume
it in a timely fashion. However, the potential size and ad
hoc nature of these environments makes the management of
communications a non-trivial task. One proposed solution
to this problem uses gossiping, a class of probabilistic rout-
ing protocol, to disseminate context information throughout
the environment. Gossiping algorithms require far less in
the way of guarantees about network structure, reliability,
and latency than alternative approaches, but are unproven
in real world scenarios. We describe the on-going develop-
ment of a framework for evaluating the performance of these
algorithms within the context of pervasive environments.
Categories and Subject Descriptors
C.2.4 [Computer-Communication Networks]: Distrib-
uted Systems
General Terms
Algorithms, Performance
1. INTRODUCTION
Pervasive computing systems require a large amount of in-
formation to be made available to devices in order to support
adaptive, context-aware, applications. Potentially consist-
ing of large numbers of sensors and applications, the com-
plexity of any inter-device communication must be managed
correctly to allow systems to scale without deterioration in
performance.
Building pervasive systems in a decentralised manner al-
lows them to grow piecemeal while retaining some extent
of co-ordinated behaviour through lateral relationships be-
tween components. This is preferable to the traditional ap-
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.
MPAC ’06 November 27-December 1, 2006 Melbourne, Australia
Copyright 2006 ACM 1-59593-421-9/06/11 ...$5.00.
proach where we rely on services and commands supplied
by global control points. However, as the number of de-
vices in any given environment increases, we face the prob-
lem of efficiently transporting information from the sensors
that provide it to the applications that consume it and use
it to adapt. Real-time interactions can only be supported
with timely responses to user and application commands.
The presence of devices with differing capabilities and con-
nectivity, and applications with varying data requirements
makes this an interesting challenge.
We may observe that data from sensors in these systems,
in addition to having a limited lifetime, are often frequently
repeated: e.g., a user’s location may be updated every 5
seconds. The implication of this, is that we do not need
100% reliability of message delivery. Our pervasive systems
architecture, Construct [2], makes use of this relaxed re-
liability constraint to use a non-deterministic mechanism
called “gossiping” — an epidemic style algorithm — to un-
derpin its communications. Gossiping requires far less in the
way of guarantees about network structure, reliability, and
latency than alternative approaches, whilst providing the
desirable properties of decentralisation, and robustness to
change, that are required for pervasive computing systems
of this nature.
However, simple gossiping algorithms do not necessarily
provide reliability and scalability guarantees. An algorithm
that has been incorrectly applied to a given pervasive sys-
tem can result in situations where an unacceptable number
of peers fail to receive a message, or message delivery is not
timely enough for a given application. Although gossiping
has potential as an approach to data distribution in perva-
sive systems, its applicability remains unproven, especially
in real world scenarios.
In order that we can characterise different gossiping al-
gorithms, we need to be able to quantify the improvements
and drawbacks that they introduce into a network. To that
end, we describe the on-going development of a framework
for evaluating the performance of these algorithms within
the context of pervasive environments. We describe how
the framework can be used to simulate variance in a set of
key variables, and measure the properties of the system that
may be affected as a result.
2. CONSTRUCT
We motivate our research into gossiping protocols by in-
troducing Construct, our middleware infrastructure for the
collection and dissemination of context information in smart

environments. Construct forms a real-world empirical test-
bed for the gossiping communication substrate.
Construct is a software platform for integrating noisy data
sources in a clean, dynamic, flexible and semantically well-
founded manner. Construct provides applications with a
uniform view of information regardless of how that infor-
mation is derived, and supports extensive inferencing and
sensor fusion within the platform [2, 5].
Construct has a fully decentralised architecture; devices
in a smart-space each run an instance. Each instance man-
ages the local data provided by sensors (physical or logi-
cal) connected to that device — a “local star” topology in
which dumb sensors are connected to more computation-
ally capable hubs which then exchange information between
themselves. Should a device fail, any wireless sensors may
re-connect to another visible device running the Construct
platform. All data are modelled using the Resource Descrip-
tion Framework (RDF), which provides a standardised way
in which to model contextual information and properties.
Construct stores and manipulates this data using the Jena
Framework [11].
Applications connect to the Query Service component of
Construct, and use its services to obtain required informa-
tion. Components with domain-specific knowledge may re-
quest and aggregate data from multiple sources in order
to contribute new, or refined information. All data stored
within the system are associated with metadata, which de-
scribe data lifespan and security restrictions.
Collectively, devices running the Construct software main-
tain a global model of the data within a smart-space. Ap-
plications then work with the local view of the data present
on the device which they are connected to. Maintaining the
global data model requires context information to be com-
municated between instances of Construct: this is accom-
plished by the randomised state-swapping algorithm, goss-
iping, described in detail in the following section.
3. GOSSIPING
This section provides an introduction to gossiping and de-
scribes a number of notable approaches using gossiping for
information dissemination in distributed systems. It should
be noted, however, that gossip-based technologies are not
limited to information dissemination and can, in fact, be
used for other purposes. Examples of such applications in-
clude data aggregation [9], load balancing [9], failure detec-
tion [15], and resource location [14].
The particular focus on information dissemination is mo-
tivated by our adoption of gossiping as the underlying com-
munication mechanism for the Construct platform. Here,
the fundamental problem gossiping addresses in the archi-
tecture of Construct is that of providing a reliable, scalable,
and fault-tolerant group communication system.
The properties of reliability, robustness, and scalability
are important design goals for Construct. In gossip-based
algorithms, these properties stem from the purely local in-
teractions between gossiping nodes, resulting in emergent
behaviour which realizes these qualities: each node in the
Construct network acts autonomously and a network built
using gossiping can cope easily with faults, reconfiguring it-
self if necessary.
3.1 Epidemic Algorithms and Gossiping
Recently, research has been conducted into using ran-
domised algorithms for communication. One class of such
algorithms is based on the spread of epidemics.
Much of the current research in this area stems from the
systems developed to maintain consistency in the Clearing-
house database servers of the Xerox Corporate Internet [4]
and includes, for example, pbcast [1], lpbcast [7], and pm-
cast [6].
The term epidemic reflects how the information is dissem-
inated through the network. Peers pass information to a
small number of others, randomly selected from their neigh-
bourhood, in much the same way that an infection spreads
from person to person. During each cycle of the epidemic an
infected host — one that has already received a particular
message — may come into contact with, and infect a given
number of, other hosts. These target hosts are chosen ran-
domly and may, or may not, already be infected. The newly
infected peers then, in turn, relay the ‘infection’ to further
peers.
Epidemics have long been studied outside the field of com-
puting and much of the mathematics predicting the be-
haviour of such processes is applicable [8].
The expression gossiping derives from another analogy to
the process by which information spreads through the net-
work. In a similar fashion to an epidemic, rumours may
be spread through a community: people gossip with their
neighbours and friends, who in turn gossip with people in
their proximity. In this way, the rumour is propagated
through the population.
3.2 Notable Gossip-based Techniques
Eugster et al. [8] identifies a number of challenges that
developers of epidemic protocols face:
Network Awareness: How can the physical characteris-
tics of the underlying network be taken into account
in order to reduce load and increase performance?
Membership Management: How is information regard-
ing peers discovered, updated and tracked?
Buffer Management: How can the storage, ageing, and
removal of messages be managed effectively?
Message Filtering: How can it be ensured that nodes re-
ceive messages that they are interested in while re-
ducing the probability that they must handle other
messages?
Directional Gossip [10] begins to address the problem of
incorporating network awareness into gossiping algorithms.
They aim to reduce the overhead and increase the reliability
of gossiping algorithms in wide-area networks. Directional
Gossip uses an estimation of the minimum link cut set (the
maximum number of link disjoint paths) between nodes to
identify important paths through the network. Messages are
always flooded across links to nodes identified with a low
link cut set and gossiped probabilistically with other peers.
The evaluation of this algorithm showed increased reliability
of message delivery, however, in more realistic network set-
tings the overhead performance was more akin to flooding
than gossiping. It should also be noted that static ‘gossip
servers’ are assigned to each LAN segment which constitute
a single point of failure and require configuration informa-
tion from network administrators to provide links between
gossip servers.

Bimodal Multicast [1] aims to build a reliable, scalable,
multicast protocol with the key properties of high and sta-
ble throughput. The algorithm they develop, termed pb-
cast, augments an initial unreliable multicast (such as IP
multicast) with a gossiping stage to provide what may be
viewed as a decentralised retransmission service. In this
way, the initial multicast provides a quick, rough coverage
of the group and the gossiping stage fills in the gaps where
messages were lost. The main focus is on the optimisations
made to the algorithm which are designed to improve per-
formance and smooth throughput, and on the evaluation of
the protocol. A cursory mention is made of network aware-
ness and membership management, commenting that it is
possible to view a wide-area network as a collection of inter-
connected local-area networks. Gossiping would then hap-
pen inside local-area networks as normal, with only certain
nodes interconnecting these.
Lightweight Probabilistic Broadcast [7] goes some way to
address the concerns of scalable membership management
and buffer management. The lpbcast protocol piggybacks
membership information on gossip messages to provide a
completely decentralised broadcast group architecture where
each node maintains a fixed size partial view of nodes in
the system. Interesting optimisations to the protocol are
the age-based message buffer purging and frequency-based
membership information removal heuristics.
Finally, the remaining challenge of the four identified ear-
lier remains very much an open issue. The provision of a
form of message filtering would move the algorithms from
pure broadcast towards a scalable multicast. Probabilis-
tic Multicast [6] defines the problem and moves towards a
scalable multicast implementation which takes into account
network awareness and the interests of each node in dissem-
inating messages. Autonomous Gossiping [3] aims to solve
a similar problem in the context of wireless mobile ad hoc
networks. Selective dissemination is based on profiling of
data items and host interests, with data items competing to
survive, migrate, and replicate between hosts.
To illustrate the possible implementation of algorithms
similar to those described, and to highlight some of the in-
teresting details of gossiping protocols, we will proceed with
a deeper examination of a simple gossiping protocol.
3.3 A Simple Gossiping Protocol
A skeleton gossip-based protocol is described with psue-
docode in Figure 1. Each node has two main responsibilities:
the first, to periodically wake up and initiate a gossip with a
number of nodes; the second, to listen for incoming gossips
and react to these. These responsibilities are illustrated in
Figure 1a and Figure 1b respectively.
Each peer participating in the network maintains a local
state which, at a minimum, consists of two tables: a list
of peers that it considers to be its neighbours and a list of
currently active messages.
Two important points may be implied from this descrip-
tion of local state. Firstly, we must have a way of discovering
and managing membership information (i.e. what peers are
our neighbours?); and secondly, we cannot infinitely buffer
all messages and therefore messages eventually should be
considered inactive and must be purged. Both of these areas
— membership management and buffer management — are
identified as challenges by Eugster et al. [8], and represent
points where intelligence could be applied to the gossiping
process.
The premise of gossiping is that, on each round of the
algorithm, the nodes involved in a gossip reconcile any dif-
ferences they have in their local states. Figure 1 represents
this notion by having nodes gossip summaries of the recent
messages that are present in their buffers (this could con-
sist simply of sequence numbers of messages the node knows
about). On the listening side, if the summary it receives dif-
fers from its own message history, the node can act to rec-
oncile the differences in two ways (which are not mutually
exclusive): either by requesting the messages it is missing
from the sender of the gossip, or by forwarding the messages
it has detected the sender is missing.
The remaining important points to note from the psue-
docode description are that gossiping occurs in unsynchro-
nised rounds — the duration of which may be tuned by
adjusting the parameter T ; the neighbours to gossip with
are chosen randomly from the list of peers a node is aware
of; the number of neighbours chosen on each round is termed
the fanout; and the notion of ageing messages is included in
order to measure how old an infection is and gauge when it
may be appropriate to remove it from the buffer.
Adjusting T and fanout have a number of effects. For
example, the load put on the network can be adjusted: in-
creasing T reduces the frequency of gossips and hence re-
duces the bandwidth used (however the corollary is that la-
tency will be increased). Also, the reliability of propagation
of messages can be affected: increasing fanout increases the
probability that nodes receive a message but also increases
the network load as more messages are sent.
The advantage of gossip-based algorithms, such as that
described heretofore, lies in the emergent behaviour: these
simple local interactions lead to a scalable, resilient network
which can react well to changes in the network structure.
4. GOSSIPING IN CONSTRUCT
The gossiping layer of Construct is a realisation of the sim-
ple gossiping protocol described in the previous section. The
implementation aims to be compact and lightweight, allow-
ing us to gain experience and make some early observations
of how the algorithm functions. However, at the same time,
we hope to make it flexible enough to allow the implementa-
tion of more complex gossiping algorithms. Important areas
where flexibility is catered for in the core algorithm include:
the choice of which node to gossip to on each round of the
algorithm; the choice of which message to request when a
received message summary differs from our own; and the
choice of which message is removed from the buffer when it
becomes full.
The gossiping subsystem itself consists of three compo-
nents:
1. the core gossiping protocol
2. the message buffer responsible for storing the messages
that we are currently gossiping about
3. the membership manager responsible for maintaining
the list of contacts that we can gossip with
In the initial version, the core gossiping protocol is state-
less. Gossiping is initiated periodically at each host by send-
ing their message buffer summary to a single, randomly se-
lected node from their contact list. When messages arrive,

(a) Initiating gossip process (b) Listening gossip process
Figure 1: Skeleton Gossiping Psuedocode
they are dealt with based on the type of message and then
discarded. The gossiping layer must handle three types of
messages: data messages, message buffer summaries and re-
quests for messages. Data messages are simply stored in
the message buffer for future gossiping then passed up the
protocol stack to the application. Message buffer summaries
contain a list of the message IDs present in the remote nodes
buffer. On comparison with our message buffer, a request
for a message which we do not have is generated. These
requests are fulfilled by sending a data message containing
the appropriate information.
We do not address the discovery of instances of Construct
to participate in gossiping at the moment. This function is
currently accomplished with the zero configuration network
protocol, Bonjour [12]. This provides a convenient solution,
allowing us to work on the core gossiping protocol, but limits
the scalability of the initial implementation, as we effectively
assume global knowledge of Construct instances.
Currently, there is no evaluation or measurement taken of
the performance of this implementation. To evaluate this
effectively is a difficult problem and, in working towards a
solution, we have developed a framework which describes
the parameters we wish to tune and the measurements we
will use to determine the effects of changing these parame-
ters. The framework we will use is described in the following
section.
5. AN EVALUATION FRAMEWORK FOR
GOSSIPING PROTOCOLS
Gossiping algorithms involve trade-offs between a number
of factors, for example: network overhead, reliability, or la-
tency. As novel gossiping techniques are designed, there is a
requirement to identify where improvements and drawbacks
have been introduced. To better understand the impact of
using new gossiping techniques we have designed a goss-
iping evaluation framework detailing parameters that may
affect performance and the measurements that we will take
to evaluate the impact.
For researchers it is a complex, even impractical, process
to perform real-world tests on their gossiping implementa-
tions. Experimentation of this sort will require the use of
large numbers of heterogeneous devices spread across the
globe to gain insight into the behaviours of their systems.
Initiatives such as Planet Lab [13] support global experimen-
tation however there is a considerable overhead in preparing
and launching such tests. To obtain the most out of a real-
world test session we can first test by simulation.
By employing a framework for evaluating gossiping sys-
tems we aim to gain a comprehensive understanding of how
algorithms will perform under varying conditions (arrange-
ment of devices, link latency, node failure etc.). Through
simulation we will have a basis to compare both existing
algorithms and our new algorithms. Through real-world ex-
perimentation we can validate our results by observing ac-
tual deployments with emergent behaviour in the wild. By
taking a systematic approach with the description of param-
eters and measurements, and evaluating at different points
in the test space using both simulation and real-world ex-
perimentation, we will gain a comprehensive understanding
of our algorithms.
5.1 Overview of the Framework
The gossiping evaluation framework takes a two-phase ap-
proach involving OMNet++ [16], a discrete event simulator
that can model computer networks and communication pro-
tocols, and real-world experimentation on Planet Lab. The
two parts of the experimentation method compliment and
provide validation against each other. The results from sim-
ulation and real-world experimentation should have com-
monalities providing a consensus on the performance of the
particular implementation.
We have identified the main set of variables that affect
the performance of the gossiping algorithm under inspection.
The impact of altering each is often seen to be combinatorial
and multiplicative. To address this, we have also identified a
set of measurements that we feel can be used to characterise
and compare the performance of gossiping algorithms.
5.2 Evaluation parameters
There are a number of parameters that can be manipu-
lated when deploying a network of gossiping nodes. Table 1
describes these.
Gossiping networks trade-off scalability with reliability
properties. As the number of nodes in a gossiping net-
work increases, the amount of data being transmitted will
increase. This adds redundancy to the network but impacts
on bandwidth utilisation. The optimal configuration of the
network depends on tuning each parameter with consider-
ation to the impact on other parameters plus the overall
affect on the network.
Node tables store lists of peers that may be gossiped with.
Smaller tables lead to a higher likelihood of node isolation
or disconnected islands. A fully connected network, where
node tables store the location of all other nodes, is infea-
sible on a global-scale. The memory footprint of a node
will be the sum of node table size with the buffer size. The
memory usage becomes highly critical in resource limited

Parameter Remarks
Number of nodes How many nodes are deployed in the network?
Node table size How many other nodes does each node know about? This may vary between nodes.
Buffer size How much data can each node hold?
Probability of nodes join-
ing or leaving
How volatile is the network configuration?
Message size How much data is gossiped in each cycle?
Message gossip frequency How often do gossiping cycles take place?
The ‘fanout’ How many nodes does a node gossip to per cycle?
Physical network How do properties such as bandwidth, link latency, link failure, and topology impact
performance?
Table 1: Tuneable parameters in a network of gossiping nodes
devices commonly found in ad hoc networks. As more data
is stored in buffers, increasing redundancy, the reliability
will increase.
Altering the probability of nodes joining or leaving the
network impacts the rate at which data is propagated. De-
pending on the network configuration this may increase or
decrease the rate at which data travels around the network
or may result in islands of connectivity being introduced into
the system.
With an increase in traffic levels, buffers will fill more
rapidly. In some cases this may result in messages having
to be dropped. Over communication may saturate the net-
work links, leading to blockages. Message gossip frequency
and fanout have a similar impact on the system. The com-
putational cost of a gossiping algorithm places a restriction
on the minimum value that the gossiping frequency can take.
The advantage of more communication is an increase in re-
dundancy as data is copied about the network.
In addition to these parameters, well understood proper-
ties resulting from the physical organisation of a network
will also impact the performance of any given algorithm.
Bandwidth, link latency and link failure are three examples
of such properties. OMNet++ has support for modelling
these network features, which we will use to test these algo-
rithms under various network conditions.
5.3 Measurements
Manipulation of the parameters defined above impacts on
the performance and behaviour of a network. To facilitate
characterisation of different gossiping algorithms we need
tangible figures for comparison. Table 2 describes these mea-
surements to be taken as an output of experimentation.
Latency and saturation form the baseline measurements
for our evaluation of gossiping systems. Latency, the aver-
age time taken for a unit of data to travel between randomly
selected points A and B, defines the simplest basis for com-
parison. From this measurement we extend to a measure-
ment of data saturation. In Eugster et al. [7] saturation is
described as the average amount of time taken for messages
within the system to reach 90% of the nodes in the network.
However, the selection of an appropriate fraction depends
on the requirements of an individual application, therefore
evaluation at a number of saturation points is desirable.
Robustness can be measured with respect to latency and
saturation measurements when subjected to network dis-
ruptions such as node or link failure, increased traffic levels,
increased latency on communications links, and increased
processing delay within nodes. Reliability, buffer and node
table size may also be key properties affecting this.
In simple gossiping systems it is common for duplicate
messages to arrive at a node. Any messages being transmit-
ted to a node which already has the data contained within
that message is wasted bandwidth. As in Eugster et al. [7],
this can be measured by a change in the ratio of received in-
formation to already known information within a gossiping
cycle over time.
Prevalence describes the proportion of nodes in a network
to which an item of data is known at a specific time. It pro-
vides a useful measure of the commonality of data within a
system. Incidence describes the number of new nodes that
learn of an item of data in a given gossiping cycle. Preva-
lence will be greatly affected by the lifespan of data, the
buffer size of nodes, and the rate that new data is gener-
ated. Incidence will be affected by both the selection of
data to be gossiped in any given cycle and the selection of
nodes to gossip with.
When investigating algorithms that are capable of using
the semantics of data to inform the gossiping process we
can provide additional measurements that relate to those
semantics. Topic coverage and topic distribution are two
such examples. Topic coverage measures the proportion of
data associated with a given topic that is known to a node in
relation to all data known about that topic. Topic distribu-
tion measures the prevalence of data in a network pertaining
to a given topic.
Local data redundancy is a measurement of the amount
of data stored by a node that can be considered redundant
when taking into account the data stored by neighbouring
nodes. A high redundancy figure implies a high tolerance
in the network for failure of that node. This figure may
be affected by many properties, including: node table size;
the rate at which new data is contributed at that node; the
rate at which data is gossiped; and the relative buffer size
of neighbouring nodes.
Finally, the scalability of the system is a combination of
changes in the other measurements as network size increases.
6. CONCLUSIONS
As pervasive computing systems grow and more sensors
are added to the system, we gain access to increasing vol-
umes of context data. A challenge arises in distributing
this data around the system to the points at which it is re-
quired by applications in an efficient, scalable, and timely
manner. Gossiping algorithms provide a promising means
of addressing this challenge but remain unproven in the
area. We have described the principles involved in goss-

Property Description
Latency A measure of how long it takes for datum X to go from A to B
Saturation The time taken for a message to reach a given fraction of the nodes
Robustness The percentage change in latency and saturation in the face of network disruptions (such
as node failure, increased traffic. . . )
Redundancy The ratio of information received by a node that is already known to it within a given
gossiping cycle
Prevalence The cumulative fraction of nodes that have received a message with respect to time
Incidence The number of nodes receiving datum X per round
Topic coverage The ratio of messages about a particular topic known to a node in relation to the total
number of messages in the network about that topic
Topic distribution As for prevalence, but for data pertaining to a specific topic
Local data redundancy The ratio of the total data known to a node that is replicated in neighbouring nodes
Scalability Defines the impact on the above measurements with increasing network size
Table 2: Measurements to be taken on gossiping algorithms
iping protocols for information dissemination and the initial
implementation of a basic gossip-style algorithm in our per-
vasive systems architecture, Construct. The complexity of
evaluating the performance of this algorithm has led us to
define a framework, which describes the parameters, mea-
surements, and approach utilising simulation and real-world
testing. We have presented an embryonic version of the eval-
uation framework; this will grow and evolve as we learn more
about gossiping algorithms and applications that are likely
to employ these techniques. Using this evaluation frame-
work, we aim to confirm the suitability of gossiping for the
dissemination of context information and allow us to gain
an understanding of the issues and trade-offs involved when
designing a gossiping protocol. Our next steps are to create
a simulation of the current implementation in Construct and
apply our evaluation framework to both the simulation and
concrete implementation.
7. ACKNOWLEDGMENTS
This material is based on works supported by Science
Foundation Ireland under Grant No. 04/RP1/I544.
8. REFERENCES
[1] K. P. Birman, M. Hayden, O. Ozkasap, Z. Xiao,
M. Budiu, and Y. Minsky. Bimodal multicast. ACM
Trans. Comput. Syst., 17(2):41–88, May 1999.
[2] L. Coyle, S. Neely, G. Rey, G. Stevenson, M. Sullivan,
S. Dobson, and P. Nixon. Sensor fusion-based
middleware for assisted living. In Proc. of 1st
International Conference On Smart homes & heath
Telematics (ICOST’2006) ”Smart Homes and
Beyond”, pages 281–288. IOS Press, 2006.
[3] A. Datta, S. Quarteroni, and K. Aberer. Autonomous
Gossiping: A self-organizing epidemic algorithm for
selective information dissemination in mobile ad-hoc
networks. In IC-SNW’04, International Conference on
Semantics of a Networked World, LNCS, pages
126–143, 2004.
[4] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson,
S. Shenker, H. Sturgis, D. Swinehart, and D. Terry.
Epidemic algorithms for replicated database
maintenance. SIGOPS Oper. Syst. Rev., 22(1):8–32,
January 1988.
[5] S. Dobson, L. Coyle, and P. Nixon. Hybridising events
and knowledge as a basis for building autonomic
systems. In Journal of Autonomic and Trusted
Computing. September 2006, 2006. To appear.
[6] P. T. Eugster and R. Guerraoui. Probabilistic
multicast. In Dependable Systems and Networks, pages
313–322, 2002.
[7] P. T. Eugster, R. Guerraoui, S. B. Handurukande,
P. Kouznetsov, and A. M. Kermarrec. Lightweight
probabilistic broadcast. ACM Trans. Comput. Syst.,
21(4):341–374, November 2003.
[8] P. T. Eugster, R. Guerraoui, A. M. Kermarrec, and
L. Massoulie. Epidemic information dissemination in
distributed systems. Computer, 37(5):60–67, 2004.
[9] M. Jelasity, A. Montresor, and O. Babaoglu.
Grassroots self-management: A modular approach. In
International Workshop on Self-* Properties in
Complex Information Systems, 2004.
[10] M. J. Lin and K. Marzullo. Directional gossip: Gossip
in a wide area network. In European Dependable
Computing Conference, pages 364–379, 1999.
[11] B. McBride. Jena: Implementing the RDF model and
syntax specification. In Proceedings of the 2nd
International Workshop on the Semantic Web, Hong
Kong, may 2001.
[12] The Bonjour Developer Connection Homepage.
http://developer.apple.com/networking/bonjour/.
[13] The PlanetLab Homepage.
http://www.planet-lab.org/.
[14] R. van Renesse, K. P. Birman, and W. Vogels.
Astrolabe: A robust and scalable technology for
distributed system monitoring, management, and data
mining. ACM Trans. Comput. Syst., 21(2):164–206,
May 2003.
[15] R. van Renesse, Y. Minsky, and M. Hayden. A
gossip-style failure detection service. Technical Report
TR98-1687, Cornell University, 1998.
[16] A. Varga. The OMNet++ discrete event simulation
system. In Proceedings of the European Simulation
Multiconference (ESM’2001), Prague, Czech Republic,
June 2001.