Real World Testing of Aggregation in Publish/Subscribe Systems
Michael Schiefer, Christoph Steup and J¨org Kaiser
Department of Distributed Systems, O.-v.-Guericke-Universitaet Magdeburg,
Universitaetsplatz 2, 39106 Magdeburg, Germany
Keywords:
Aggregation, Wireless Sensor Network, Publish/Subscribe, Network Topology, Ubiquitous Computing.
Abstract:
Wireless sensor networks collecting data to monitor real life processes gain increasing attention from the
scientific community. These systems promise ubiquitous computing in a digitalized world. However many
problems need to be solved to achieve this dream. One of these problems is the limited battery power of the
nodes and the limited bandwidth of the communication. Existing work tackle the problem of limited bandwidth
by merging communication packets within the network. Based on this approach we investigate the use of
application specific aggregation in publish/subscribe wireless sensor systems. We expect this approach to
overcome overload situations in the network and decrease packet loss due to bandwidth exhaustion. This paper
describes our architecture and evaluates the properties of our approach. We investigated specific topologies
theoretically and through real-life experiments. Our results quantify the achievable event count and loss rate
reduction.
1 INTRODUCTION
Sensor networks are an emerging field of technology
which provides easy and cheap surveillance of com-
plex and spacious systems. These networks consist
of cheap nodes with limited computational power and
energy storage. Therefore the collaboration of these
nodes is crucial to the final service.
This technology is applicable to many scenarios
ranging from forest fire detection to remote surveil-
lance of industrial environments. This paper consid-
ers a typical wireless sensor network scenario includ-
ing mobile and failing nodes. Therefore the resulting
network includes dynamic changes in the topology as
well as unreliable links between nodes.
The dynamic nature of such a network induces
challenges for the reliable dissemination of the data.
The constant changes of the topology complicates
the identification of nearby stations through names
or addresses. A solution to this problem is the pub-
lish/subscribe communication paradigm to decouple
the nodes from each other as described by (Eugster
et al., 2003).
Another problem is the limited processing capa-
bility and bandwidth of the hardware. This may cause
simple strategies like ”forward-to-sink and process”
to fail because of overload in the sink or during com-
munication. There are two possible solutions for this
problem: re-routing to nodes with free capacity or re-
duction of used bandwidth. Re-routing is a complex
problem due to mobility and energy constraints. It
also fails if only one path to the sink exists. The re-
duction of bandwidth can either be achieved by drop-
ping or combining individual packets. Dropping al-
ways induces a loss of information and needs rules to
describe which packets should be dropped (cf. (Chen
and Kotz, 2005)), whereas combining may conserve
the data but increases the dissemination delay.
A problem specific to event driven sensor systems
occurs whenever multiple sensors observe the same
real event. The sensors will create individual events
shortly after the observation. These sensor are gen-
erally close to each other, which creates a short-term
overload in the network due to the concurrent event
dissemination. Additionally necessary forwarding of
the created events may heighten the problem. This
can either be overcome by spreading the event publi-
cation in time, which the CSMA/CA algorithms will
do or by combining multiple events to prevent the in-
dividual event transmissions. However the spreading
done by CSMA/CA is error-prone since it depends on
statistical properties.
This paper tries to combine the publish/subscribe
communication paradigm with an application-specific
aggregation mechanism to lower the used bandwidth,
overcome the short-term overload through concurrent
event production and provide a load balancing mech-
anism within the wireless sensor network. Addition-
199
Schiefer M., Steup C. and Kaiser J..
Real World Testing of Aggregation in Publish/Subscribe Systems.
DOI: 10.5220/0004533401990206
In Proceedings of the 10th International Conference on Signal Processing and Multimedia Applications and 10th International Conference on Wireless
Information Networks and Systems (WINSYS-2013), pages 199-206
ISBN: 978-989-8565-74-7
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
ally the properties of our approach will be evaluated
on hypothetical topologies as well as real life experi-
ments.
2 RELATED WORK
Plenty of work exists on aggregation of data in net-
works. The first approach especially tailored to-
wards wireless sensor networks is ”Tiny AGgre-
gation” (TAG) (Madden et al., 2002) on TinyOS
which afterwards lead to TinyDB (Madden et al.,
2005)(Madden, 2003). This work views the sen-
sor network as a database, which executes SQL like
queries. A tree based routing is used to spread the
request ”from a powered, storage-rich basestation” as
root to every node. Afterwards the data flows back
from the leaves level by level. The parent nodes first
gather the data of all children and aggregatethem with
their own data before they transfer the result one tree
level up. Thus the nodes can sleep most of the time.
Additionally each node carries only one message per
request, which ”dramatically decrease[s] the amount
of communication required to compute an aggregate”
in most cases. This also provides an almost uniform
energy consumption of the nodes. TAG is limited by
its centralized routing tree, which makes it impossible
to use two queries at the same time and always needs
an external base station as the source of the request.
Another approach targets the known meta data of
publish/subscribe middlewares. Application-Specific
Integrated Aggregation for Publish/Subscribe Sys-
tems (ASIA) by Margera et. al(Frischbier et al., 2012)
views the network as a topic-based publish/subscribe
system instead of a distributed database. The publish-
ers and subscribers are linked with brokers thus every
message has to be transmitted over at least one broker.
This generates meta data on the network itself or the
sensor data. ASIA provides an aspect oriented inter-
face to determine and aggregate those meta data in a
application-specific way. However it is not intended
to aggregate the data itself. The current design and
implementation of ASIA is unsuitable for resource
constraint systems.
To use aggregation in complex systems
application-specific mechanisms are needed since
primitive merging functions are not enough. Systems
like ASIA allow the users to create customized merg-
ing functions only for meta data. On the other hand
TAG and TinyDB allow user-specified application-
specific aggregation functions, but are incompatible
with the publish/subscribe communication paradigm
due to their rigid time and routing scheme. Therefore
this paper explores semantic aggregation in pub-
lish/subscribe systems and evaluates it theoretically
and practically.
3 ARCHITECTURE
Our approach is based on Contiki(the Contiki project,
2012) as the underlying operating system and FA-
MOUSO(Zug et al., 2010)(Schulze, 2009) as pub-
lish/subscribe middleware.
Contiki provides hardware abstractions and thread
implementations tailored towards deeply embedded
systems like sensor nodes. Whereas the FAMOUSO
middleware provides a topic-based publish/subscribe
communication independent of the underlying OS
through template meta-programming.
Hardware
Contiki OS
Routing
FAMOUSO
Application Layer
RF230 Driver
Multi. Network Layers
Rime
Network Layer
Abstract Network Layer
Event Layer
Publish-Subscribe
Interface
Aggregation
Interface
Figure 1: Architecture of the topic-based aggregation using
the FAMOUSO middleware and CONTIKI.
As depicted in Figure 1 the middleware uses the
Rime-Stack by (Dunkels, 2007) of Contiki to access
the underlying wireless network. The Rime-Stack en-
capsulate the wireless hardware as well as the MAC-
Layer of the IEEE 802.15.4 Standard(IEEE, 2011). It
also provides a node id, which is used by the network
layer of the middleware to translate publish/subscribe
topics to network addresses in run-time.
Since Rime does not provide a routing mechanism
FAMOUSO integrates its own, which is based on Di-
rected Diffusion by (Intanagonwiwat et al., 2000).
The subscribers periodically send their subscription
through the network by flooding it. Every node saves
the first neighbour sending this message, whichever
is closer to the subscriber. Thus every node knows
where received events need to be forwarded to. The
interval of renewal of subscriptions is chosen depend-
ing on the mobility of the nodes. This provides an ap-
propriate routing mechanism for content-based com-
munication, which fits to the communication model
of the middleware. Through periodic renewals of sub-
scriptions limited mobility can be tolerated.
WINSYS2013-InternationalConferenceonWirelessInformationNetworksandSystems
200
4 TOPIC BASED AGGREGATION
There are mainly two ways to achieve a semantic,
application-specific aggregation in publish/subscribe
systems.
The first one exploits the existing pub-
lish/subscribe API. One node subscribes to a
topic of specific data, aggregates the data and pub-
lishes the results in another topic. All interested
nodes subscribe for the new topic containing the
aggregated data.
The second approach uses an application specified
aggregation function and links it to the routing via
a defined interface. If an aggregatable event arrives
the routing forwards it up to the specified application
function so that it may be merged with other events.
Thus the events are aggregated on the fly on their way
from publisher to subscriber.
The first approach has the drawback that the
application-layer of the publish/subscribe system has
no information on the topology of the network.
Therefore explicitly placed aggregation nodes are
needed. In case of a static network the optimal posi-
tions may be pre calculated. For dynamic networks
however it is needed to reposition these nodes dy-
namically. The topology data for this repositioning
needs to be deduced through additional mechanisms
like service discovery. The second approach circum-
vent this problem and is used in our approach.
We force the application designer to convert the
event data in an aggregation aware format before pub-
lishing it. This enables the subscriber to evaluate the
received aggregated events before delivering them to
the application. This induces no additional overhead
to the designer since publishingand subscribing needs
application-specific data anyway. Therefore we can
assume that FAMOUSO either transfers data in an ag-
gregation aware format or the data itself is marked as
not aggregatable.
The application registers its specific merge func-
tion with the middleware as a channel attribute. The
middleware passes this merge function down to the
routing layer. Afterwards every received, aggregat-
able packet of the appropriate topic is passed to the
merge function. To link the application event in-
terface and the routing layer packet interface every
packet is converted to an event. To support storing
and merging of events each merge function owns an
individual packet buffer. To avoid long hold times ev-
ery data in the buffer is send after a specified period of
time, which is changeable by the merging function. It
is although possible to hold up events forever by peri-
odically changing the time-out.
The resulting architecture can even handle par-
tially deployed merge function. This might occur if
nodes cannot aggregate data because of limited mem-
ory or computing time. This paves the way for a
dynamic distribution of aggregation nodes within the
network based on individual capabilities and the cur-
rent topology of the network.
On the other hand the advantages and disadvan-
tages heavily depend on the used routing function.
An aggregation of data before the subscriber cannot
be guaranteed, since parallel data flows like depicted
in dark blue in Figure 2 are not aggregatable. A more
aggregation aware routing tries to collapse the differ-
ent event paths sooner and may be able to increase the
benefit of the aggregation.
P
P
S
Figure 2: Example network with parallel flow (dark arrows)
and an aggregation optimized flow (lighter arrows).
The routing tries to delay the cloning of events as
long as possible to minimize the amount of packets to
be transmitted to deliver a published event to the sub-
scribers. Therefore a route must be found that enables
early aggregation as well as late cloning.
Applications need to be aware that they may re-
ceive individual events multiple times as parts of ag-
gregated events. We therefore force applications to
handle such situation explicitly.
The publish/subscribe middlewarecreates individ-
ual routes for the events of each topic, as visible in
Figure 3. Here two different topics exist with only
one subscriber for both. At node R
5
events for chan-
nel one are routed to R
3
to enable aggregation. Events
of topic two are routed through R
6
to be aggregated
with events from P
2,2
. Hence packets arriving at the
same node with the same destination are routed differ-
ently depending on their content to enable optimal ag-
gregation. If both channels would be routed through
R
3
this would result in more sent packets in a global
view. This shows that cost and benefit of aggregation
and routing cannot be viewed independently. Useful
estimations are only possible for specified scenarios
with known routing mechanisms and topologies.
So why is the aggregation neglecting routing al-
gorithm of FAMOUSO used? Most publish/subscribe
systems do not incorporatean aggregation aware rout-
ing and thus using aggregation would lead to similar
results. Furthermore by designing the network topol-
ogy it is possible to create worst case but also best
RealWorldTestingofAggregationinPublish/SubscribeSystems
201
S
P
2,1
P
2,2
P
1,1
P
1,2
R
2
R
4
R
6
R
1
R
3
R
5
Figure 3: Example network with different routes based on
the topics of the data. P
i,k
is the k
th
publisher for topic i.
R
j
are forwarding relay nodes. The topic data is displayed
through the lighter and darker arrows respectively.
case scenarios. These results can be validated with
practical experiments.
5 EVALUATION
This section first analyse the possible cost and bene-
fits of an aggregation in a theoretical way. Afterwards
the real world test cases are described.
5.1 Topological Analysis
Unless otherwise stated we consider the routing as
well as the links between nodes to be optimal. The
goal of the analysis is an estimation of bandwidth re-
duction as well as induced latency for some general
topologies.
We found no typical topology, which approxi-
mates WSN well enough. Therefore we will discuss
some topologies with special properties and try to rea-
son their relevance towards real networks.
5.1.1 Worst Case
P
1
R
1
P
2
S
R
2
P
3
Figure 4: A topology not benefiting from aggregation. The
only extension conserving the worst case property is shown
in dashed grey.
To achieve worst case packet counts disjoint rout-
ing paths are needed. Worst case delays can only be
observed if no parallel communication to multiple ag-
gregating nodes are possible.
The black graph in Figure 4 showssuch a topology
without any possible time and event reduction. This
topology enables aggregation only at node R
1
, which
induces an additional packet from P
2
to R
1
. However
the gain is only one packet less from P
2
to S. There-
fore no reduction of packet count is possible. Addi-
tionally P
1
and P
2
may not transmit data to R at the
same time, which lengthens the delay of the events.
This worst case scenario is extensible, Figure 4 in-
cluding the grey parts. However the worst case prop-
erty only holds if each publisher has its own disjoint
routing path to the subscriber. In the end the topology
may only be extended through single publisher chains
connected directly to S.
5.1.2 Best Case
S
P
1
P
2
P
3
P
4
P
5
Figure 5: The best topology for aggregation.
The best case for aggregation is a chain of nodes
consisting of a subscriber at one end and publishers
(see Figure 5). Under the assumption that all n pub-
lishers publish at nearly the same time a minimal de-
lay is to be expected. The rightmost publisher P
5
transmits its event to its neighbour. The neighbour in-
stantly aggregates it with its own event and forwards
it to the next one. In the end the complete aggregated
event arrives at the subscriber. The packet count in
this example is reduced from 15 to five. In general
aggregation could decrease the number of transferred
packets from (n·(n+ 1)/2) to n. This is the optimum
concerning delay and packet count.
If on the other hand the leftmost publisher P
1
starts
to transmit followed by P
2
and so on, only one event
arrives at the subscriber after five sent packets. Be-
cause of the time-out the left publisher P
1
sends the
saved event of P
2
to the subscriber etc. Depending
on the time of publication and the chosen time-outs it
is possible for every event to arrive at the subscriber
individually. Thus even this best case topology may
provide no benefits and introduce additional delays.
This is a general problem. For every aggregationnode
in every topology such a bad case could be generated.
It is therefore necessary to chose waiting time-outs
depending on application’s constraints.
5.1.3 Simple Aggregation
The simplest non linear topology enabling aggrega-
tion consists of multiple publishers and subscribers
connected to one relay node (cf. Figure 6). If pub-
lishers transmit their events shortly after each other
the relay node may send the aggregated event to the
subscriber instantly. Thus the delay is small and the
event count is minimized.
WINSYS2013-InternationalConferenceonWirelessInformationNetworksandSystems
202
R
P
P
S
Figure 6: A topology with the possibility of aggregation.
On the other hand if the transmissions of the pub-
lishers are widely spread in time the relay node will
withhold events until timed-out. As described in Sec-
tion 5.1.2 this is a general limitation. The simplicity
as well as the extensibility of this topology are indi-
cations for its occurrence in real networks
5.1.4 Conclusion
As seen worst case scenarios are difficult to build
and unlikely in reality. Topologies with expectable
good aggregation results, e.g. chain like structures in
mine drifts(Harms, 2011), are more likely. Regular
networks will be composed of positive and negative
structures like Figures 4 to 5. In summary benefits
are expected with respect to routing parameters and
timing configuration.
5.2 Test Cases
To validate the theoretical results of Section 5.1 real
life experiments were conducted aiming at bandwidth
saving.
5.2.1 Configuration
As physical motes the RCB128RFA1 and the
deRFmega128-22A platforms by dresden elektronik
ingenieurtechnik gmbh(Pietschmann, 2013) are used.
They are quite similar in hardware with the main dif-
ference being the antenna. As software our archi-
tecture as described in Section 3 is used. Contiki is
configured to use ”nullmac” as MAC-Protocol result-
ing in no retransmissions and ”sicslowmac” as ”Radio
Duty Cycling”-Protocol underneath it. Furthermore
the maximum tx power is reduced and the rx treshold
is increased to achieve a smaller and more control-
lable test area.
Each test contains one additional control node,
which is not represented in the figures. No packet is
ever routed through this node. It is used to initialize
and verify the topology at the beginning and controls
the tests.
The topology is configured statically in the EEP-
ROM of the node. Even though nodes may receive
packets via links not part of the topology, these will
be dropped before being processed. To achieve repro-
ducible test results a verification of the routing paths
between publishers and subscriber is done in the be-
ginning of each test.
5.2.2 Aggregation Function
The used aggregation function is the mean of 16 bit
values. To enable identification of individual events
within the aggregatedevents a list of individual events
is attached to the resulting value. The forwarding of
stored events( aggregated or individual ones) happens
in three situations.
The first one is a time-out. After five seconds the
stored event is forwarded to the next node. This value
is a trade-off between delay of event dissemination
and time to aggregate incoming events.
The second situation consists of a maximum num-
ber of individual events incorporated within one ag-
gregated event. The maximum number is inferred
from the publishers active within the network.
The third situation occurs if an event is received
from a publisher already contained in the publisher
list of the stored aggregated event. This results in for-
warding the stored event and using the newly received
event as the new aggregated event. This mainly hap-
pens in high throughput tests when a packet is lost
and the original publisher of the lost event produces
the next data.
5.2.3 Bandwidth Tests
Our test considers only the best case topology for
bandwidth reduction to quantify the possible benefit
of using aggregation. This enables us to estimate the
possible benefit for different types of real life topolo-
gies. Thus the best case topology, see Figure 5, is
used.
We considers four different configurations: high
and low network throughput with a size of five as well
as ten publishers. To quantify the benefit of the aggre-
gation each configuration is tested with and without
aggregation. Network throughput will be emulated
by the transmission of pseudo events, which contain
randomly generated sensor data. This generated sen-
sor data will be aggregated by the application specific
mechanism described in Section 5.2.2 attached to the
used publish/subscribe channel. The high throughput
scenario publishes repeatable randomly two to four
pseudo events per publisher per second, whereas the
low throughput test publishes only one pseudo event
per publisher per minute. The nodes are configured
to publish at nearly the same time. The experiments
with high network throughput will run 15 minutes and
the low throughput ones 30 minutes.
RealWorldTestingofAggregationinPublish/SubscribeSystems
203
The bandwidth test is used to measure transmitted
as well as received packets for individual nodes and
published and received events for the whole test. Each
node measures the data on its routing layer. After the
test ends the statistics are saved in the EEPROMof the
node and later transmitted to pc via a serial data con-
nection. Thus the data gathering is guaranteed to not
interfere with the measurements. To prevent result-
biasing additional packets there is no retransmission
mechanism activated in the network.
To evaluate the problem of concurrent event pro-
duction we additionally conducted low throughput
tests with an id-based delay before the publication in
each node. The results should approximate a network
not observing the same phenomenonand therefore not
publishing at the same time.
6 RESULTS
The following sections will discuss the results of our
real-life experiments as described in section 5.2. We
will use abbreviations like H10AD, which stands for
a high throughput test (H) of size ten (10) with aggre-
gation (A) and id-based delay (D). Another example
would be L5, which represents a low throughput test
of size 5 without aggregation and without delay.
6.1 Packet Count
The Figure 7a depicts the bandwidth results of the
H10A experiment showing an increasing amount of
packets related to the distance of the publishing node
P
i
to the subscribing node S. This is caused by the
routing based retransmission of each packet of the
predecessor of each node additionally to its own data.
Consequently a linear growth is to be expected. How-
ever the diagram shows a non-linear growth from pub-
lisher P
10
to P
1
. This is caused by the high loss rate,
which will be discuss in Section 6.4.
The nodes of the H10A experiment, as visible in
Figure 7b, transmit a nearly constant amount of pack-
ets only. This is caused by the combination of the
received packets to a single event. Node S transmits
some packets to establish the channel and initialize
the routing. Publisher P
10
only receives these prop-
agating through the network. Therefore both nodes
have small values for transmitted and received pack-
ets respectively.
Figure 7c shows the bandwidth results of the
L10AD experiment. In this experiment the publish-
ers additionally had an ID dependent delay before
publishing. Here the expected linear growth of pack-
ets based on the distance to the subscriber is visible.
Compared to the basic test, this is due to the additional
delay between publications. The delay is signifi-
cantly larger than the initial delay of the CSMA/CA
of 802.15.4. This spreads the publications in time and
lessens significantly the collision rate of the packets
within the network. This shows the problem of short-
term overload on event generation of multiple nodes.
6.2 Aggregation Efficiency
Aggregation efficiency can be measured through the
amount of individual events contained in aggregated
events. In the experiment H10A 2372 aggregated
events contained all possible ten individual events.
Only 49 aggregated events contained less then ten
events. Therefore during the experiment in 95% of
all cases an optimal aggregation was possible. This is
to be expected since the publishers transmitted shortly
after each other.
Experiment L10AD created quite different results,
as visible in Figure 7d. The x axis shows the number
of individual events the received aggregated event is
build of. The y axis corresponds to the occurrence
of an aggregated event containing a specific number
of individual events. This diagram shows that 85,5%
of all events were part of aggregated events, but none
aggregated event contained the maximum number of
ten individual events. This is caused by the additional
delay destroying the alignment of the publishers. This
disabled the aggregation in some nodes, due to time-
outs.
6.3 Total Event Counts
The Figures 7e and 7f illustrates the total number of
transmitted events within the network.
The total number is the sum of all events gener-
ated by publishers as well as all the forwarded events
while routing. The numbers clearly show the correla-
tion of the saving with the chain size. This supports
the theoretical evaluation in Section 5.1. The H10 and
H5 experiments showed an event reduction of 50,4%
and 32, 1%.The L10 and L5 experiments showed sim-
ilar values with 45,3% and 34,1%. Therefore it can
be assumed that the event reduction does not depend
on the throughput of the network. This fits to the the-
oretical evaluation.
6.4 Loss Rate
The loss rate of the application is calculated as the
fraction of events received by the subscriber of all
published events. Aggregated events are multiplied
by the number of individual events they contain. The
WINSYS2013-InternationalConferenceonWirelessInformationNetworksandSystems
204
P
10
P
9
P
8
P
7
P
6
P
5
P
4
P
3
P
2
P
1
S
0
2,000
4,000
6,000
8,000
Node
#Packets
(a) Packet count of experiment H10.
P
10
P
9
P
8
P
7
P
6
P
5
P
4
P
3
P
2
P
1
S
0
2,000
4,000
6,000
8,000
Node
#Packets
received packets
transmitted packets
(b) Packet count of experiment H10A.
P
10
P
9
P
8
P
7
P
6
P
5
P
4
P
3
P
2
P
1
S
0
100
200
300
Node
#Packets
(c) Packet count of experiment L10D.
2 4
6
8 10
0
5
10
15
20
17
13
8
4
7
1
8
7
6
0
#aggregated events in received event
#events received by S
(d) Histogram of number of aggregation events received by S
containing x events in experiment L10AD.
H10A H10 H5A H5
0
10,000
20,000
30,000
40,000
50,000
24,282
46,242
12,130
17,867
#Events
(e) Total Event Count of high throughput test cases.
L10A L10 L5A L5 L10AD L10D
0
400
800
1,200
1,600
2,000
155
326
251
642
1,629
591
#Events
(f) Total event count of low throughput test cases.
H10A H10 H5A H5 L10A L10 L5A L5 L10ADL10D
0
20
40
60
80
1.73
2
1.33
54.67
69.9
73.33
0.34
1.11
64.72
1.36
percentage
(g) Event loss rate of all experiments.
H10A H10 H5A H5 L10A L10 L5A L5 L10ADL10D
0
20
40
60
80
0.6
1.08
1.05
29.18
36.62
31.56
0.06
0.4
43.79
0.31
percentage
(h) Packet loss rate of all experiments.
Figure 7: Results of different test cases as described in Section 5.2.
RealWorldTestingofAggregationinPublish/SubscribeSystems
205
loss rates for events and packets can be seen in Fig-
ures 7g and 7h respectively.
Tests with aggregation showed a quite low loss
rate compared to tests without. A cause may be the
nearly synchronous event production of the publish-
ers. This causes a temporal overload in the network,
which increases packet loss. The aggregation over-
comes this problem by combining multiple packets to
a single event. Therefore the short-term overload of
the network is circumvented and the additional colli-
sions are prevented.
This assumption is supported by the loss rate of
experiments L10D and L10AD. In this case the loss
rate of aggregation is higher, even though both ex-
periments showed a very low loss rate for packets as
well as events. The negative benefit might be caused
by the additionally delays of the forwarding, which
shifts forwarded events closer to the next published
event increasing collision probability.
7 CONCLUSIONS
In this paper we presented an extended pub-
lish/subscribe architecture to support application
specific aggregation. We described our architec-
ture based on the publish/subscribe middleware FA-
MOUSO and the embedded operating system Contiki
as well as the topic based aggregation.
To estimate the benefits of general aggregation
mechanisms in publish/subscribe systems we evalu-
ated different possible network topologies theoreti-
cally. Additionally we have done experiments on best
case topologies to evaluate the possible benefit. This
evaluation showed that in chain like topologies an
event count reduction of 32% for five nodes and 50%
for ten nodes is possible. This supports the theoret-
ical evaluation, which promises an increasing bene-
fit depending on the size of the chain. However the
theoretical benefits were higher because they omitted
loss rate. Additionally application-specific aggrega-
tion emerged as a solution for the temporal overload
of the network during concurrent event production of
close nodes. Finally we reasoned why aggregation
beneficial topologies will be increasingly likely de-
pending on the size of the WSN.
Following this work we want to evaluate the influ-
ence of an aggregation aware routing, which promises
additional benefits in non-optimal topologies. Fur-
thermore we want to extend the evaluation to other
metrics like energy and resource consumption.
REFERENCES
Chen, G. and Kotz, D. (2005). Policy-driven data dis-
semination for context-aware applications. In In Pro-
ceedings of the Third IEEE International Conference
on Pervasive Computing and Communications, page
283–289.
Dunkels, A. (2007). Poster abstract: Rime — a lightweight
layered communication stack for sensor networks.
Eugster, P. T., Felber, P. A., Guerraoui, R., and Kermarrec,
A. (2003). The many faces of publish/subscribe. ACM
Comput. Surv., 35(2):114–131.
Frischbier, S., Margara, A., Freudenreich, T., Eugster, P.,
Eyers, D., and Pietzuch, P. (2012). Asia: application-
specific integrated aggregation for publish/subscribe
middleware. In Proceedings of the Posters and Demo
Track, Middleware ’12, pages 6:1–6:2, New York, NY,
USA. ACM.
Harms, H. (2011). Sensornetzwerke im Bergwerk als
Untertage-Ortungssystem. Diploma thesis, Otto-von-
Guericke University Magdeburg.
IEEE (2011). IEEE standard for local and metropolitan area
networks–Part 15.4: Low-rate wireless personal area
networks (LR-WPANs). Technical Report 802.15.4-
2011, IEEE Computer Society.
Intanagonwiwat, C., Govindan, R., and Estrin, D. (2000).
Directed diffusion: A scalable and robust communi-
cation paradigm for sensor networks. In MOBICOM,
pages 56–67. ACM.
Madden, S. (2003). Tinydb: A declarative
database for sensor networks. website
http://telegraph.cs.berkeley.edu/tinydb/index.htm,
date: 2012-07-25.
Madden, S., Franklin, M. J., Hellerstein, J. M., and Hong,
W. (2002). Tag: a tiny aggregation service for
ad-hoc sensor networks. SIGOPS Oper. Syst. Rev.,
36(SI):131–146.
Madden, S. R., Franklin, M. J., Hellerstein, J. M., and
Hong, W. (2005). Tinydb: an acquisitional query
processing system for sensor networks. ACM Trans.
Database Syst., 30(1):122–173.
Pietschmann, L. (2013). dresden elektronik webseite.
http://www.dresden-elektronik.de.
Schulze, M. (2009). Famouso – family of adaptive mid-
dleware for autonomous sentient objects. website
http://famouso.sourceforge.net/, date: 17.07.2012.
the Contiki project (2012). Contiki: The open source op-
erating system for the internet of things. website
http://www.contiki-os.org/, date: 2012-07-17.
Zug, S., Schulze, M., Dietrich, A., and Kaiser, J. (2010).
Programming abstractions and middleware for build-
ing control systems as networks of smart sensors and
actuators. In Proceedings of Emerging Technologies
in Factory Automation (ETFA ’10), Bilbao, Spain.
WINSYS2013-InternationalConferenceonWirelessInformationNetworksandSystems
206
