Proactive Adaptation in Service Composition using a Fuzzy Logic
Based Optimization Mechanism
Silvana de Gyvés Avila and Karim Djemame
School of Computing, University of Leeds, LS2 9JT, Leeds, U.K.
Keywords: Web Service Composition, Proactive Adaptation, Fuzzy Logic, Optimization, Quality of Service.
Abstract: The importance of Quality of Service management in service oriented environments has brought the need of
QoS aware solutions. Proactive adaptation approaches enable composite services to detect in advance,
according to their QoS values, the need for a change in order to prevent upcoming problems, and maintain
the functional and quality levels of the composition. This paper presents a proactive adaptation mechanism
that implements self-optimization based on fuzzy logic. The optimization model uses two fuzzy inference
systems that evaluate the QoS values of composite services, based on historical and freshly collected data,
and decide if adaptation is needed or not. Experimental results show significant improvements in the global
QoS of the use case scenarios, providing reductions of up to 8.9% in response time and 14.7% in energy
consumption, and an improvement of 41% in availability; this is achieved with an average increment in cost
of 11.75 %.
1 INTRODUCTION
A composite service is a software solution with
specific functionalities that can be seen as an atomic
component in other service compositions, or as a
final solution to be used by a consumer. The process
of developing a composite service is called service
composition, which consists in combining, in a
structured way, the features provided by different
services (Dustdar and Schreiner, 2005).
The nature of service composition, dynamicity
offered by the environments where services are
executed and growing amount of available services
(that may provide the same functionality), have
brought the need of mechanisms focused in ensuring
that the consumer will obtain the expected results
when invoking a composite service. To achieve this
goal, it is important to consider the QoS aspects of
the services involved in the composition, as their
drawbacks will be inherited by the composite
service. However, knowing the QoS of the
components is not enough to warranty the behaviour
of the composition, as unexpected events may occur
at runtime, for example, services becoming
unavailable or showing discrepancies in their QoS
(Châtel et al., 2010). As a result, various adaptive
mechanisms have been proposed in order to restore
and maintain the functional and quality aspects of
the composition. The aim of adaptive mechanisms is
to provide composite services with capabilities that
enable them to morph and function in spite of
internal and external changes, searching to maximize
the composition potential and reducing as much as
possible human involvement (Zeginis and
Plexousakis, 2010). Based on the moment when
adjustments take place, adaptation approaches are
classified as either reactive or proactive. The former
corresponds to adaptation actions performed in
response to an incident, while the later is related to
actions taken in advance, before an incident impacts
the system (Metzger, 2011).
This paper introduces a proactive adaptation
approach for service composition that implements a
self-optimization solution based on fuzzy logic.
Fuzzy logic is an approximate reasoning technique
suitable to deal with uncertainty (Zadeh, 1994),
which can be used to support decision making in
software systems. Current work in proactive
adaptation for service composition is mainly focused
on dealing with the decrease of the QoS values and
service failures (Aschoff and Zisman, 2012,
Yuelong et al., 2012, Leitner et al., 2010). On the
other hand, approaches related to self-optimization
are focused on the selection of services that provide
the most appropriate QoS levels for the composition
257
de Gyvés Avila S. and Djemame K..
Proactive Adaptation in Service Composition using a Fuzzy Logic Based Optimization Mechanism.
DOI: 10.5220/0004820902570267
In Proceedings of the 4th International Conference on Cloud Computing and Services Science (CLOSER-2014), pages 257-267
ISBN: 978-989-758-019-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
(Ardagna et al., 2011, Calinescu et al., 2011,
Cardellini et al., 2012).
The proposed optimization model combines the
analysis of historical QoS data and fresh data
(collected at runtime from the different stages of the
composite service execution) in order to identify the
need of adaptation, which can be related to prevent a
decrease in the global QoS of the composition, but
also, the possibility of improving the global QoS
levels. Composite services are considered to be
workflows formed by tasks, and tasks to be Web
service invocations. The use of the fuzzy support
systems enables the evaluation of the QoS
parameters and helps deciding whether adaptation is
needed or not. If adaptation is needed, the fuzzy
systems provide the parameters to be used during the
service selection process.
The approach has been implemented in a service
composition framework and evaluated through the
execution of two test cases. Results were compared
with a non-adaptive approach. The major
contributions for this paper are:
The optimization approach for service
composition that evaluates the benefit of
adaptation.
The use of fuzzy logic as a decision making tool
to determine the need of adaptation in the context
of proactive adaptation in service composition.
The remainder of the paper is structured as follows:
background is briefly described in Section 2. The
proposed framework, service selection and
optimization models are described in Section 3.
Section 4 presents the experimental description and
results. Section 5 discusses some related work.
Conclusion and future work are given in Section 6.
2 BACKGROUND
2.1 Adaptation in Service Composition
Adaptive mechanisms provide software systems
with capabilities to: self-heal, self-configure, self-
optimize, self-protect, etc., which are implemented
considering the objectives the system should
achieve, the causes of adaptation, the system
reaction towards change and the impact of
adaptation upon the system (Cheng et al., 2009).
Adaptation in service composition aims to mitigate
the impact of unexpected events that take place
during the execution of composite services,
maintaining functional and quality of service levels.
Important aspects that can be considered as part of
adaptation solutions in service composition are listed
as follows (Cardellini et al., 2012):
Adaptation goal is the purpose of adaptation,
functional and/or non-functional (QoS).
Adaptation level defines those elements that will
change in order to achieve the adaptation goal.
Adaptation actions are those used to solve the
adaptation problem.
Adaptive mechanisms correspond to the
approaches applied to implement the adaptation
actions (e.g. agent-based, policy-based, rule-
based, etc.).
Stage of adaptation is the time when adaptation is
performed (development time, compile/link time,
load time and runtime).
Awareness levels describe the scope of
information that will be available in order to
adapt (Dustdar et al., 2010).
2.2 Reactive vs Proactive Adaptation
In service-based applications, reactive adaptation is
triggered after problems have occurred, when
situations like the use of faulty services or services
that present undesirable QoS have already affected
the application (Hielscher et al., 2008). The use of
reactive mechanisms may cause increases in the
execution time and financial loss, which can lead to
user and business dissatisfaction (Aschoff and
Zisman, 2012). Proactive approaches aim to deal
with some of these drawbacks by detecting the need
for a change, before reaching a point where a
problem may occur.
Situations that can be predicted in proactive
adaptation approaches for service composition
include: the impact of a new requirement,
misbehaviour of a service and the existence of new
services (Aschoff and Zisman, 2012). Techniques
such as data mining, online testing, statistical
analysis, runtime verification and simulation are
applied during the prediction stage of the process
with the aim of accurately predict the future
behaviour of the system (Metzger, 2011).
2.3 Fuzzy Logic
Fuzzy logic is a method based on multi-valued logic
which aims to formalize approximate reasoning
(Zadeh, 1994). It is used to deal with different types
of uncertainty in knowledge-based systems. Some of
the relevant characteristics of fuzzy logic are fuzzy
sets, linguistic variables and fuzzy rules. A fuzzy set
is a collection of objects characterized by a
membership function with a continuous grade of
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
258
membership which can be ranged between zero and
one (Zadeh, 1965). A linguistic variable is a type of
variable that uses words instead of numbers to
represent its values (e.g. slow, medium, fast) (Zadeh,
1994). The values used to define linguistic variables
are called terms and the collection of terms is called
the term set. A fuzzy rule is used to represent human
knowledge using the form of IF-THEN within a
fuzzy system (Li-Xin Wang, 1997).
During the execution of a fuzzy system, crisp
inputs are converted to linguistic variables, this
process is known as fuzzification. The variables’
values are then evaluated using fuzzy rules,
generating the linguistic values for the outputs.
Finally, the defuzzification method uses these values
to obtain crisp outputs values.
3 SYSTEM MODEL
An overview of the system model considered in this
work is illustrated in Figure 1, which shows its core
components: composition engine, adaptation
manager, service binder, service selector, predictor
and the sensors; and their interactions. This model
was implemented with the aim of evaluating the
proposed approach, enabling the execution of QoS
aware service composition in an environment with
proactive capabilities.
The composition engine is the software platform
responsible for executing the composite services
(processes’ definitions) and hosting the components
in charge of the adaptation process. Composite
services are considered to consist of a series of
abstract tasks that will be linked to executable
services at runtime.
The adaptation manager works semi-independent
of the rest of the components and is constantly
monitoring and analyzing not only information
collected by the sensors, but also the historical data.
The use of historical data helps understanding the
service behaviour and enables the detection of any
possible deviation in the values of the QoS
parameters. During the execution of a composite
service, sensors collect fresh data, looking at activity
and service levels, and send this information to the
monitor. The monitor queries the historical database
to obtain information about previous executions and
states of the current service, then, sends this
information to the analyzer, which evaluates both,
fresh and historical data, in order to determine the
need of adaptation. If adaptation is needed, the
analyzer sends a request of adaptation to the planner,
which obtains the adaptation values that will be sent
Figure 1: System model.
to the adapter. This information is forwarded to the
service binder, in order to maintain/improve the QoS
of the composition.
For each task in the composite service, the service
binder invokes the service selector with the desired
characteristics that the component service should
provide. The service selector performs a search in the
service registry based on the provided functional
requirements. For each of the pre-selected services
(candidates), the service selector invokes the
predictor to obtain its estimated QoS. This
information is sent to the service binder, which
compares the candidates and selects the service that
suits the request. If the need of a change was
identified by the adaptation manager, the binder uses
the adaptation values to perform the ranking and
selection tasks.
It is considered that at the time of invoking a
composite service, the system has available data from
previous executions of the different possible
components, in order to obtain accurate predictions
about these components’ quality characteristics.
Also, for each task of the composite service, there
exist at least two concrete services to invoke.
3.1 QoS Model
Services that offer the same functionality may be
associated with several QoS attributes (Cardoso et
al., 2004, Zeng et al., 2004), providing different QoS
levels. By evaluating these attributes within a set of
services that share the same goals, consumers can
search/select components to be used in their
applications.
The QoS attributes of a service can be evaluated
during design and execution time. At design time,
these attributes help in order to build a composite
service based on the QoS requirements of the user.
While at execution time, they can be monitored to
maintain the desired QoS level. Information about
Sensors
ProactiveAdaptationinServiceCompositionusingaFuzzyLogicBasedOptimizationMechanism
259
these attributes can be obtained from the service’s
profile (Hwang et al., 2007), nevertheless, when this
information is not available, it can be obtained by
analyzing data collected from past invocations
(Cardoso et al., 2004).
In this work, the quality attributes that will be
considered for each service are response time, cost,
energy consumption and availability.
Response time: time consumed between the
invocation and completion of the service
operation (Dai et al., 2009);
Cost: fee charged to the consumer when
invoking a service (Cardellini et al., 2012);
Energy consumption: amount of power
consumed by a server over a period of time
(Buyya et al., 2010);
Availability: probability that the service is up
and ready for immediate consumption (W3C
Working Group, 2003).
Considering response time and cost enables the
selection of faster and cheaper services, providing a
competitive advantage (Cardoso et al., 2004). Both
parameters have been used in other approaches, like
those presented in (Dai et al., 2009, Ying et al.,
2009, Ardagna and Mirandola, 2010, Cardellini et
al., 2012).
The amount of energy used by data centres has
not only economical but also environmental impacts.
Energy efficiency is becoming a key topic due to
high energy costs and governments’ pressure to
reduce carbon footprints (Kaplan et al., 2009).
Energy consumption has been selected as the third
parameter because of the importance of energy
efficiency when managing computing infrastructure
and services.
The last parameter that has been selected is
availability. By knowing the availability values of
the different services, it is possible to select a subset
of components that will provide a composition with
higher probabilities to be fulfilled. Work that
considers availability has been presented in (Huang
et al., 2009, Canfora et al., 2008, Zeng et al., 2004).
To compute the values of these parameters at
execution time, three situations have been
considered within the composite service structure:
single, sequential and concurrent service
invocations. When computing the QoS parameters of
a single service invocation, the QoS values of the
activity that performs the invocation corresponds to
the QoS values of the invoked service. For activities
in a sequential structure, the values of response time,
cost and energy consumption are summed for the
different activities with service invocations, while
availability is obtained by multiplying them.
∑
(1)
∑
(2)
∑
(3)
∏
(4)
For activities in a concurrent/parallel structure,
the value of response time is considered as the
maximum response time of the completed activities;
values of cost and energy consumption are summed;
and availability is the minimum availability value
among the service invocations within the structure.
,..,
(5)
∑
(6)
∑
(7)
,..,
(8)
For equations (1) to (8),
corresponds to an
activity with a service invocation within the
composite service .
3.2 Service Selection Model
Estimation of QoS values is a key step during service
selection process. Estimated values are calculated
using historical QoS data recorded from previous
executions. This data is filtered, discarding values
considered as outliers and the average of the last
executions of the remaining subset is obtained.
Concrete services are searched in the registry by
name, assuming that this parameter
includes/describes the service’s functionality. The
resulting set of candidate services is sorted according
to the relationship between their estimated QoS
values. Due to these attributes having different units
of measure, their raw values are normalized before
being processed and ranked. The following formula
is used to normalize response time, cost and energy
consumption, which are negative parameters (lower
the value, higher the quality).
(9)
A different formula is used for availability, as it is
a positive parameter (higher the value, higher the
quality).
(10)
In both equations,
and
correspond to
the maximum and minimum values of the evaluated
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
260
QoS parameter, respectively; and
correspond to
the estimated value for the next execution. When
=
, then
= 1. After normalizing the
corresponding values, results are computed using the
Simple Additive Weighting formula:
(11)
Where
is the service estimated response time;
is the service estimated cost;
is the service
estimated energy consumption;
is the service
estimated availability; and
,
,
and
correspond to assigned weights where
,
,
,
≤ 1 and
+
+
+
= 1.
3.3 Optimization Model
The proposed optimization model works as part of a
proactive adaptation mechanism. It combines the
analysis of historical and fresh data. QoS information
of the different services and states of the composition
is collected from service, task and process
perspectives, where service corresponds to concrete
Web services; task to elements within the composite
service that invoke services; and process to the entire
composition (service workflow). Based on this
information, it is possible to take decisions about
future actions.
The QoS parameters are obtained when the
service invocation is performed. Response time is
measured during the service’s execution; the values
of cost and energy consumption are retrieved from
the service’s WSDL file; while the value of
availability is obtained based on historical data.
According to the structures of the composite service,
the QoS values of each task are computed using
equations (1) to (8) and stored in the historical QoS
data base, considering both individual values and
accumulated. These values are used in order to obtain
the global QoS of the composite service.
The service selection model previously described
uses as weights for equation (11) the results of the
optimization model evaluation. This model is based
on two fuzzy support systems, which assess the QoS
of the composition, determine the need of adaptation
and, when adaptation is needed, obtain the weight
values to be used during service selection. The
optimization mechanism identifies when the QoS of
the composition is decreasing. It also considers
situations where a number of the accumulated QoS
values of the previous activity in the process are
better than expected, which provides the possibility
of improving other QoS parameters.
The idea of using fuzzy logic is to understand the
relationship between the QoS values of the
composite service and the need of adaptation. In this
context, QoS parameters can be expressed using
linguistic variables. Two inference engines have been
defined to 1) obtain the benefit of adaptation, 2)
obtain the weights to be used during service
selection. Each of these systems uses its own
linguistic variables and rules.
The first fuzzy support system evaluates the QoS
of the composite service every milliseconds, in
order to identify as soon as possible the need of
adaptation. It uses as inputs the measured QoS values
collected from the composite service execution. The
defined input variables are response time, cost,
energy consumption and availability, which are
expressed with three terms low, medium and high.
To establish these terms for each of the linguistic
variables, an interval is defined at runtime using data
collected from previous executions. Historical data is
analyzed, obtaining maximum/minimum values and
standard deviations from each of the QoS parameters.
Sigmoidal functions (open to the left and right) are
used to define the low and high terms, while Gauss
function is used to define the medium term. The
system takes the inputs and based on the
corresponding fuzzy rules, provides the estimated
benefit of adaptation. Four different levels of benefit
of adaptation (low, medium, high and very high)
were established, falling in the interval [0, 1], and
defined with Gauss functions.
The second fuzzy support system uses the value
of the benefit of adaptation (output of the first
system) and the errors between the estimated and the
measured QoS as inputs. The error value is computed
per each parameter using the following formula:
(12)
Where
is the estimated data; and
is
the real measured data.
Input variables corresponding to the QoS errors
are expressed with three terms, low, medium and
high, falling in the interval [-1, +1]. Benefit of
adaptation is expressed with four terms, as defined in
the first fuzzy system.
By evaluating the different errors and the benefit
of adaptation, the system provides the values to be
used as weights during the service selection process.
Output variables (response time weight, cost weight,
energy consumption weight and availability weight)
are expressed with five terms, very low, low,
medium, high and very high, falling in the interval
[0,1] and are defined using Gauss functions.
ProactiveAdaptationinServiceCompositionusingaFuzzyLogicBasedOptimizationMechanism
261
The algorithm presented in Table 1 describes the
QoS evaluation method applied during the
optimization process, which involves the use of the
fuzzy systems previously described. Once the
execution of a composite service starts, the
adaptation manager constantly evaluates its QoS and
obtains the errors between estimated and measured
values (steps 1 to 12). The measured QoS values are
used as inputs for the first fuzzy system. The benefit
of adaptation is obtained (step 13) and evaluated
(step 14); if it is medium or higher then there is a
need of adaptation. When adaptation is needed, the
system determines the adaptation weights. This
action is performed by the second fuzzy system
(steps 15 to 18). Weights are then adjusted, to fulfil
the restriction
++ += 1 (step 19). Finally,
the algorithm returns the weight values
,and (step 20). These values are sent to
the service binder to be used at the moment of
selecting the next service. When adaptation is not
needed, the service binder ranks the services using
fixed weight values.
4 EVALUATION
To evaluate the proposed optimization approach, an
experimental environment was setup and two
composite services were developed as test cases.
Experiments were carried out to address the
following question:
Does the use of a proactive adaptation
approach based on self-optimization helps
improving the global QoS of a composition?
4.1 Experimental Environment
The experimental environment consists of 4 nodes
configured on a WAN, distributed between United
Kingdom and Germany, with estimated values for
bandwidth and latency around 32Mbit/s and 29ms,
respectively. Node 1 is a computer with Windows
Vista, 4GB RAM and one Intel core2 duo 2.1GHz
processor (located in United Kingdom). This node
hosts the BPEL engine (Apache ODE 1.3.4), service
registry (jUDDI 3.0.4) and historical data base
(MySQL 5.1.51). It is in charge of coordinate the
execution of the compositions and record all the
gathered information. Nodes 2 to 4 are virtual
machines setup on remote servers (located in
Germany), each of the VM’s uses Debian Squeeze
x86 and 1GB RAM. These nodes host one
application server (Tomcat 6.0.35.0) each, which
contains 3 sets of Web services. In total there are 9
Web services deployed per composition’s activity.
The initial values of the QoS parameters were
established based on the node where the service is
running and the corresponding set. Delays are
inserted on some of the service sets, to obtain
different response times, not only based on the
network latency, but the Web services performance.
This information is shown in Table 2.
Values of the cost and energy consumption
change over time, or between services’ executions.
This adds dynamicity to the test environment and
helps obtaining sensible results; also avoiding the
invocation of only one service per each of the
tasks in the composite services. To turn cost into a
dynamic QoS value, the number of times a service
Table 1: QoS evaluation algorithm.
Input:
rt response time
cost cost
ec energy consumption
av availability
eRt response time error
eCost cost error
eEc energy consumption error
eAv availability error
Output:
benefit of adaptation
response time weight
cost weight
energy consumption weight
availability weight
(1) Sort by response time
(2) rt Obtain measured response time
(3) eRt Obtain response time error
(4) Sort by cost
(5) cost Obtain measured cost
(6) eCost Obtain cost error
(7) Sort by energy consumption
(8) ec Obtain measured energy consumption
(9) eEc Obtain energy consumption error
(10) Sort by availability
(11) av Obtain measured availability
(12) eAv Obtain availability error
//fuzzy system 1
(13) Obtain benefit of adaptation
(14) if >= medium then
//fuzzy system 2
(15) Obtain response time weight
(16) Obtain cost weight
(17) Obtain energy consumption weight
(18) Obtain availability weight
(19) Adjust weights
(20) return ,and
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
262
has been invoked within a period of time is evaluated
continuously. Based on this information, it is possible
to establish a new cost based on the demand (number
of times the service is invoked), assuming that higher
the demand, higher the cost. Cost value is updated on
the WSDL file of each service.
Regarding energy consumption, each of the
servers where the Web services are executed is
assumed to have different hardware and software
configurations. Servers information and their
characteristics were selected from the Energy Star
report (Energy Star, 2012). Using the model
proposed in (Buyya et al., 2010), which is based on
the percentage of CPU usage, it is possible to
determine an approximate value to the server energy
consumption.
∙1∙
∙ (13)
(14)
Where is the power consumed in an
instance of time;
is the power consumed when
the server is fully utilized; is the utilization level;
and is the fraction of power consumed by the idle
server. is the total energy consumed by a node over
a period of time .
Servers’ utilization is considered to be variable
over time. The power consumed by a server is
obtained periodically and exposed on the WSDL files
of the corresponding services. It is computed using
equations (13) and (14) and the data presented in
Table 2.
Table 2: QoS parameters initial setup.
Server Set
Time
delays
(ms)
Cost
Energy
Consumption
(W/sec)
A
vailabilit
y
Idle Load
Node 2
S1 0 120
50.75 129.5
0.9
S2 350 80 0.9
S3 200 100 0.9
Node 3
S1 0 150
45.27 81.9
0.64
S2 350 100 0.62
S3 200 120 0.63
Node 4
S1 0 100
210.85 388.3
0.5
S2 350 60 0.46
S3 200 80 0.48
4.2 Experiment Description
Two test cases have been modelled in order to asses
the proposed approach. These models are BPEL
(OASIS, 2007) processes that represent typical
examples for service composition scenarios. Test
case 1 is illustrated in Figure 2(a), it implements an
order booking process that validates the product
availability, obtains the best price of the product from
two different providers, selects the best provider,
performs the payment, and finally completes the
order. Test case 2 implements a travel planning
process, as shown in Figure 2(b). It validates a credit
card, performs flight and hotel reservations in
parallel, and finally invokes a car rental operation.
For matter of simplicity, both diagrams only depict
those activities that involve service invocations.
Per each of the tasks in the processes, there are 9
candidate services, distributed among the servers
(nodes) that fulfil the required functionality and offer
different QoS, giving a total of 45 candidate services
to be used in test case 1 and 36 for test case 2. These
services were previously registered into the service
registry (UDDI), and executed several times to
populate the historical data base and enable the
estimation of their QoS attributes. Both processes are
hosted and invoked from Node1.
In order to evaluate the proposed approach, both
test cases where executed 100 times. These
executions were performed from two different
perspectives: 1) using the proactive optimization
(b)
(a)
Product
availability
Manufacturer
price
Distributor
price
Order
fullfilment
Paypal
payment
Card
payment
Credit card
validation
Hotel
reservation
Flight
reservation
Car rental
Figure 2: Test cases. (a) Order booking process. (b)
Travel planning process.
ProactiveAdaptationinServiceCompositionusingaFuzzyLogicBasedOptimizationMechanism
263
mechanism; 2) using a non-adaptive method. The
experiment was repeated 5 times to assess the
consistency of the results based on statistical
analysis.
4.3 Evaluation Results
The behaviour of the proactive optimization
mechanism was compared with a non-adaptive
approach, where service selection was performed
using fixed weights set to 0.25. Initial results show
improvements in the global QoS values of the
composition when using the proposed approach.
Global QoS refers to the final values of the different
QoS properties (response time, cost, energy
consumption and availability).
The plots shown in Figure 3 depict the behaviour
of the order booking process, showing the mean
values of the different QoS parameters after
performing 5 sets of 100 runs. For the proposed
approach, the values of cost and energy consumption
change over time, as previously described, while for
the non-adaptive approach, remain constant. For both
cases, the value of availability changes according to
the behaviour of the component services.
After analyzing the value of each of the QoS
parameters, in both processes, it was identified that,
in order to improve response time, energy
consumption and availability, there was an increment
in the composition’s cost.
In test case 1, results show that the proposed
approach provides a mean reduction of 2% with a
standard deviation of 6.7% in the measured response
time values. Also, it can be noticed from Figure 3(a),
that it presents a more stable behaviour, without
showing high peaks, as compared to the non-adaptive
approach. This is due to the constant evaluation of
the QoS parameters during execution.
In terms of energy consumption, it is important to
notice that this value is not only based on power
consumption, but also influenced by time. As a
result, a small response time may produce a small
energy consumption value. Figure 3(b) shows the
values corresponding to energy consumption, which
have a similar behaviour to response time, and
provide a mean reduction of 14.7% with a standard
deviation of 18.9%. Results also indicate that there is
a significant improvement in the processes’
availability, presenting a mean increase of 41% with
a standard deviation of 35%. The availability values
corresponding to the order booking process are
illustrated in Figure 3(c). Regarding cost, it can be
1009080706050403020101
1900
1800
1700
1600
1500
1400
1300
1200
1100
Number of executions
Response time (ms)
Proactive
No proactive
1009080706050403020101
480
420
360
300
240
180
120
Number of executions
Energy consumption (W/sec)
Proactive
No proactive
1009080706050403020101
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Number of executions
Availability
Proactive
No proactive
1009080706050403020101
600
550
500
450
400
Number of executions
Cost
Proactive
No proactive
(a) (b)
(
d
)
(
c
)
Figure 3: Order booking process results. (a) Response time. (b) Energy consumption. (c) Availability. (d) Cost.
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
264
noticed from Figure 3(d) that the use of the proposed
approach turns into more expensive composite
services. It shows a mean increase of 11% with a
standard deviation of 8.4%.
Results obtained from test case 2 show a similar
behaviour; where response time, energy consumption
and availability values are improved, while cost
increases. In terms of response time, it shows a mean
reduction of 8.9% with a standard deviation of 16%.
For energy consumption, the obtained mean
reduction is 4.6% with a standard deviation of 29%.
Regarding availability, it provides an improvement of
18% with a standard deviation of 25%. Finally, in
terms of cost, there is an increment of 12.5% with
standard deviation of 6.8%.
Based on the analysis of the weight values
obtained by the optimization model and sent to the
service binder, the parameter that had the higher
impact within the adaptation process was energy
consumption, followed by response time. Because of
this, at the moment of selecting new services to be
invoked, priority would be given to those that are
being executed on servers with lower energy
consumption, and lower response time. Which, based
on the QoS configuration, are the services that also
involve higher costs. Different QoS configurations
may give different results; however, because of the
use of multiple QoS criteria, it is likely to find that
not all the parameters can be improved.
5 RELATED WORK
Several works have been proposed to mitigate the
impacts of unexpected events during the execution of
services, ensuring/maintaining the functional and
quality levels. These approaches can be classified
based on the moment when adaptation takes place
into the categories: reactive and proactive. Reactive
adaptation occurs after the appearance of an
undesired event, while proactive adaptation aims to
predict and prevent the occurrence of the problem
(Aschoff and Zisman, 2012).
Some of the approaches that support reactive
adaptation implement self-* properties. Self-healing
mechanisms aim to prevent composite services from
failing, from functional and non-functional
perspectives. Projects like those presented in
(Canfora et al., 2008, Erradi and Maheshwari, 2008,
Dai et al., 2009, Wu et al., 2009, Ying et al., 2009,
Ardagna et al., 2011, Wenjuan et al., 2010, Baker et
al., 2013) apply self-healing approaches, where new
services are selected and invoked after a functional
failure or a QoS constraint violation. Self-
optimization mechanisms are closely related to the
selection of services at runtime, in order to maintain
the expected QoS of the entire composition.
Examples of works belonging to this category are
described in (Ardagna et al., 2011, Calinescu et al.,
2011, Cardellini et al., 2012).
Approaches that support proactive adaptation are
presented in (Hielscher et al., 2008, Tosi et al., 2009,
Leitner et al., 2010, Aschoff and Zisman, 2012,
Metzger, 2011, Yuelong et al., 2012, Sammodi et al.,
2011). The work presented in (Aschoff and Zisman,
2012) introduces a proactive adaptation approach
that enables service replacement (1-1, 1-n, n-1, n-m)
when it detects situations that may cause the
composition to stop its execution (unavailable or
malfunctioning services); or that allow the
composition to continue its execution, but not in its
best way. Also it considers the emergence of better
services and new requirements. The approach uses a
composition template as start point and selects a set
of candidate services to be used in the composition
and their replacements. The approach introduced in
(Yuelong et al., 2012) combines runtime information
with design-time specifications (of each component
service within a composition), in order to construct a
k-step model of the current service states. The
resulted model can be used to be compared with the
desired behaviour of the composition. The work in
(Leitner et al., 2010) aims to minimize Service Level
Agreement (SLA) violations. It uses predictions of
SLA violations generated with regressions of
monitored and estimated data. These predictions are
evaluated at defined checkpoints. In (Hielscher et
al., 2008), a framework that uses online testing to
trigger proactive adaptation in service-based
applications is described. Test objects can be single
or composite services. While performing online
testing, if an online test fails or deviates from its
expected behaviour, the framework will trigger
adaptation to avoid undesirable consequences. One
of the application scenarios for this approach is
composite services. The work presented in (Tosi et
al., 2009) proposes a self-adaptive mechanism based
on the use of test cases to obtain possible
mismatches between requested and provided
services. When the diagnosis mechanism reveals
mismatches, it triggers adaptation strategies that
update the structure and the behaviour of the client
application to solve the identified problems. Even
though this approach is not mainly focus in service
composition, it presents a proactive mechanism that
works in service-based applications. In (Sammodi et
al., 2011) it is presented a proactive approach for
adaptation in service-based applications that
ProactiveAdaptationinServiceCompositionusingaFuzzyLogicBasedOptimizationMechanism
265
combines monitoring, online testing and quality
prediction. When a service is likely to be used with a
high frequency, it is selected to be tested. The use of
pre-defined test cases (concrete data inputs) enables
the system to collect information about the
behaviour of the services and complement the data
gathered during monitoring. Finally, the work
described in (Metzger, 2011) discusses two main
directions than can be followed in order to perform
proactive adaptation in service oriented systems. The
first direction is to improve the failure predictions
techniques. Some prediction techniques identified by
the authors include data mining, online testing,
runtime verification, statics analysis and simulation.
The second direction is by dynamically estimating
the accuracy of the predicted failures during the
operation of the service-oriented system.
Although these approaches are closely related
with the work described in this paper, there are
significant differences:
The QoS parameters considered in this study
are response time, cost, energy consumption
and availability, while related approaches
mainly focus on response time.
Adaptation is not limited to failure prevention,
but also considers the possibility of
improvement in the QoS levels.
The use of the benefit of adaptation, obtained
from the measured QoS values, to determine
whether adaptation is needed or not.
The use of fuzzy logic as a decision making
tool to determine the need of adaptation in the
context of proactive adaptation in service
composition.
6 CONCLUSION AND FUTURE
WORK
This paper presents a proactive adaptation approach
for service composition that implements a self-
optimization mechanism, which aims to
improve/maintain the quality levels of the composite
services. At runtime, this mechanism performs a
continuous evaluation of the composite services’
behaviour and triggers adaptation when the benefit of
performing this action is considered to be significant.
In order to perform this decision making process, it
uses fuzzy support systems and carries out an
analysis on historical and fresh QoS data.
Evaluation results show that the proposed
mechanism improves the global QoS values of the
compositions, showing significant improvements
regarding response time, energy consumption and
availability. However, it was not possible to improve
all the considered quality parameters, as there was an
increment in the cost of the composite services. This
behaviour is shown due to the relationship between
the values of quality parameters exhibit by the
services, as those services with lower energy
consumption and higher availability, also display
higher costs.
Future work includes the use of Service Level
Agreements on top of the composite services, in
order to provide a contract between the provider and
consumer. Also, it is considered the extension of the
adaptation mechanism, by adding features that enable
service replacements considering different structures
with the forms 1-n, n-1 and n-m. Finally, with the
aim of performing a comparison, it is considered the
implementation and evaluation of mechanisms
proposed in the related work, using the same
composition environment.
REFERENCES
Ardagna, D., L. Baresi, et al., 2011. A Service-Based
Framework for Flexible Business Processes. IEEE
Software 28(2): 61-67.
Ardagna, D. and R. Mirandola, 2010. Per-Flow Optimal
Service Selection for Web Services Based Processes.
Journal of Systems and Software 83(8): 1512-1523.
Aschoff, R. and A. Zisman, 2012. Proactive Adaptation of
Service Composition. ICSE Workshop on Software
Engineering for Adaptive and Self-Managing Systems
(SEAMS'12). Zürich, Switzerland: 1-10.
Baker, T., O. F. Rana, et al., 2013. Towards Autonomic
Cloud Services Engineering via Intention Workflow
Model. Proceedings of the 10th International
Conference on Economics of Grids, Clouds, Systems,
and Services (GECON'13), Zaragoza, Spain.
Buyya, R., A. Beloglazov, et al., 2010. Energy-Efficient
Management of Data Center Resources for Cloud
Computing: A Vision, Architectural Elements, and
Open Challenges. Proceedings of the 2010
International Conference on Parallel and Distributed
Processing Techniques and Applications, las Vegas,
USA.
Calinescu, R., L. Grunske, et al., 2011. Dynamic QoS
Management and Optimization in Service-Based
Systems. IEEE Transactions on Software Engineering
37(3): 387-409.
Canfora, G., M. D. Penta, et al., 2008. A Framework for
QoS-Aware Binding and Re-Binding of Composite
Web Services. Journal of Systems and Software
81(10): 1754-1769.
Cardellini, V., E. Casalicchio, et al., 2012. MOSES: A
Framework for QoS Driven Runtime Adaptation of
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
266
Service-Oriented Systems. IEEE Transactions on
Software Engineering 38(5): 1138-1159.
Cardoso, J., A. Sheth, et al., 2004. Quality of Service for
Workflows and Web Service Processes. Journal of
Web Semantics 1(3): 281-308.
Châtel, P., J. Malenfant, et al., 2010. QoS-based Late-
Binding of Service Invocations in Adaptive Business
Processes. Proceedings of the International
Conference on Web Services (ICWS'10), Miami, USA,
IEEE Computer Society.
Cheng, B., R. Lemos, et al., 2009. Software Engineering
for Self-Adaptive Systems: A Research Roadmap.
Software Engineering for Self-Adaptive Systems 5525:
1-26
Dai, Y., L. Yang, et al., 2009. QoS-Driven Self-Healing
Web Service Composition Based on Performance
Prediction. Journal of Computer Science and
Technology 24(2): 250-261.
Dustdar, S., C. Dorn, et al., 2010. A Roadmap Towards
Sustainable Self-Aware Service Systems. Proceedings
of the 2010 ICSE Workshop on Software Engineering
for Adaptive and Self-Managing Systems, Cape Town,
South Africa, ACM.
Dustdar, S. and W. Schreiner, 2005. A Survey on Web
Services Composition. International Journal on Web
and Grid Services 1(1): 1–30.
Energy Star, 2012. Computer Servers Product List -
Families.
Erradi, A. and P. Maheshwari, 2008. Dynamic Binding
Framework for Adaptive Web Services. Proceedings
of the 2008 Third International Conference on Internet
and Web Applications and Services, Athens, Greece,
IEEE Computer Society.
Hielscher, J., R. Kazhamiakin, et al., 2008. A Framework
for Proactive Self-adaptation of Service-Based
Applications Based on Online Testing. Proceedings of
the 1st European Conference Service Wave, Madrid,
Spain, Springer-Verlag.
Huang, A., C.-W. Lan, et al., 2009. An Optimal QoS-
based Web Service Selection Scheme. Information
Sciences 179(19): 3309-3322.
Hwang, S.-Y., H. Wang, et al., 2007. A Probabilistic
Approach to Modeling and Estimating the QoS of
Web-Services-Based Workflows. International
Journal of Information Sciences 177(23): 5484-5503.
Kaplan, J., W. Forrest, et al., 2009. Revolutionizing Data
Center Energy Efficiency, McKinsey.
Leitner, P., A. Michlmayr, et al., 2010. Monitoring,
Prediction and Prevention of SLA Violations in
Composite Services. Proceedings of the IEEE
International Conference on Web Services (ICWS'10),
Miami, USA.
Li-Xin Wang, 1997. A Course in Fuzzy Systems and
Control, Prentice Hall.
Metzger, A., 2011. Towards Accurate Failure Prediction
for the Proactive Adaptation of Service-Oriented
Systems. Proceedings of the 8th Workshop on
Assurances for Self-adaptive Systems, Szeged,
Hungary, ACM.
OASIS, 2007. Web Services Business Process Execution
Language Version 2.0. Retrieved Dec. 2013, from
http://docs.oasis-open.org/wsbpel/2.0/OS/wsbpel-
v2.0-OS.html.
Sammodi, O., A. Metzger, et al., 2011. Usage-Based
Online Testing for Proactive Adaptation of Service-
Based Applications. Proceedings of the IEEE 35th
Annual Computer Software and Applications
Conference (COMPSAC'11), Munich, Germany.
Tosi, D., G. Denaro, et al., 2009. Towards Autonomic
Service-Oriented Applications. International Journal
of Autonomic Computing 1(1): 58-80.
W3C Working Group, 2003. QoS for Web Services:
Requirements and Possible Approaches. Retrieved
Dec. 2013, from http://www.w3c.or.kr/kr-
office/TR/2003/ws-qos/.
Wenjuan, L., Z. Qingtian, et al., 2010. A Framework to
Improve Adaptability in Web Service Composition.
Proceedings of the 2nd International Conference on
Computer Engineering and Technology (ICCET'10),
Chengdu, China.
Wu, G., J. Wei, et al., 2009. Towards Self-Healing Web
Services Composition. Proceedings of the First Asia-
Pacific Symposium on Internetware, Beijing, China,
ACM.
Ying, Y., Z. Bin, et al., 2009. A Self-Healing Composite
Web Service Model. Proceedings of the IEEE Asia-
Pacific Services Computing Conference (APSCC'09),
Biopolis, Singapore.
Yuelong, Z., W. Xiaobin, et al., 2012. Predicting Failures
in Dynamic Composite Services with Proactive
Monitoring Technique. Proceedings of the IEEE
Eighth World Congress on Services (SERVICES'12),
Honolulu, USA.
Zadeh, L. A., 1965. Fuzzy Sets. Information and Control
8(3): 338-353.
Zadeh, L. A., 1994. The Role of Fuzzy Logic in Modeling,
Identification and Control. Modeling, Identification
and Control 15(3): 191-203.
Zeginis, C. and D. Plexousakis, 2010. Web Service
Adaptation: State of the Art and Research Challenges.
Heraklion, Crete, Greece, Institute of Computer
Science.
Zeng, L., B. Benatallah, et al., 2004. QoS-Aware
Middleware for Web Services Composition. IEEE
Transactions on Software Engineering 30(5): 311-327.
ProactiveAdaptationinServiceCompositionusingaFuzzyLogicBasedOptimizationMechanism
267
