icst transactions on ubiquitous environments research article

ICST Transactions on Ubiquitous Environments Research Article

Towards energy-aware semantic publish/subscribefor wireless embedded systemsDavy Preuveneers1,∗, Yolande Berbers1

1IBBT-DistriNet, Department of Computer Science, KU LeuvenCelestijnenlaan 200A, B-3001 Leuven, Belgium

Abstract

In this article, we present our lightweight semantic publish/subscribe system µC-SemPS that is targetedtowards wireless battery-powered embedded systems in ubiquitous computing environments. The keychallenge that we address is the minimization of the overall energy consumption, which involvescomputational aspects for semantic subscription matching, and communication aspects for the delivery ofthe publications as events in a wireless network. Our pub/sub system relies on a compact mathematicalrepresentation of the semantic subscriptions and publications, as well as fast semantic matching and routingalgorithms. From an energy efficiency point of view, it favors computation over the more expensive wirelesscommunication for the semantic matching and routing of events. When compared to more conventionalpub/sub routing implementations, experimental results from network and energy simulations with MiXiM,an OMNeT++ modeling framework extension for wireless and mobile networks, show that our approach forrouting semantic events significantly reduces energy consumption.

Received on 18 August 2011; accepted on 2 July 2012; published on 14 November 2012

Keywords: energy-awareness, publish/subscribe, semantic routing, wireless embedded systems

Copyright © 2012 Preuveneers and Berbers, licensed to ICST. This is an open access article distributed under the terms ofthe Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/), which permits unlimiteduse, distribution and reproduction in any medium so long as the original work is properly cited.

doi:10.4108/trans.ubienv.2012.10-12.e1

1. Introduction

The publish/subscribe (pub/sub) communicationparadigm [5] provides a many-to-many anonymousevent communication model that is decoupled intime and space, and as such an ideal distributednotification platform for ubiquitous computingsystems. Traditionally, an event is a collection ofattributes as key-value pairs, structured according toan a-priori known event schema, and communicated bymeans of push/pull interactions between possiblymobile publishers and subscribers. An eventnotification service acts as the mediator to notifysubscribers about published events that match theconstraints in at least one of their subscriptions. Forscalability reasons, the event notification service isoften implemented as an overlay network of eventbrokers [9] that match and route events/subscriptionsto other brokers and clients.

∗Corresponding author. Email: [email protected]

The major challenge that we address in this articleis the energy-efficiency of semantic pub/sub systemsfor wireless and battery-powered embedded systems.This class of pub/sub systems [2, 3, 11, 12] isgaining importance given the fact that (1) emergingcomputing paradigms like Machine-to-Machine (M2M)and the Internet of Things (IoT) will generate ever-increasing amounts of data, and that (2) participantsin loosely-coupled heterogeneous systems will use avariety of terminologies. The fact that all events mustbe unambiguously interpreted by the publishers andsubscribers is the main motivation why semantics mustbecome a first class entity of the pub/sub system.

In this article, we investigate how a pub/sub sys-tem can be optimized for battery-powered wirelesssensor nodes and embedded microcontrollers (µC), bymatching and routing semantic events in such a waythat reduces computation and communication energyusage. Addressing both challenges at the same timeis not straightforward. As previous research [4] hasshown, communication can be much more expensive

EAI European Alliancefor Innovation 1

ICST Transactions on Ubiquitous EnvironmentsOctober-December 2012 | Volume 12 | Issue 10-12 | e1

http://creativecommons.org/licenses/by/3.0/

mailto:<[email protected]>

D. Preuveneers, Y. Berbers

ab

b

c

a

a

b

a

c

b

x

Semanticoverlay network

Wireless sensor network

Inefficient transmission

Efficient transmission

EEE

E'

E

E' E'

E'

E'

EEamp

~ d2

1"

d

Figure 1. Mapping of a semantic overlay structure onto a peer-to-peer network

than computation, and therefore power-efficient net-work protocols [6] are often part of the big picture as faras minimizing energy consumption in wireless sensornetworks is concerned. However, merely implementinga semantic pub/sub system as an overlay structure ontop of a wireless peer-to-peer network and relying on anpower efficient communication protocol − hence treat-ing both problems as separate concerns − can still beinefficient from an energy perspective. A semantic over-lay network typically imposes a network structure withrouting paths from event producers to their respectivesubscribers where the order of the subscribers is definedby the specificity of their subscriptions. When routingevents from nodes with generic subscriptions to nodeswith more specific subscriptions (i.e. imposing a subsetof matching events), the forwarding of events stops atthe node for which the event no longer matches with itssubscriptions. The semantic overlay network provides ameans to route events while avoiding passing throughnodes without any matching subscriptions. However,strictly following the routing path imposed by the over-lay network may not be efficient from an energy pointof view depending on the distance between the nodesinvolved. This is illustrated in Figure 1 and explainedfurther below.

A direct unicast between two nodes may theoreticallybe more expensive than sending over intermediatenodes with minimum transmission energy (MTE) asrouting criteria, because the energy consumption of thewireless signal amplifier is quadratically proportionalto the distance. On the left of Figure 1, a messageis transmitted with direct unicast from a to brequiring E′ nJ/bit, or through m − 1 intermediate

nodes requiring m.E nJ/bit. For transmission overlarger distances, the latter is often more efficient(depending on the number of intermediate nodes, thedistance between them, and the energy dissipated in thetransmitter/receiver/amplifier electronics). However,the decision is not clear cut. If the batteries of theintermediate nodes are almost depleted, a direct unicastcan improve the network lifetime. See [6] for an analysisof when direct unicast is cheaper than MTE. As such,there is a clear trade-off between being more energyefficient and reducing the overhead of routing throughnodes without matching subscriptions.

The problem is worsened for multi-hop communica-tion in the semantic overlay network, if each hop in thesemantic overlay network is mapped onto direct unicastor MTE in the physical network. On the right in Figure1, we see how a direct mapping of the routing in thesemantic overlay network can lead to inefficient routingpaths in the physical network. The event routing from bto c can be avoided by a multicast/broadcast at node xto reach nodes b and c.

In this article, we present µC-SemPS, a semanticevent pub/sub system for battery-powered micro-controllers in mobile and ubiquitous computingenvironments that aims to optimize the networklifetime. The main contributions are:

• An event subscription model for efficient seman-tic pub/sub matching

• A semantic-based clustering approach for efficientevent routing



Towards energy-aware semantic publish/subscribe for wireless embedded systems

• A significant energy usage reduction vs. moreconventional pub/sub schemes

The article is structured as follows. After discussingrelated work in section 2, we present in section 3 oursemantic subscription model. The semantic pub/submiddleware and algorithms are explained in section 4.We elaborate on the MiXiM-based experimental resultsin section 5. A concluding overview of our contributionsand opportunities for further work are highlighted insection 6.

2. Related workThe publish/subscribe paradigm has been extensivelystudied as a way for selective information dissemi-nation. Most of the work in this area deals with theefficient dissemination of simple subscription formatsin subject-based and content-based systems. Recently,semantic matching in pub/sub systems is receivingmore and more attention to achieve higher eventexpressiveness.

Current solutions attempt to build semantic-awareness with a simple hierarchical topology ofconcepts (taxonomies of classes and properties).S-ToPSS [12] extends the conventional key-value pair-based systems with methods to process syntacticallydifferent, but semantically equivalent information.S-ToPSS uses an ontology with synonyms, a taxonomyand transformation rules to deal with syntacticallydisparate subscriptions and publications.

The Ontology-based Publish/Subscribe system(OPS) [18] represents subscriptions as RDF graphpatterns and uses a subgraph isomorphism algorithmfor efficient matching. Follow-up research on semanticpub/sub systems of some of the authors resulted in theSemantic Publish/Subscribe System (SPS) [11], whichleverages more of the expressive power of OWL Lite butrelies on a centralized broker with a powerful server todo the semantic matching.

Chand et al. present interest similarity metrics [2]for peer-to-peer overlays of pub/sub networks totackle the lack of expressiveness in many subscriptionlanguages. The containment hierarchy tree that isimplied by the containment-based and similarity-basedproximity metrics is similar to class and propertysubsumption relationships in ontologies. Subsumptionis a description logics concept which also defines apartial order of containment on the individuals. Othermetrics like those proposed in [21] use fuzzy semanticmatching to eliminate the need for exact similarities,but go beyond the capabilities of embedded devices.

Many of the above semantic pub/sub systems gobeyond the capabilities of embedded devices. TheDSWare [10] middleware offers publish/subscribeservices to distributed wireless sensor networks and

supports the specification and detection of patterns ofcompound events.

Resource-aware publish/subscribe in wireless sensornetworks is studied in [17]. The authors propose aprotocol that aims to extend the network’s lifetime byoffering tradeoffs between fixed event disseminationpaths that increase communication efficiency, andresource-awareness that provides freedom for eventrouting. Network hops are used as a measure forthe delivery cost. With energy efficiency being themain concern, hops are not a good measure forquantifying energy usage. Energy consumption of awireless amplifier increases quadratically with the hopdistance, so that transmitting over smaller distances(i.e. with more hops) could be more optimal.

In [8], Demeure et al. present an energy-awaremiddleware for MANETs of handheld devices whichincludes a publish/subscribe event system. The middle-ware is made energy-aware by providing policies thatspecify adaptations in the middleware for a variety ofbattery thresholds. These adaptations may trigger theuse of more energy-friendly encryption algorithms orchange communication by using a non-acknowledgedprotocol to reduce network activity. The authors do notimmediately focus on optimizing the event matchingand routing. For an overview of energy-aware routingin wireless sensor networks, we refer to the survey ofAkkaya et al. [1].

3. Semantic event subscription modelBefore we discuss the event and concept models usedin µC-SemPS, our semantic pub/sub system, we willbriefly describe the energy model that we use tosimulate energy consumption due to computation andcommunication. The energy model is fairly simpleand assumes ideal operational circumstances, buthelps us to simulate large scale sensor networks andcompare different algorithms under the same workingconditions.

3.1. Energy usage approximation modelWe reuse the same first order radio model forcommunication presented in the widely referencedwork [6] by Heinzelman et al. This is a simplemodel where the radio dissipates Eelec = 50nJ/bit torun the transmitter and receiver circuitry and εamp =100pJ/bit/m2 for the transmit amplifier. To transmit ak-bit message a distance d away, the radio dissipates:

ET x(k, d) = Eelec ∗ k + εamp ∗ k ∗ d2

To receive this message, the radio dissipates:

ERx(k) = Eelec ∗ k




Citizen

Foreigner

Person

Male

Female

Student Employee

TeachingAssistant

AssociateProfessor

TenuredProfessor

UndergradStudent

MastersStudent

PhDStudent

alicebobpeggysteve

isParentOf isFatherOf

Inheritance

Property/role

isFatherOf

isParentOf

Class

Property

Individuals

Stranger

Equiv. class

Staff

Equivalence

2

2

3

3 5 5

7

17

11

13

19

2323

29

3131

37

41

Figure 2. Inheritance of ontology classes and properties with prime number assignment

For a publisher to send a message of 64 bytes (or 512bits) with direct unicast to a subscriber 50 meters away,the total energy consumption for communication wouldbe:

Edirect = ET x(512, 50) + ERx(512)

= 2 ∗ 50nJ/bit ∗ 512bits

+ 100pJ/bit/m2 ∗ 512bits ∗ 2500m2

= 132µJ

Routing the message over the same distance but with re-transmissions by four intermediate nodes at 10 metersapart would require:

Emte = 5 ∗ {ET x(512, 10) + ERx(512)}= 5 ∗ {2 ∗ 50nJ/bit ∗ 512bits

+ 100pJ/bit/m2 ∗ 512bits ∗ 100m2}= 46µJ

For the energy dissipated due to computation,measurements on a variety of microprocessors haveshown that the energy usage varies between 0.5∼4nJ/instruction. Shnayder et al. measured power con-sumption of the Mica2 mote in [15] with similar charac-teristics. For our simulated experiments, we will use an

estimate of 2 nJ/instruction which matches pretty wellwith what some ARM processors consume.

Note that this energy usage approximation modelmakes certain assumptions that may be hard to achievein a particular implementation or setting. For example,we assume ideal circumstances for communicationwhere sending and receiving nodes are awake duringtransmission so that messages do not require anyretransmissions between these nodes. This will be aconcern when nodes are not always on at the sametime. When a receiving node goes to sleep to saveenergy, then a sending node may have to retransmitto get the message across. The impact of timingand coordination issues will be more outspoken fortransmitting messages between wireless nodes overlonger distances and/or multiple hops.

3.2. Concept modelConcept-based semantic publish/subscribe systemsdescribe events and subscriptions at a higher levelof abstraction than simple key-value pairs. Thesemantic relationships are defined in ontologies for anunambiguous interpretation of the event structure. InFigure 2, we illustrate this with a familiar example oftenused for educational purposes. The concept model to




define the semantic meaning of a subscription or eventconsists of the following constructs:

• Inheritance hierarchy of named classes: A class(or concept) can have multiple parent classes andsubclasses. Any individual (entity or instance of aclass) of a class is also a member of its ancestorclasses:

Student ⊆ P erson

• Inheritance hierarchy of properties: An (objector datatype) property defines a role of a class bypointing to another class. A similar inheritancerelationship can also be defined for properties (orroles):

isFatherOf ⊆ isP arentOf

• Equivalence of classes and properties: Classesand properties are defined semantically equiva-lent if they represent the same set of individuals.For example, the class Car could be a synonym forthe class Automobile:

Car ≡ Automobile

Compared to other semantic pub/sub systems,our model also supports equivalence betweenproperties and classes. For example, the classAircraf t is equivalent to the class V ehicle forwhich a property hasW ings is defined (We assumethat Aircraf t ⊆ V ehicle and that hasW ings hasother domains like the class Animal).

Aircraf t ≡ (V ehicle ∧ hasW ings)

• Anonymous classes: These are classes for whichno name is specified. They define a set ofindividuals that comply with a set of constraints.Examples include the union or intersection ofclasses, property restrictions on domain or range,complements of classes, etc.

• Custom mappings: These are equivalence rela-tionships defined on the individual level ratherthan on the class level. They allow for express-ing types of semantic similarity that cannot beexpressed in the class or property hierarchies.

Synonyms and the two types of inheritance relation-ships are typically found in most semantic pub/sub sys-tems. The acyclic, reflexive and transitive inheritancerelationship is in the literature sometimes referred toas containment, type inclusion or subsumption.

3.3. Event model and subscription languageThe event model specifies the organization of datainside the events and provides an expressive subscrip-tion language for subscribers to define their interestsin certain events. Rather than using a directed labeledgraph representation for the semantic relationships ofan event of interest, we use the user-friendly compactManchester syntax for OWL [7] to express semanticsubscriptions. For example,

Student or Person that hasAge all xsd:integer[ <= 25 ]would represent with the usual precedence rulesan interest in students and persons that are 25years or younger. Rather than relying on subgraphisomorphisms, we make use of an efficient encoding ofsemantic relationships based on the properties of primenumbers, as explained in the following subsection.

3.4. Semantic event matchingThe encoding presented in this article is proposed tocounter one of the major disadvantages of ontologies:the reasoning tools and technologies underlying thesemantic modeling with ontologies are particularlyexpensive in terms of memory and processing time.They are not designed with the limited computationalresources of an embedded system environment inmind. The prime number based encoding and matchingalgorithms we presented in [14] allow for embeddingsemantic awareness in resource constrained devices.The algorithms were developed with the aim toreduce the memory and processing complexity withseveral orders of magnitude compared to otherstandard ontology reasoning engines. The compactrepresentation of the semantic relationships presentedin this paper is also useful from a communicationperspective because the message lengths of theevents are significantly smaller. We will brieflysummarize some terminology before highlighting thebasic principles of our event matching algorithms:

Hierarchy: This is the partial ordered set of classes(or properties) that reflects the subsumptionrelationships and is represented with a symbol χ.

Child: A class A ∈ χ is a child of a class B ∈ χ if andonly if class A has a direct connection with class Band inherits from class B in the hierarchy. A Leafclass is a special child in the hierarchy that hasno children of its own. The child relationship isdenoted as: A <:d B

Descendant: A class A ∈ χ is a descendant of classB ∈ χ if and only if it is either a child of class Bor a descendant of one of the children of class B.This recursive definition of descendant is denotedas: A <: B. A child is a direct descendant, hence thesubscript d in the child notation.




A

B E

γ(A): 2

γ(B): 2×3 = 6

γ(C): 2×3×5 = 30

γ(E): 2×7×11 = 154

γ(D): 7D

C

Hierarchy χ = { A, B, C, D, E }Roots(χ) = { A, D }Leafs(χ) = { C, E }Parents(E) = { A, D }Children(A) = { B, E }Ancestors(C) = { A, B }Descendants(A) = { B, C, E }

11

72

3

5

Figure 3. Hierarchy encoding using prime numbers

Equivalence: Two classes A and B in the hierarchyχ are defined equivalent if both classes have thesame set of instances or individuals at all times.The equivalence relationship is denoted as: A ≡ B

Subsumption: Refers to the reflexive, transitive andanti-symmetric relationship of classes in thehierarchy χ, which states that a class A ∈ χ issubsumed by a class B ∈ χ if and only if theset of instances of class A are also included inthose of class B. This is the same as statingthat class B subsumes class A. The subsumptionrelationship is denoted as: γ(A) ⊆ γ(B). Note that:γ(A) ⊆ γ(B)⇔ A ≡ B ∨ A <: B.

Gene: A gene is a unique symbolic or numericidentifier that is used to define a subsumptionrelationship between two classes. If a class Ainherits from a class B, then it inherits allthe genes of class B. Each class has a newdifferentiating gene that it only shares with itsdescendants. The differentiating gene of class A isrepresented as gA.

Code: A class is encoded into a compact representationthat allows efficient subsumption testing. Forbit vector representations, the encoding is oftendetermined by first assigning to each gene adistinguishing bit position in the bit vector, andthen for each class setting the correspondingbits of its genes to 1. The code of a class A isrepresented as γ(A).

Given these definitions and notations for specifyingrelationships between classes in a multiple inheritancehierarchy, we declare the following functions for a classA ∈ χ:

parents(A) = {B ∈ χ|A <:d B}children(A) = {B ∈ χ|B <:d A}ancestors(A) = {B ∈ χ|A <: B}descendants(A) = {B ∈ χ|B <: A}roots(A) = {B ∈ ancestors(A)|parents(B) = ∅}leaves(A) = {B ∈ descendants(A)|children(B) = ∅}

Semantic event matching requires an efficient encodingfor subsumption testing. Bitvectors do not work wellfor ontologies because publishers and subscribersdo not know all the concepts in advance (closedworld assumption). Also, representing inheritancerelationships as bitvectors results in a sparse binarymatrix, i.e. the matrix contains many 0’s, because thesize of the hierarchy is usually much larger than thebranching factor of a class in the hierarchy. Therefore,we will only encode a reference to the ancestors inthe representation of a class by assigning a numberas differentiating gene g ∈ G to each class that is co-prime with all the other genes. The encoding γ(A) ofclass A can then be defined as the multiplication of itsdifferentiating gene gA with the inherited genes of itsancestors. This way, a classA ∈ χ subsumes a class B ∈ χif and only if the gene gA = ϕ(A) of class A divides theencoding γ(B) of class B.

The encoding of a multiple inheritance hierarchyχ = {A, B, C,D, E} is illustrated in Figure 3. Thedifferentiating gene of each class is underlined. Forexample, class E inherits the genes 2 and 7 fromits ancestors and is assigned 11 as its differentiatinggene gE . The encoding of class E becomes γ(E) = 2 ×7 × 11 = 154. As classes A and D have no ancestors,their encoding γ(A) and γ(D) is solely based on theirdifferentiating gene gA = 2 and gD = 7. The degreeof compaction solely depends on the prime numberp that is assigned to a class C. The objective isto assign prime numbers to classes in such a waythat after multiplication the length of the encoding− either for the individual class with the longestencoding or the total encoding length for all classes




S1

S3

S4

Female S2

S5

Student

Female ∩ PhD Student

Female ∩ Employee

Female ∩ Teaching Assistant

S0

Employee

S4

Foreigner ∩ Employee

S6

Foreigner ∩ Employee ∩ Male

A female PhD student and teaching assistant?

Figure 4. Organizing nodes according to the specificity of their subscriptions

together − is minimized. See [13, 14] for more detailson the heuristics we used to optimize the encodinglength of each class. For the very simple hierarchy inFig. 3, the OWL ontology file is more than 1500 byteslong, whereas our encoding generates an equivalentrepresentation of less than 100 bytes:

A=2,2

B=3,6

C=5,30

D=7,7

E=11,154

The advantages of our semantic encoding are:

• A class (or property) can be identified by a simpleprime number rather than a possibly long stringthat represents its URI.

• All the relevant semantic relationships areembedded in the encoding of each class andproperty.

• Partial imports of an ontology are possiblebecause you only need the class and propertyencodings of the terminology you are interestedin.

These characteristics have an immediate effect on themessage length of events that have to be routed, astextual descriptions are being replaced with a simpleprime number. Message size reductions of a factor of50 or more were not uncommon in our experiments.Also, because all the relationships are already explicitlyrepresented in the class encoding, we can eliminatethe computational overhead of classifying an ontology(a logical inference process that makes all implicitsemantic relationships explicit) at runtime. The factthat our representation is line oriented makes it alsoeasier to parse than XML-based file formats.

The above improvements also help to reduce the sizeof the subscriptions. For example, a subscription for

tenured professors who are the grandfather of a student,would be in the compact OWL Manchester syntax:

TenuredProfessor thatisFatherOf some (Person that isParentOf some Student)

This is twice an intersection of a class and propertyrestriction. Each label is replaced with the correspond-ing prime number (see Fig. 2 for the prime numberassignment):

13 ∩ ∃ 3 (2 ∩ ∃ 2 3)

Any individual with a class encoding that can bedivided by 3 is a Student (or a subclass thereof). Forexample, a PhD Student has encoding 2 ∗ 3 ∗ 23 = 138and can be verified to be an instance of both Personand Student. The same reasoning can be applied with2 and 13 for the classes Person and Tenured Professorrespectively, as well as for properties. Note that inmany cases we do not have to carry out the division,but use faster heuristics to rule out a semantic match.For example, to optimize the encoding length weensure that the differentiating gene of a subclass isalways bigger than that of its parent. Therefore, wecan quickly rule out that C does not inherit from Dby comparing two prime numbers. See our matchingalgorithms in [14] for more details.

4. Publish/subscribe for low-power µC systemsAs computation consumes less energy than wirelesscommunication and to reduce the impact of retrans-missions on the energy consumption, it is clear thatthe most energy savings can be gained by optimiz-ing the wireless communication. Therefore, our µC-SemPS pub/sub subsystem for microcontrollers willfavor spending more effort in the semantic event match-ing if it can further reduce the energy impact of seman-tic event routing.




Algorithm 1 BuildSubscriptionHierarchy(in: fromPeer, subscription)

1: (subscriptionRelevant, subscriptionForwarded) = (false, false)2: if (AlreadyReceived(subscription)) then3: FeedbackSubscription(fromPeer, DUPLICATE, subscription)4: else5: MarkReceived(fromPeer, subscription)6: if (IsMatchingAncestor(subscription)) then7: subscriptionRelevant = true8: else9: LabelSubscription(subscription, OUT_OF_INTEREST)

10: if (subscription.hopsLeft > 0) then11: subscription.hopsLeft = subscription.hopsLeft - 112: for each Peer p in ForwardFilter(adjacentPeers, subscription) do13: subscriptionForwarded = true14: ForwardSubscriptions(p, MatchingSubscriptions(subscription))15: if (not subscriptionForwarded) then16: if (not subscriptionRelevant) then17: FeedbackSubscription(fromPeer, OUT_OF_INTEREST, subscription)

4.1. Semantic event subscription polyhierarchyThe event notification routing protocol aims toreduce the delivery of events to nodes withouta matching subscription, as well as the amountof undelivered events for nodes with a matchingsubscription (i.e. maximize precision and recall). Notethat reducing energy consumption does not necessarilymean less messages between peers (as illustrated inFigure 1). The challenge is to organize the peersaccording to their interests in events (i.e. theirsubscriptions) to improve the event routing efficiency(both in terms of event delivery accuracy and energyusage for communication).

Our event routing protocol will first establishsemantic routing tables by organizing the peers into asubscription polyhierarchy according to the specificityof their subscriptions. A subscription for PhD Studentevents is more specific than a subscription for Studentevents (see Figure 2), which in turn is more specificthan a subscription for Person events. Note that thisexample ontology is not that practical for wirelesssensor applications, but it is sufficient to the specificityof subscriptions. All the subscriptions of all nodes areclassified this way in the subscription polyhierarchy.See Figure 4 for an example of a subscriptionpolyhierarchy of named classes and anonymous classes(intersections of named classes).

Nodes that have a very specialized subscription(e.g. Foreign male employee) will be at the bottomof the polyhierarchy, whereas nodes with a verygeneral subscription (e.g. Female) are at the top of thepolyhierarchy. This means if an event matches one ofthe nodes in the polyhierarchy, it will also match witha subscription of all the ancestors of the node (due tothe specificity relationship). Furthermore, if a node has

more than one subscription, it appears multiple timesin the polyhierarchy.

A similar classification can be carried out for thepublishers by comparing the specificity of the eventsthey produce. If we would include the publishers inthe subscription polyhierarchy, then all the descendantsubscriptions of a published event would be morespecific and hence the produced event would notmatch. All the ancestor subscriptions of a publishedevent would be more general, and would consumeit. However, as the specificity of the produced eventis often unknown in advance and typically evolvesquicker than the subscriptions, we do not incorporatethe publishers in the subscription polyhierarchy.

The algorithm for constructing this polyhierarchy(and semantic routing tables) is partially shown inAlgorithm 1. For every subscription a node receives, itbasically checks if it already received the subscriptionvia other routes and whether it has a matching localsubscription (equivalent or subclass). If so, it forwardsits own matching subscriptions to its adjacent peersfor which it did not receive a matching out of interestmessage before. The purpose of the algorithm is toensure that any event matching the subscription of anode always matches those of its ancestors, and that anout of interest event will not match the subscriptions ofthe node’s descendants.

4.2. Energy based clustering and event routingIn principle, routing events from the root nodestowards the leaf nodes (until they stop matching thesubscriptions) is a way to minimize sending events tonodes that do not match their subscriptions. However,there are a few drawbacks with strictly following thisrouting scheme:




ENERGY EFFICIENT ROUTING SCHEMES FOR EVENTS

S3

S2

S1

Energy budget every node:100 units

5 units3 units

S3

S2

S1

S3

S2

S1

4 units

a) Minimum energy:100 units / 5 units = 20 events

Total energy:20*5 units = 100 units

b) Minimum energy path:100 units / 4 units = 25 events

Total energy:25*(3+4) = 175 units

S3

S2

S1

c) Maximum lifetime:3x+5y ≤ 100, 4x ≤ 100 →x=25,y=5 or 30 events

Total energy:25*(3+4)+5*5 = 200 units

x

x

y

Figure 5. Simplified model for energy usage in different routing scenarios

• The root nodes with broader interests will havea lot more events to forward than the nodes atgreater depths in the polyhierarchy.

• Strictly following the routing path may not beefficient from an energy point of view dependingon the distance between the nodes involved.

Depending on how we want to optimize the energyconsumption, we can define different routing schemesas shown in Figure 5.

This figure shows different scenarios for the casewhere an event is published by S1 to which S2 and S3are subscribed:

• (a) This scenario minimizes the total amount ofenergy by sending the event to all subscribers inone transmission.

• (b) This scenario defines the most energyfriendly routing path by minimizing the energyconsumption of the transmit amplifier on eachnode.

• (c) The third scenario attempts to maximizethe lifetime of the network by mixing bothcombinations in order to transmit as many eventsas possible.

These scenarios are somewhat simplified as theyassume communication takes place under ideal oper-ational circumstances. In each of the three scenarios,

every node had an interest in the published event. Evenin the case where S2 is not subscribed, these approachescould still be useful to improve the lifetime of thenetwork even if it means a node received and needs toforward an event it is not interested in. However, if noneof the subscriptions of the nodes are considered for therouting of the events, we fall back to energy efficientrouting protocols for plain communication (LEACH,PEGASIS, TEEN and derivatives [1]). From a pub/subpoint of view, such an approach would be wastingenergy to routing out-of-interest events, and suffer froma possibly low precision and recall.

Cluster-based data routing [19] is a way to improveenergy consumption. To improve traffic confinementwe cluster subscribers with similar interests in sucha way that a subscriber that receives an event canbroadcast it to all the nodes of the cluster it is part of.This guarantees that within a cluster all of the abovescenarios can be applied in order to balance the energyconsumption over the different nodes in the cluster. Italso implies that we need an event delivery protocolto route an event from any publisher to at least onesubscriber of every matching cluster.

Our clustering algorithm is based on the sameprinciples of the LEACH protocol in which a percentageP of the nodes (here 5%) become cluster heads:

• Cluster heads are self-selected at the beginningof a cycle. In the r-th cycle, a node that has




Clusters(Mixed)

Cluster overlay(MTE)

cluster

cluster head

Mixed:MTE+unicast

Hierarchy partitioning

Figure 6. Grouping into clusters of nodes with similar subscriptions

not become a cluster head during the previous1/P cycles decides to become a cluster head withprobability P /(1 − P (r mod b1/P c))

• Other nodes join all clusters for which thecluster head is an ancestor in the subscriptionpolyhierarchy.

The first step rotates cluster heads in such a waythat current cluster heads have a lower probability tobecome cluster heads in the next cycle. The second stepgroups nodes with similar interests. It is possible that anode does not find a matching cluster, especially if thesenodes have subscribed to very general concepts (locatednear the top in the subsumption polyhierarchy). SeeFigure 6 for an illustration of this grouping intosemantically similar clusters. So rather than having toroute events in the network according to subscriptionpolyhierarchy of Figure 4, we can now route on twodifferent levels:

• Clusters: within a cluster, we use a mixtureof unicasting and minimum transfer energy tomaximize the lifetime of the cluster.

• Cluster overlay: between cluster heads andnodes without a cluster, we route events withminimum transfer energy (MTE) strategy, becauseit increases the lifetime better than unicasting andevent matches are more likely for nodes with verygeneral subscriptions anyway.

This approach basically breaks down into a partitioningof the subscription hierarchy into two parts: (1) thesubhierarchies imposed by the cluster heads and theirdescendants that will form the clusters, and (2) thehierarchy imposed by the ancestor classes of the clusterheads and the nodes without a cluster. We use twodifferent routing strategies in the two partitions, and byreselecting the cluster heads we balance the energy loadamongst all nodes. The efficient encoding and semanticmatching makes the overhead of managing the clusterslow. This is important for mobile nodes that connect toa variety of nodes and for nodes that frequently changetheir subscriptions. However, theoretical situationsexist where the majority of communication is spent onregrouping messages.




Figure 7. Experimental results with the MiXiM wireless network modeling framework

5. Experimental performance evaluationIn this section, we will analyze both energy usageas well as the publish/subscribe effectiveness of theevent delivery (in terms of event routing throughnon-subscriber nodes). To test the effectiveness of thesemantic routing part of our µC-SemPS pub/sub systemon larger setups (500 nodes), we conducted simulationswith MiXiM1, an OMNeT++ modeling frameworkcreated for mobile wireless and ad-hoc sensor networkswith built-in energy framework. We simulated a ZigBeewireless network using energy characteristics of a IEEE802.15.4 low-power digital radio.

In our 500 node setup, we randomly selected 20% of the nodes as publishers each producing singleevents, and another 40 % played the role as subscriberswith up to 3 subscriptions. The remaining unselectednodes were merely used for routing purposes. Ourimplementation supports mobility of nodes and a-periodic event publications, but in order to compareour results with other techniques, we simplified ourexperimental setup to static ad hoc networks withperiodic events. In ongoing work, we are investigatingthe effects of different topology change rates and packetarrival rates, as well as new decision mechanisms forproactive and reactive routing schemes.

1http://mixim.sourceforge.net/

The simulated network depicted in Figure 7 illus-trates publishers, represented as antennas, and sub-scribers represented as personal digital assistants. Thecolor of these symbols represents the type of event thatis being published or subscribed to. The network itselfis generated by means of a weighted location dependentwiring. For k nodes, the first one n1 is situated in thecenter of a unit square. The other nodes nj , j = 2...k, arerandomly placed in the unit square and connected to atmost m existing nodes ni that each minimize a differentfunction Fm:

Fm(ni , nj ) = H(n1, nj ) + wm.D(ni , nj )

with H = number of hops to node n1, D = Euclideandistance, and m different weights wm = parameter thatinfluences the geographical dependency.

5.1. Energy cost of semantic matchingRegarding the computational complexity, we estimatedbased on [14] that verifying for a semantic matchbetween two concepts in a very large ontology withmore than 25000 concepts requires on average about50 instructions on an ARMv4 CPU, which at 2nJ/instruction results in an energy cost of about 100 nJ.

This average was computed by comparing each pairof concepts and verifying whether they semanticallymatch. In case there is no match, we exploit various



http://mixim.sourceforge.net/


0 100 200 300 400 500 600 700 800 900 1000 1100

0

20

40

60

80

100

120

Mobile wireless network lifetime

Direct unicastMTEμC-SemPS

Time steps

No

de

s a

live

Figure 8. Measuring the network lifetime

efficient heuristics to detect this and only in about 5-10% of the cases we actually need to carry out the moreexpensive operation to check whether a big integernumber can be divided by a prime number. Given thatontologies with more than 1000 unique concepts arerather rare, the cost for semantic matching for a realisticpub/sub application will be lower.

5.2. Energy cost of semantic routing

In Figure 8, you can see the advantage of using oursemantic clustering approach with cluster head rotationon the lifetime of the network.

For this particular experiment, the routing protocolsfrom our µC-SemPS semantic publish/subscribe systemare more efficient compared to direct unicasting andminimum energy transmission. The energy for compu-tation (the semantic matching) is quite negligible com-pared to the energy dissipated due to communication(less than 10% for some nodes).

For nodes that were not transmitting themselves,either because they were not a publisher or becausethey did not forward any event, there was stillsome communication energy spent for creating andmaintaining the clusters, but the significant differencebetween communication and computation was lessoutspoken.

Note that these results are very specific for thisexperimental setup. If the average distance between thenodes grows, then this will have an effect on the totalenergy dissipation for all protocols.

5.3. Pub/sub efficiency of event routingWe compared in another experiment the efficiency ofevent routing. In the experiment, we measured howmany out-of-interest events were delivered (countingthe events received by a node that did not have amatching subscription) and missed events (events thatshould have been delivered but were not). These metricsboil down to precision and recall from an informationretrieval point of view.

precision ={relevant event }

⋂{ received event }

{ received event }(1)

recall ={ relevant event }

⋂{ received event }

{ relevant event }(2)

The precision metric computes the fraction of thereceived events that are relevant, i.e. semanticallymatch one of the subscriptions of the node. The recallmetric computes how many of the relevant eventspublished in the network were successfully received bythe node.

A low precision means that a node is spending(and wasting) energy on receiving, processing andforwarding events that do not match any of itssubscriptions. A low recall means that a node misseda lot of relevant events. A 100% recall is easy toachieve by making sure every node receives every event(e.g. through broadcasting every received event), butthis greatly affects the precision. Obviously, from anenergy perspective it would be better if a node did nothave to deal with event at all and this would result in




Direct unicast

MTE

μC-SemPS

0 10 20 30 40 50 60 70 80

Semantic routing efficiency

PrecisionRecall

Precision and recall of delivered events (in %)

Ro

uti

ng

pro

toc

ol

Figure 9. Measuring the delivery of out-of-interest and missed events

an undefined precision and recall, but would also breakthe purpose of the semantic pub/sub system.

See Figure 9 for the results. Due the fact thatMTE uses more intermediate nodes, it is clear that itsprecision is smaller. The max hop count for deliveryalso had an effect on its lower recall. The direct unicastperformed better, but our µC-SemPS system performedeven better due to the semantic clustering approach.

5.4. Bridging the simulation gapIn this section, we briefly discuss the gap that existswhen going from simulations to experiments witha prototype implementation on a real test setup.Although we used the MiXiM energy framework ofthe OMNeT++ discrete event network simulator forwireless and mobile networks to realistically modelconnections between nodes and interference, there arestill some assumptions that we need to revise before wecan get an accurate picture of power consumption on awireless sensor network in the real world.

• Sender and receiver clocks are synchronized withno clock drift

• Nodes can send and receive concurrently, i.e. full-duplex

• Simplified transmission and receiving powermappings for faster simulations

• No interference influences from the environment

Obviously, these assumptions are hard to guaranteein the real world. Time synchronization in wirelesssensor networks are a big concern [16] and mostof the sensor nodes are half-duplex. To get a veryaccurate power consumption, we need to consider allthe above aspects and have a more realistic power

consumption model of the physical layer in the wirelessradio [20] than what we have used in our experiments.One possible approach would be to combine theOMNeT++/MiXiM simulations with state-of-the-artwireless sensor emulators so that we have accuratemeasurements for the overall power consumption −including the energy spent by the CPU − using thenative semantic pub/sub code cross-compiled for thatparticular platform.

However, our goal with µC-SemPS was to havea framework for initial validation of the pub/subalgorithms before trying to implement a prototype fora particular wireless sensor network. Further workon smaller test setups will look into tackling thesechallenges in order to validate the results of oursimulations.

6. ConclusionsAs the publish/subscribe community expands towardsnew computing paradigms like the Internet of Things,it is paramount that semantic awareness and energyconsumption is taken into account in the matchingand routing of events. While many energy-efficientrouting protocols have been proposed for wirelesssensor networks, they only focus on optimal deliverypaths and do not consider optimizing the routing pathsfor precision and recall of the pub/sub system.

This article addresses energy-efficient semanticmatching and routing for low-power wireless andmobile devices. We presented our µC-SemPS seman-tic publish/subscribe for battery-powered microcon-trollers and sensor nodes. Rather than relying on a cen-tralized broker for ontology-based semantic matching,we use an encoding scheme that represents semanticrelationships in an efficient compact representation thathelps reduce the packet size of the network messages.




We have demonstrated our clustering technique thattakes both semantic similarity as well as energy aware-ness into account for the efficient delivery of eventsto subscribers. We have shown that our approach issignificantly better in terms of energy consumption andnetwork lifetime while maintaining a high precisionand recall of delivered events.

A key direction of our future work, will be toinvestigate the effects of node mobility and churnto maintain the subscription polyhierarchy, and tocreate a larger testbed with many more energy-efficientwireless routing protocols. While none of them focuson semantic publish/subscribe systems, it will beinteresting to see how well they fare in terms ofprecision and recall. This could provide interestinginsights into how energy efficiency can be traded formore accurate routing to increase precision and recallfor the semantic events.

AcknowledgmentsThis research is partially funded by the InteruniversityAttraction Poles Programme Belgian State, BelgianScience Policy, and by the Research Fund KU Leuven.

References[1] Akkaya, K. and Younis, M.F. (2005) A survey on routing

protocols for wireless sensor networks. Ad Hoc Networks3(3): 325–349.

[2] Chand, R. and Felber, P.A. (2005) Semantic peer-to-peer overlays for publish/subscribe networks. In Euro-Par 2005 parallel processing, 30 August - 2 September 2005,Lisbon, Portugal - Also published in LNCS Volume 3648.doi:10.1007/11549468_130.

[3] Cilia, M., Bornhövd, C. and Buchmann, A.P. (2003)Cream: An infrastructure for distributed, heterogeneousevent-based applications. In Meersman, R., Tari, Z. andSchmidt, D.C. [eds.] CoopIS/DOA/ODBASE (Springer),Lecture Notes in Computer Science 2888: 482–502.

[4] de Meulenaer, G., Gosset, F., Standaert, F.X. andPereira, O. (2008) On the energy cost of communicationand cryptography in wireless sensor networks. InWIMOB ’08: Proceedings of the 2008 IEEE InternationalConference onWireless &Mobile Computing, Networking &Communication (Washington, DC, USA: IEEE ComputerSociety): 580–585. doi:10.1109/WiMob.2008.16.

[5] Eugster, Felber, Guerraoui and Kermarrec (2003) Themany faces of publish/subscribe. ACM Comput. Surv.35(2): 114–131. doi:10.1145/857076.857078.

[6] Heinzelman, W.R., Chandrakasan, A. and Balakrish-

nan, H. (2000) Energy-efficient communication protocolfor wireless microsensor networks. In HICSS ’00: Pro-ceedings of the 33rd Hawaii International Conference onSystem Sciences-Volume 8 (Washington, DC, USA: IEEEComputer Society): 8020.

[7] Horridge, M. and Patel-Schneider, P. (2008) Manch-ester syntax for owl 1.1. In International Workshop OWL:Experiences and Directions (OWLED 2008).

[8] I. Demeure et al. (2008) An energy-aware middlewarefor collaboration on small scale MANets. In Proceedingsof the Autonomous and Spontaneous Networks Symposium(ASNS’08) (Paris, France).

[9] Jafarpour, Hore, Mehrotra and Venkatasubramanian

(2009) Ccd: efficient customized content disseminationin distributed publish/subscribe. In Middleware ’09:Proceedings of the 10th ACM/IFIP/USENIX InternationalConference on Middleware (New York, NY, USA: Springer-Verlag New York, Inc.): 1–20.

[10] Li, S., Lin, Y., Son, S., Stankovic, J. and Wei, Y. (2003)Event detection services using data service middlewarein distributed sensor networks.

[11] Ma, J., Xu, G., Wang, J. and Huang, T. (2006) A semanticpublish/subscribe system for selective disseminationof the rss documents. In GCC ’06: Proceedings of theFifth International Conference on Grid and CooperativeComputing (Washington, DC, USA: IEEE ComputerSociety): 432–439. doi:10.1109/GCC.2006.19.

[12] Petrovic, M., Burcea, I. and Jacobsen, H. (2003)S-ToPSS: semantic Toronto publish/subscribe system.In VLDB ’2003: Proceedings of the 29th internationalconference on Very large data bases (VLDB Endowment):1101–1104.

[13] Preuveneers, D. and Berbers, Y. (2006) Prime Num-bers Considered Useful: Ontology Encoding for EfficientSubsumption Testing. Tech. Rep. CW464, Departmentof Computer Science, Katholieke Universiteit Leuven,Belgium.

[14] Preuveneers, D. and Berbers, Y. (2008) Encodingsemantic awareness in resource-constrained devices.IEEE Intelligent Systems 23(2): 26–33.

[15] Shnayder, V., Hempstead, M., Chen, B.r., Allen, G.W.

and Welsh, M. (2004) Simulating the power consump-tion of large-scale sensor network applications. In SenSys’04: Proceedings of the 2nd international conference onEmbedded networked sensor systems (New York, NY, USA:ACM): 188–200. doi:10.1145/1031495.1031518.

[16] Sivrikaya, F. and Yener, B. (2004) Time synchronizationin sensor networks: a survey. Network, IEEE 18(4): 45–50.doi:10.1109/MNET.2004.1316761.

[17] Taherian, S. and Bacon, J. (2007) A Publish/SubscribeProtocol for Resource-Awareness in Wireless SensorNetworks. In Proceedings of the International Workshopon Localized Algorithms and Protocols for Wireless SensorNetworks (Springer), Lecture Notes in Computer Science(LNCS) 4549: 27–38.

[18] Wang, Jin and Li (2004) An ontology-based pub-lish/subscribe system. In Middleware ’04: Proceedingsof the 5th ACM/IFIP/USENIX international conference onMiddleware (Springer-Verlag): 232–253.

[19] Wang, H.L. and Chao, Y.Y. (2009) A cluster-based datarouting for wireless sensor networks. In ICA3PP ’09: Pro-ceedings of the 9th International Conference on Algorithmsand Architectures for Parallel Processing (Berlin, Heidel-berg: Springer-Verlag): 129–136. doi:10.1007/978-3-642-03095-6_13.

[20] Wessel, K., Swigulski, M., Köpke, A. and Willkomm,

D. (2009) Mixim: the physical layer an architectureoverview. In Proceedings of the 2nd InternationalConference on Simulation Tools and Techniques,



http://dx.doi.org/10.1007/11549468_130

http://dx.doi.org/10.1109/WiMob.2008.16

http://dx.doi.org/10.1145/857076.857078

http://dx.doi.org/10.1109/GCC.2006.19

http://dx.doi.org/10.1145/1031495.1031518

http://dx.doi.org/10.1109/MNET.2004.1316761

http://dx.doi.org/10.1007/978-3-642-03095-6_13

http://dx.doi.org/10.1007/978-3-642-03095-6_13


Simutools ’09 (ICST, Brussels, Belgium, Belgium: ICST(Institute for Computer Sciences, Social-Informaticsand Telecommunications Engineering)): 78:1–78:8.doi:10.4108/ICST.SIMUTOOLS2009.5555.

[21] Zhang, W., Ma, J. and Ye, D. (2008) Fomatch: Afuzzy ontology-based semantic matching algorithm of

publish/subscribe systems. Computational Intelligence forModelling, Control and Automation, International Confer-ence on 0: 111–117. doi:10.1109/CIMCA.2008.111.



http://dx.doi.org/10.4108/ICST.SIMUTOOLS2009.5555

http://dx.doi.org/10.1109/CIMCA.2008.111

icst transactions on ubiquitous environments research article

Documents