S-Discovery: A Behavioral and Quality-based Service Discovery on the
Cloud
Ahmed Gater
1
, Fernando Lemos
2
, Daniela Grigori
1
and Mokrane Bouzeghoub
2
1
Universit´e Paris-Dauphine, Pl. Mal de Lattre de Tassigny, 75775 Paris, France
2
Universit´e de Versailles Saint-Quentin en Yvelines, 45 Avenue des Etats-Unis, 78035 Versailles Cedex, France
Keywords:
Business Process as a Service, Service Oriented Architecture, Service Discovery, Business Process Matching.
Abstract:
Cloud computing recently emerged as a paradigm providing computer power and storage as a utility that is
consumed on demand (following the footsteps of other utilities, like electricity). Recently, a new service de-
livery mode emerged: Business Process as a Service (BPaaS). As a consequence, process models repositories
will be developed allowing this new type of services to be published by process providers and discovered by
enterprises wanting to outsource some of, or parts of, their processes. In this paper we present the S-Discovery
framework allowing to find in such repositories processes that could satisfy user functional and no-functional
requirements.
1 INTRODUCTION
Cloud computing emerged recently as a paradigm
providing computer power and storage as a utility
that is consumed on demand (following the footsteps
of other utilities, like electricity). It allows typi-
cally three delivery modes: Infrastructure as a Service
(IaaS), Platform as a Service (PaaS), and Software as
a Service (SaaS).
The development of cloud computing offers new
opportunities for enterprises to outsource their pro-
cesses and thus a new service delivery mode emerged:
Business Process as a Service (BPaaS) (Anstett et al.,
2009; Pathirage et al., 2011). As a consequence, pro-
cess models repositories will be developed allowing
this new type of services to be published by pro-
cess providers and discovered by enterprises want-
ing to outsource some of, or parts of, their processes.
Similar to service registries, process repositories will
contain process model descriptions, but also business
specifications (non-functional descriptions like cost,
quality of service, etc). We argue that process model
discovery techniques taking into account functional
and non-functional requirements will be required.
Besides business processes outsourcing, other ap-
plication scenario for BPaaS can be found in the area
of scientific workflows (Pathirage et al., 2011). Many
large scale collaborative science projects use work-
flows to automate computation steps. However, defin-
ing and running such workflow systems are often a
challenge. Workflow engines in the cloud would fa-
cilitate scientist’s work and reduce overhead on their
projects. In the context of web-based scientific work-
flow repositories, scientists expressed the need to
have workflow similarity search capabilities (Goderis
et al., 2006). Moreover, in the new context of BPaaS
repositories, they may be interested in finding, among
the retrieved workflows the one satisfying some busi-
ness criteria (e.g., cost) and some global or local qual-
ity requirements, e.g., the workflow that takes the
shorter time or which guarantees a given correctness
for the results of a specific activity.
These scenarios show the need for a discovery ap-
proach taking into account both process model and
non-functional requirements. The contribution of this
paper is a framework able to efficiently query a repos-
itory to find processes that could best fulfill user struc-
tural and non-functional requirements. We suppose
that users express their query as a process model ac-
companied by non-functional requirements. Thus, the
technical challenges in the discovery approach are
at two levels. At the description level, (i) provide
a formal model that allows one to specify, at differ-
ent granularity levels, non-functional attributes as an-
notations of the functional specification; and (ii) al-
low the user to enrich his query with non-functional
requirements. At the discovery level, (i) filter the
repository to efficiently find matching candidates (ii)
104
Gater A., Lemos F., Grigori D. and Bouzeghoub M..
S-Discovery: A Behavioral and Quality-based Service Discovery on the Cloud.
DOI: 10.5220/0004378201040109
In Proceedings of the 3rd International Conference on Cloud Computing and Services Science (CLOSER-2013), pages 104-109
ISBN: 978-989-8565-52-5
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
combine the structure-based matching algorithms and
non-functional factors matching and (iii) define a sim-
ilarity measure that aggregates both functional and
non-functional similarities.
The remainder of the paper is structured as fol-
lows. Section 2 presents our model to annotate ser-
vice process models. Section 3 describes an overview
of our approach, which is composed of (i) a filtering
task, described in Section 4; (ii) a structural similarity
evaluation task (Section 5); (iii) a preference satisfac-
tion evaluation and ranking tasks (Section 6). Related
work are discussed in Section 7. Finally, the conclu-
sions are presented in Section 8.
2 BACKGROUND AND
NOTATIONS
Our S-Discovery framework is based on matching
techniques that operate on process models. The pro-
cess model consists of a set of related activities that
are organized using control flow structures to con-
struct complex behavior models. To abstract as much
as possible from any existing notation formalism, we
represent a process model as a directed graph p =
(A,C, E,Q), called process graph (p-graph for short),
where A is a set of activity nodes, C is a set of con-
nector nodes, E is a set of directed edges and Q is
a set of quality annotations. An activity node repre-
sents an atomic task which is described by its name
(N), a set of inputs (In), a set of outputs (Out) and a
set of quality annotations (Q). Notice that activity in-
puts and outputs are annotated unsing a domain ontol-
ogy. Connector nodes describe control flows between
activities, and represent Split and Join operators of
types XOR or AND. Split connectors have multiple
outgoing edges, while Join connectors have multiple
incoming edges. Quality annotations are of the form
(m,r), where r is a value for a QoS attribute m. They
can characterize the process as a whole or specific ac-
tivities.
A user query is a p-graph q = (A,C,E,P), where
A, C, E are as defined before and P is a set of QoS
preferences, which are specified as expressions us-
ing the following constructors: around, between, max,
min, like and dislike.
A preferred order between preferences can be de-
fined using the complex constructors pareto (⊗) and
prioritized (&). The semantics of the terms of this
vocabulary were taken from the PreferenceSQL ap-
proach (Kießling, ), howeverthe user may personalize
these semantics by means of a membership function
µ. User can label a preference as hard or soft, the
difference being that a target p-graph must satisfy all
hard preferences, while the satisfaction of soft prefer-
ences is optional.
Figures 1(a) and 1(b) show, respectively, a process
model and a sample user query annotated with hard
and soft preferences.
3 OVERVIEW OF THE
S-DISCOVERY FRAMEWORK
Given an extensive repository of published p-graphs,
the goal of our framework is to retrieve a ranked list of
p-graphs that best fulfill a p-graph query. Our frame-
work is a multi-stepped approach, as illustrated by
Figure 2.
Given a query p-graph q, the Filter module selects
the p-graphs that can most likely answer the query
(p-graphs T
1
,... ,T
n
); it avoids scanning the whole
repository to compare each target p-graph against the
query. This module retrieves all the p-graphs sharing
at least one activity with the query p-graph, and this
is done by relying on an index built on an offline step.
This set is subsequently passed to the Structural
Similarity Evaluator that measures the structural sim-
ilarity λ
struc−i
between each selected p-graph (target)
and the query. Furthermore, a set of activity map-
pings between query and target p-graphs is estab-
lished (mappings M
i
, 1 ≤ i ≤ n).
Next, the Preference Satisfaction Evaluator com-
putes the degree of satisfaction λ
pref−i
of the QoS
preferences at the basis of the mappings computed in
the previous stage. At the end, the retrieved p-graphs
are ranked according to their structural similarity and
the preference satisfaction degrees using a set of ag-
gregation metrics. The following sections present our
contributions to each stage of the S-Discovery frame-
work.
4 REPOSITORY FILTERING
As mentioned previously, a target p-graph is a poten-
tial match of a query p-graph when they have simi-
lar activities. To avoid comparing each query activity
against all the activities stored in the repository, we
defined an index structure over the activities of the
repository p-graphs.
This index is built by assuming that activities hav-
ing similar inputs/outputs are similar (Gater et al.,
2011a). The structure we defined indexes the activ-
ities stored in the repository by attaching to each con-
cept C of the ontology two sets C
In
and C
Out
record-
ing the identifiers of the activities where it appears,
S-Discovery:ABehavioralandQuality-basedServiceDiscoveryontheCloud
105
start
end
XOR
join
XOR
split
returnError
out: errorMsg
createPDF
in: file, fileExt
out: pdfFile
preflight
in: file, fileExt
out: status
[status=ok]
[status=ko]
createLink
out: webLink
start
end
AND
join
AND
split
linkGeneration
out: link
PDFConvertion
in: file, fileExt
out: pdfFile
convertionCheck
in: file, fileExt
out: status
HARD PREFERENCES
: ,
SOFT PREFERENCES
: ,
:
: &
,
: ,
: &
,
SOFT PREFERENCE
:
,
HARD PREFERENCE
: ,
: , $15 ,
:
, 90
: (, 10)
:
, 60%
: (, )
: (, 75)
: (, 25)
: , 15
: (, )
0 ()
1
10 60
0 ()
1
5 20
0 (%)
1
75 100
(a) (b)
Figure 1: Mapping between a (a) sample target p-graph T
1
and a (b) sample query p-graph Q
1
.
Filter
Structural
Similarity
Evaluator
Preference
Satisfaction
Evaluator
Ranker
query’s matches
index
p-graph
repository
query p-graph
Q
…
(
T
1
)
⋮
(
T
n
)
(
T
1
,
M
1
, λ
struc
-1
)
⋮
(
T
n
,
M
n
, λ
struc
-n
)
(
T
1
,
M
1
, λ
struc
-1
, λ
pref
-1
)
⋮
(
T
n
,
M
n
, λ
struc
-n
, λ
pref
-n
)
…
Figure 2: S-Discovery framework architecture.
respectively, as an input or an output. Furthermore,
since we are not only interested in retrieving inexact
matches, the technique should be able to retrieve the
activities that don’t match exactly the query activity
but also those having similar inputs/outputs, i.e. the
activities having inputs/outputs that are semantically
related to those of the query activity. To avoid mak-
ing this computation in real time, the set C
In
(resp.
C
Out
) of a concept C record also the activities having
as input (resp. output) a concept which is semanti-
cally related (its descendants, ascendant, ...) to C.
Thus given a query activity A
q
, the set of its poten-
CLOSER2013-3rdInternationalConferenceonCloudComputingandServicesScience
106
tial activity matches is composed of all the activities
belonging to the set of annotations attached to the in-
puts and outputs of A
q
. Straightforwardly, the set of
process-match candidates of a query is the set of tar-
get p-graphs with whom it shares at least one activity.
5 STRUCTURAL SIMILARITY
EVALUATION
The problem of process matching is reduced to a
graph matching problem and an error-correcting sub-
graph isomorphism (ECSI for short) detection algo-
rithm (Messmer, 1995) was adapted to this purpose.
The principle behind ECSI algorithm is to apply edit
operations (node/edge insertions, deletions and sub-
stitutions) over the target graph until there exists a
subgraph isomorphism to the query graph. Each edit
operation is assigned a cost function, on the basis of
which the quality of an ECSI is estimated. The goal
is then to find the sequence of edit operations leading
to the ECSI between the query and target graphs that
has the minimal cost. To find this sequence, the ECSI
detection algorithm relies on an exhaustive A* search
space algorithm.
The adaptation of this algorithm in order to handle
the p-graphs concerns the definition of: (i) measures
for evaluating the similarity of two activities that inte-
grate the similarity of their names, inputs and outputs;
(ii) measures for evaluating the behavioral/structural
similarities; (iii) heuristics for detecting the granular-
ity level differences.
This algorithm allows evaluating the similarity of
two p-graphs as well as finding a set of correspon-
dences between their activities. Experimental results
showed the effectiveness of this approach in terms of
precision/recall of the found matches. However, the
time complexity induced by the combinatorial search
space limits their application in practice to p-graphs
of relatively small sizes (55 activities). To make this
algorithm tractable, we defined two heuristics that
aim to find matches having acceptable qualities in rea-
sonable execution times. More details are given in
(Gater et al., 2010; Gater et al., 2011b).
6 PREFERENCE SATISFACTION
EVALUATION
After the calculation of the structural similarity be-
tween query and each candidate p-graph, the most
similar ones are subjected to the preference satisfac-
tion evaluation, which calculates the degree to which
the QoS annotations of the p-graphs satisfy the QoS
preferences of the query. The procedure first calcu-
lates the satisfaction degrees of atomic preferences
and, then, it recalculates these degrees based on the
complex preferences. At the end, a global satisfac-
tion degree is obtained from the aggregation of the
preference degrees. The mapping between query and
target p-graphs is used in the evaluation procedure to
recalculate the QoS attributes of target p-graphs and
to evaluate the satisfaction degree of atomic prefer-
ences.
6.1 Atomic and Complex Preference
Evaluation
For each activity correspondence (w,v), the degree
to which each atomic preference of v is satisfied by
its corresponding annotation in w is calculated. The
same is similarly done for the profile preferences.
For a preference p of the type around, between,
min or max, given its corresponding annotation a, the
satisfaction degree δ(p,a) between them is given by
the normalized satisfaction distance d (p,a), which
measures how far is the value r in annotation a from
those favored by preference p. For a preference of the
type likes or dislikes, the satisfaction degree is based
on the semantic similarity between concepts given by
the classic edge counting technique proposed in (Wu
and Palmer, 1994). When a membership function µ
defines the semantics of the preference, the satisfac-
tion degree is a simple application of the function over
the corresponding quality attribute.
When a hard preference is not satisfied, the target
p-graph is eliminated from the discovery result. As a
consequence, the verification of the hard preferences
at activity level may be interwoven with the structural
matching when the latter checks weather two activi-
ties match. This may eliminate the mappings express-
ing structural match but not expressing preference sat-
isfaction and, thus, pruning the search space.
The satisfaction degrees of the atomic preferences
are reevaluated according to the order of importance
defined by the complex preferences (paretoand prior-
itized ). The goal is to assign weights to the satisfac-
tion degrees of atomic preferences to capture the order
of importance defined by complex preferences. This
is done with the help of a preference graph, which is
a rooted directed graph whose nodes represent atomic
preferences, edges represent a prioritized preference
from source to destiny, and each node of the graph
has weight ω
p
=
1
/i, where i is the edge distance from
the node to the graph root. Then, the reevaluation of
a preference p is done by δ
′
(p,a) = δ(p, a) × ω
p
.
S-Discovery:ABehavioralandQuality-basedServiceDiscoveryontheCloud
107
6.2 Preference Satisfaction Degree
Calculation
The global preference satisfaction degree λ
pref
indi-
cates the degree to which the QoS annotations of a
target satisfy the QoS preferences of a query. This
degree is calculated with the help of a preference sat-
isfaction metric, which receives as input the prefer-
ence satisfaction degrees {δ
1
,..., δ
k
} of the target and
aggregates them to provide the global preference sat-
isfaction degree λ
pref
. Our framework provides three
different metrics.
The average-based metric calculates the prefer-
ence satisfaction degree λ
pref
as the average of the
preference satisfaction degrees {δ
1
,..., δ
k
}.
The linguistic quantifier-based metric calculates
the degree λ
pref
by measuring the truth degree of
the sentence γ : “almost all preferences are satisfied”.
This sentence is a fuzzy quantified proposition de-
fined using a relative quantifier (e.g., almost all, at
least, around half, etc.) (Gl¨ockner, 2004).
The bipolar-based metric calculates the degree
λ
pref
by evaluating the bipolar condition (Dubois and
Prade, 2008) “all hard preferences are satisfied and
if possible at least one soft preference is satisfied”.
This method returns a bipolar degree of the form
λ
pref
=
δ
hp
,δ
sp
meaning that “all hard preferences
are satisfied to at least a degree of δ
hp
and at least one
soft preference is satisfied to at least a degree of δ
sp
”.
More details are given in (Lemos et al., 2012).
Once the structural similarity and quality satisfac-
tion degrees are computed, the retrieved p-graphs are
subsequently ranked according to the structural and
quality satisfaction degrees using aggregation tech-
niques. Our framework proposes a set of aggregation
techniques (lexicographic order, weighted average,
fuzzy-based techniques) that are detailed in (Lemos
et al., 2012).
7 RELATED WORK
Our work addresses an important topic in the area of
service oriented architecture, which is the discovery
of services based on their process models. Several
work have been proposed similarity measures (Wom-
bacher et al., ; de Medeiros et al., 2008; Dijkman
et al., ) for the evaluation of the similarity of two
service process models. These approaches proposed
similarity measures that consider either the struc-
tural or behavioral perspectives of the process mod-
els. While structure-based approaches consider the
process topologies, behavior-based approaches con-
sider the execution semantics of the process models.
In this case, the service process discovery is done
by comparing the query against each target service,
and subsequently ranking target processes according
to their closeness to the query. To avoid browsing
the whole process repository, some approaches rely
on indexing structures (Gater et al., 2011a; Awad and
Sakr, 2010; Yan et al., 2010) to quickly retrieve the
processes that are the most likely similar to a (part of)
process query.
With regard to the quality-basedservice discovery,
current approaches (Mokhtar et al., ; Agarwal et al.,
2009) consider services as black boxes, so quality re-
quirements are defined over the service profile. Gen-
erally, they specify quality preferences as relational
expressions, fuzzy sets, linguistic variables, or utility
functions. These approaches do not propose prefer-
ence constructors to help user better define and com-
pose his preferences and they are not abstract enough
to be adapted to different non-functional contexts.
While process similarity search is an active field in
the domain of business process management research
area, little attention was given until now to the dis-
covery of the services hosted in the cloud (Goscin-
ski and Brock, 2010); the existing techniques are lim-
ited in what information can be used when publish-
ing and discovering services (Microsoft Azure (Mi-
crosoft, Inc., )). To the best of our knowledge there
are no process discovery framework allowing to com-
bine functional and non-functional requirements.
8 CONCLUSIONS
In this paper we presented a framework for process
discovery taking into account both functional and
non-functional criteria. User query is expressed as a
process model adorned with quality annotations ex-
pressing user preferences and requirements. Our ap-
proach will allow searching process repositories of-
fered by BPaaS providers.
In our past work, we implemented basic match-
ing operators and evaluated them in terms of effi-
ciency and effectiveness. Our future work consists
in building a prototype implementing the framework
presented in this paper by reusing and adapting our
matching operators.
REFERENCES
Agarwal, S., Lamparter, S., and Studer, R. (2009). Mak-
ing Web services tradable: A policy-based approach
for specifying preferences on Web service properties.
Journal of Web Semantics, 7(1):11–20.
CLOSER2013-3rdInternationalConferenceonCloudComputingandServicesScience
108
Anstett, T., Leymann, F., Mietzner, R., and Strauch, S.
(2009). Towards bpel in the cloud: Exploiting dif-
ferent delivery models for the execution of business
processes. In SERVICES’2009), pages 670–677.
Awad, A. and Sakr, S. (2010). Querying graph-based repos-
itories of business process models. In DASFAA Work-
shops’10.
de Medeiros, A. K. A., van der Aalst, W. M. P., and Weijters,
A. (2008). Quantifying process equivalence based on
observed behavior. In Journal DKE, pages 55–74.
Dijkman, R., Dumas, M., and Garc´ıa-Ba˜nuelos, L. Graph
matching algorithms for business process model sim-
ilarity search. In BPM’09).
Dubois, D. and Prade, H. (2008). Handling bipolar
queries in fuzzy information processing. In Hand-
book of Research on Fuzzy Information Processing in
Databases, pages 97–114. IGI Global.
Gater, A., Grigori, D., and Bouzeghoub, M. (2010). Com-
plex mapping discovery for semantic process model
alignment. In IIWAS’10, pages 317–324.
Gater, A., Grigori, D., and Bouzeghoub, M. (2011a). In-
dexing process model flow dependencies for similar-
ity search. In COOPIS’2012.
Gater, A., Grigori, D., Haddad, M., Bouzeghoub, M.,
and Kheddouci, H. (2011b). A summary-based ap-
proach for enhancing process model matchmaking. In
SOCA’11, pages 1–8.
Gl¨ockner, I. (2004). Fuzzy Quantifiers in Natural Lan-
guage: Semantics and Computational Models. Der
Andere Verlag.
Goderis, A., Li, P., and Goble, C. A. (2006). Workflow
discovery: the problem, a case study from e-Science
and a graph-based solution. In ICWS’06, pages 312–
319.
Goscinski, A. and Brock, M. (2010). Toward dynamic and
attribute based publication, discovery and selection
for cloud computing. FGCS’10.
Kießling, W. Foundations of preferences in database sys-
tems. In VLDB’02.
Lemos, F., Abbaci, K., Grigori, D., Hadjali, A.,
Bouzeghoub, M., Li´etard, L., and Rocacher, D.
(2012). Int´egration de pr´ef´erences dans la d´ecouverte
et la s´election des services web: Une approche fond´ee
sur les ensembles flous. Ing´enierie des Syst`emes
d’Information.
Messmer, B. T. (1995). Graph Matching Algorithms and
Applications. PhD thesis, University of Bern, Switzer-
land.
Microsoft, Inc. Microsoft Azure.
http://www.windowsazure.com.
Mokhtar, S. B., Preuveneers, D., Georgantas, N., Issarny,
V., and Berbers, Y. EASY: Efficient semAntic Service
discoverY in pervasive computing environments with
QoS and context support. JSS’08.
Pathirage, M., Perera, S., Kumara, I., and Weerawarana, S.
(2011). A multi-tenant architecture for business pro-
cess executions. In ICWS’11, pages 121–128.
Wombacher, A., Fankhauser, P., Mahleko, B., and Neuhold,
E. Matchmaking for business processes based on
choreographies. In EEE’04.
Wu, Z. and Palmer, M. S. (1994). Verb semantics and lexi-
cal selection. In 32nd Annual Meeting of the Associa-
tion for Computational Linguistics (ACL 1994), pages
133–138.
Yan, Z., Dijkman, R. M., and Grefen, P. (2010). Fast busi-
ness process similarity search with feature-based sim-
ilarity estimation. In CoopIS’10.
S-Discovery:ABehavioralandQuality-basedServiceDiscoveryontheCloud
109
