Increasing Energy Saving with Service-based Process Adaptation

Alessandro Miracca and Pierluigi Plebani

Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria,

Piazza Leonardo da Vinci 32, 20133 Milano, Italy

Keywords:

Complex Event Processing, Data Prediction, Proactive Adaptation.

Abstract:

The aim to reduce the energy consumption in data centres is usually analyzed in the literature from a facility

and hardware standpoint. For instance, innovative cooling systems and less power hungry CPUs have been

developed to save as much more energy as possible. The goal of this paper is to move the standpoint to the

application level by proposing an approach, driven by a goal-based model and a Complex Event Processing

(CEP) engine, that enables the adaptation of the business processes execution. As several adaptation strategies

can be available to reduce the energy consumption, the selection of the most suitable adaptation strategy is

often the most critical step as it should be done timely and correctly: adaptation has to occur as soon as a

critical point is reached (i.e., reactive approach) or, even before it occurs (i.e., proactive approach). Finally,

the adaptation actions must also consider the inﬂuence on the performance of the system that should not be

violated.

1 INTRODUCTION

The adoption of the PUE (Power Usage Effective-

ness) to measure how much energy a data centre con-

sumes to run facilities instead of the IT equipments,

has affected the methods proposed in the literature.

Indeed, most of the proposed work are focused on the

reduction of the power consumed by the cooling sys-

tems, or by the optimization of the IT equipments.

On the contrary, less attention has been paid to the

optimization of the applications running in the data

centres. With this work, we want to focus on this as-

pect leveraging on the an adaptive system that detect

anomalies in energy consumption and react by mov-

ing the application to a greener state.

The design and implementation of systems able to

adapt their own behavior with respect to new and un-

foreseen requirements, or to deal and solve problems

that involve the underlying hardware and software,

are some of the most challenging and interesting re-

search questions in the computer science nowadays.

As also discussed in the Autonomic Computing man-

ifesto (Kephart and Chess, 2003), only with a deep

knowledge of the system that we would like to adapt,

the adaptation is possible. This knowledge concerns

the components of the system and how they are orga-

nized, along with the values of indicators that are able

to capture the performance of the system.

According to the MAPE (Monitor - Analyze - Plan

- Execute) cycle, a monitoring infrastructure that con-

tinuously collects information about the status of the

system is required. The analysis of the gathered data

leads the system to better identify any malfunction-

ings and which are the possible adaptations to elimi-

nate them. Finally, the most suitable adaptation action

has to be selected and enacted. In case of service-

based processes, KPIs (Key Performance Indicators)

can be used as a way to evaluate a process in terms of

a set of metrics capturing the most relevant aspects of

the process performance. Among these metrics, en-

ergy saving is gaining more importance. For this rea-

son, we introduce the concept of GPIs (Green Perfor-

mance Indicators) as a way to monitor the greenness

of the process.

The monitoring system should be able to compute

these indicators through the data previously collected.

Actual problems are how their current values can be

used to assess the correctness of the process execu-

tion, how it is possible to react to a critical situation

and, ﬁnally, if it is also possible to anticipate the adap-

tation predicting the next values of KPIs and GPIs.

This paper presents a methodology for energy ef-

ﬁciency in a service-based process context, with the

aim of reducing the impact of the process execution as

much as possible via a set of techniques combined to-

gether. The proposed approach is composed of three

main elements: (i) a goal-based model used to spec-

ify which are the KPIs and GPIs that must be fulﬁlled

201

Miracca A. and Plebani P..

Increasing Energy Saving with Service-based Process Adaptation.

DOI: 10.5220/0004379802010209

In Proceedings of the 2nd International Conference on Smart Grids and Green IT Systems (SMARTGREENS-2013), pages 201-209

ISBN: 978-989-8565-55-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Server4&

Server2&

Server3&

Server1&

Browse

products

Receive

order

information

Review order

(optional)

Receive

payment

Cancel

order

Invoke

delivery

Figure 1: e-Commerce sample process.

and, if this is not possible, the actions that might be

performed to adapt the execution; (ii) a prediction sys-

tem used to anticipate possible violations of the indi-

cators; and (iii) an architecture that includes a CEP

(Complex Event Processing) engine (Luckham, 2001)

and combines it with the other elements in order to

detect both the occurring and the predicted violations

and to produce a ranked list of the adaptation actions

that could solve the problem.

Our approach has been validated for a service-

based process running on a testbed where a moni-

toring system exists. The validation shows how the

system is able to detect anomalies in terms of energy

consumption. Moreover, the system is able to support

the selection of the proper adaptation action that will

reduce the energy consumption of the running process

without affecting the performance parameters, as the

response time. Finally, the proposed approach is also

useful for collecting information to further improve

improve the initial goal-based diagram.

The paper is organized as follows. Section 2 in-

troduces a running example used along the paper to

better describe our approach. Section 3 describes the

details of the approach and, in particular, the adopted

goal-based model and the data prediction techniques.

The CEP-based architecture, that combines these ele-

ments, is described in Section 4. A validation of the

approach is the goal of Section 5. After a compari-

son of the work with the state of the art in Section 6,

the paper concludes with a discussion about possible

future extensions in Section 7.

2 RUNNING EXAMPLE

To better understand and provide a validation of our

approach, we use an e-commerce example, inspired

by the TPC-C (Raab et al., 2001) benchmark, where

services are invoked during the execution of a process

for buying a good. Figure 1 shows the structure of the

process and a possible deployment of the services on

a data center (we assume that the orchestrator is run-

ning on another server, i.e., Server0 not included in

the ﬁgure). It is worth noting, as it will be also dis-

cussed during the validation, that an alternative “cen-

tralized” deployment is possible: in such a scenario

all the services are executed on the Server1.

For this case study, Table 1 reports some of the

relevant indicators and the metrics to calculate them.

These ones concern both with typical performance di-

mensions (e.g., response time) and with green-aware

indicators (e.g., application green performance). To

compute these indicators, we assume that a proper

monitoring infrastructure is available.

The goal of our approach is to monitor these pre-

deﬁned KPIs and GPIs and to properly adapt the ap-

plication deployment in order to satisfy all the goals

agreed with the users. The achievement of this ob-

jective requires the mediation between two perspec-

tives that often clash. On the one side, the satisfaction

of performance indicators, as the response time, de-

mands more and more resources. On the other side,

the increasing of resources usage raises the energy

consumed; so that the green-aware indicators could

be not satisﬁed.

3 PROPOSED APPROACH

3.1 Goals and Adaptation

The goal-based model proposed by (Asnar et al.,

2011) has been adapted to represent what we need

to monitor, the occurrences that could affect our ob-

servations, and to exploit the characteristics of adap-

tivity of the process. This model encompasses three

conceptual structures that are organized in three cor-

responding layers, named respectively Asset, Event

and Treatment layer:

• Goals (depicted as ovals) are strategic interests

that actors intend to achieve for generating values;

• Events (depicted as pentagons) are uncertain cir-

cumstances, typically out of the control of actors,

which can have an impact, positive or negative, on

the fulﬁllment of goals;

• Treatments (depicted as hexagons) are sequences

of actions used to achieve goals or to treat events.

Figure 2 shows an application of the model for our

running case study: (i) Goals represent a set of de-

SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems

202

CPU

temperature

[60,70]

Response

time [<5]

Application

green

performance

[>700]

CPU usage

 [80,90]

Process

power

consumption

[<65]

Increase

CPU

temperature

Reduce

Increase

CPU payload

Increase

server power

consumption

Increase

no. servers

involved in

the process

Reduce

Increase

Dynamic CPU

frequency

scaling

Asset Layer

Event Layer

Treatment Layer

Figure 2: Goal-based model.

Table 1: KPIs and GPIs calculation.

Indicator Threshold Metric

CPU temperature ∈ [60,70] temperature(

◦

CPU usage ∈ [80,90] CPUusage (%)

Process power consumption < 65

∑

server

(power

(W ))

Application green performance > 700 #transactions/kW h

Response time < 5

∑

(task responsetime(s))

ﬁned constraints on the KPIs and GPIs of the system;

(ii) Events embody possible situations that could af-

fect the integrity of KPIs or GPIs constraints and any

other context that requires the enactment of an adap-

tation action; ﬁnally, (iii) Treatments symbolize the

available repair actions.

An added value of the adopted goal-based model

is the possibility to deﬁne relationships among goals.

This means that an indicator may depend on other cor-

related indicators. In order to react to the violation of

an indicator, it is possible to take actions inﬂuencing

the correlated indicators and, consequently, indirectly

improve the considered indicator. Indicators’ thresh-

olds have to be deﬁned at accurately. It is also possi-

ble to describe how events affect the goals.

It is worth noting that, in order to ensure a pro-

active adaptation, the deﬁned events are determined in

terms of value trends that can have positive or nega-

tive impacts on the goal. A non-trivial task is to ﬁgure

out which trends are occurring and when the values

are considered so critical to require a treatment. The

goal of the prediction, as described in the next sec-

tion, is twofold. On the one side, the time series anal-

ysis (a necessary preliminary stage for the prediction

itself) is able to identify and to model the trends of

the monitored data. On the other side, the prediction

is capable of anticipating the achievement of critical

values, so that an adaptation action can be activated

before the indicator violation occurs.

3.2 Data Prediction

In general, a series is deﬁned as a sequence of several

observations of a phenomenon with respect to a quali-

tative nature. If this feature is time, the series is called

historical or, commonly, time series.

Given the phenomenon Y , any observations col-

lected at time t is represented by the notation Y

. So,

formally, Y = [Y

: 1 ≤ t ≤ L] where L is the total num-

ber of observations gathered.

The classical approach for time series analysis

provides a model that describes any observation as:

= f (t) + u

in which the value of the observation at time t is

the result of the composition between a determinis-

tic component f (t), called systematic part, and a se-

quence of random variables u

, named stochastic part.

It is worth noting that, before any kind of analy-

sis, time series data need to be examined in order to

make some adjustments and replace outliers. These

adjustments are necessary when discontinuity and ef-

fects due to the different duration of time intervals af-

fect the series. Figure 3 shows the effect of an outlier

IncreasingEnergySavingwithService-basedProcessAdaptation

203

Figure 3: Outlier elimination samples.

elimination procedure on a generic time series.

In this paper, time series analysis is used to pre-

dict the values of a KPI, or a GPI. That means using

a model to anticipate the future behavior of the ob-

served variable, based on its past values. Indeed, the

sooner we can ﬁgure out a possible violation of an

indicator, the sooner we can react and possibly avoid

that such a violation occurs. In particular, considering

our running example, we noticed that several outliers

arise. As a consequence, and in order to obtain a reli-

able prediction, such outliers need to be deleted.

Models for time series can have many forms and

represent different stochastic processes. AutoRegres-

sive (AR) models, Integrated (I) models and Moving

Average (MA) models are some of the most popular

way to theoretically deﬁne a time series. Combina-

tions of them produce well known AutoRegressive

Moving Average (ARMA) and AutoRegressive Inte-

grated Moving Average (ARIMA) models.

In our case, we use autoregressive model to repre-

sent time series data. The notation AR(p) indicates a

p-order autoregressive model which is deﬁned as:

= φ

t−1

+ φ

t−2

+ φ

t−p

+ η

where φ

,φ

,...,φ

are the parameters of the

model which act for the linear regression coefﬁcient

of the random variable y

, compared to its past values,

and η

is white noise.

The most common manner to identify a model is

through its transfer function W (z). Called u(t) the

function that describes the input values in terms of

time and y(t) the function that represents the output,

knowing that any autoregressive and moving average

model provides an output that is a linear combination

of previous values of both the input and the output,

y(t) is obtained by (1):

y(t) + α

y(t − 1) + ... + α

y(t − p) = β

u(t) + 1

+β

u(t − 1) + ... + β

u(t − q) (1)

Applying the Discrete Fourier Transformation, (1)

can be rewritten introducing lag operator z, which has

the aim to delay or to anticipate a value:

y(t) =

+ z

−1

+ ... + z

−q

1 + z

−1

+ ... + z

−p

· u(t) =

C(z)

A(z)

· u(t) = W (z) · u(t) (2)

W (z) is called transfer function and, in order to

compute a reliable prediction, we assume it is already

in the canonical spectral factor form, as required by

the time series analysis theory.

Starting from (2), we can get the one-step ahead

predictor transfer function

ˆy(t + 1 | t) =

z · (C(z) − A(z))

C(z)

· y(t) (3)

So, in order to estimate the future value of the

series, we need to calculate the difference between

the numerator C(z) and the denominator A(z) of the

autoregressive model transfer function, multiply this

with the lag operator z and then divided by C(z).

Regarding multi-step ahead prediction, the r-step-

ahead predictor transfer function

(z) is the result

of the method of Polynomial Long Division between

C(z) and A(z), applied r times (Cheng et al., 2006).

The prediction accuracy always depends on the

value of an error indicator. In case of ARMA(p,q)

model, the Final Prediction Error (FPE) is used as a

statistical measure of goodness-of-ﬁt. The same indi-

cator can be also used in case of an AR(p) model, as

any autoregressive model is equal to an ARMA model

where q = 0. The Final Prediction Error (FPE) is

given by:

FPE

(L + (p + 1))

(L − (p + 1))

· σ

where σ

is the variance of model residuals, L is

the length of the time series, and p is the number of

estimated parameters in the model. The variance of

model residuals is calculated iteratively by:

= σ

p−1

(1 − (

· p)

)

in which

= 0 and σ

∑

t=1

−Y )

is the

variance of the time series.

Several candidate models can be calculated using

different criteria, but only the one with the lowest FPE

value is selected as the best-ﬁt model.

Due to the limited number of pages, in this paper we

omitted the complete procedure to obtain the predictor. For

more details, we suggest to read (Hamilton, 1994)

SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems

204

Service-based process adaptor

Data monitor

Complex Event

Processing

Engine

Data predictor

Adaptation actions

ranker

Goal-based model

instances

repository

Ranked list of

adaptation

actions

Figure 4: System architecture.

4 ARCHITECTURE

After having introduced the goal-based model and the

data prediction in time series, as the cornerstones of

our approach, the Figure 4 shows the architecture that

combines them to support the service-based adapta-

tion. Here, we assume that a monitoring infrastruc-

ture, composed by several probes, is properly de-

ployed to gather information about the execution of

the service-based process. As described in the follow-

ing, the goal-based model instances repository con-

tains information that is useful to understand all the

modules composing our architecture. Indeed, it con-

tains information regarding which are the constraints

that must be satisﬁed (i.e., the goals) and the avail-

able adaptation actions (i.e., the treatments), along

with information with which instrument the CEP en-

gine to detect the violations (i.e., the events). Finally,

the arrows reported in the ﬁgure reﬂects the data ﬂow

among the modules.

4.1 Data Monitor

Goal of the Data Monitor is to collect all the data

made available by the probes and to compute the indi-

cators used to deﬁne the goals to be fulﬁlled. For this

reason, this module reads the goal-based model to ob-

tain the list of the indicators that have to be calculated

and the metrics to be used (see Table 1). In some

cases, the computation directly reﬂects a monitored

value (e.g., power consumption is equal to the value

returned by a probe installed on a server). In some

other cases, the computation is more complex and in-

volve several probes (e.g., the MHz/Watt). Data col-

lected by the Data Monitor are made available for the

CEP module and the Data Predictor. In our testbed,

a set of Nagios

plugins have been implemented to

collect the required data.

Monitoring system usually has an impact on the

overall performances. As a consequence, the setting

of the monitoring frequency is a critical step. Hav-

ing a high monitoring frequency means greater over-

head and a lot of data to manage. On the contrary,

low monitoring frequency means less impact on the

performances but an higher probability to miss sig-

niﬁcant variations and also violations.

4.2 Data Predictor

For all the monitored indicators, the goal of this mod-

ule is to identify the time series model and to predict

future values starting from the data monitored up to a

given time. The theoretical background of this mod-

ule has been discussed in Section 3.2. Considering the

monitoring frequency, the prediction can have a sig-

niﬁcant impact for the adaptation as it makes possible

to anticipate the violation of several minutes. This

module relies on GNU Octave Engine

and, in partic-

ular, on the following functions:

• arburg - it calculates the coefﬁcients of an autore-

gressive (AR) model of complex data using the

whitening lattice-ﬁlter method of Burg;

• detrend - it removes the trend from data, handles

NaN’s by assuming that these are missing values

and unequally spaced data;

• interp1 - it helps in replacing the outlier with val-

ues that do not alter the characteristics of the se-

ries, through the interpolation method.

http://www.nagios.org/

http://www.gnu.org/software/octave

IncreasingEnergySavingwithService-basedProcessAdaptation

205

In order to obtain an accurate prediction, the function

initially performs a deep analysis of the series, cal-

culating the mean and the variance ﬁrst of all, just be-

fore moving to the elimination of outliers and possible

trend components, cyclicity or seasonality. After that,

the series is rendered with zero mean value, in order to

decrease the prediction error. Now the series is ready

for the calculation of the auto-covariance, a key ele-

ment for the computation of the FPE, and then, ﬁnally,

predictions for each model and possible criterion are

estimated. At this point, among all the predictions

calculated, the function will select the one that corre-

sponds to the criterion passed as an input parameter

and, at the same time, with the least value of FPE.

4.3 CEP Engine

Both the real monitored data and the predicted ones

are inputs for the CEP engine

, as the goal of this

module is to monitor all the events related to changes

in current, and also in future, values of the indicators.

Each event is included in the initial goal-based model

and might have positive or negative impact on the ful-

ﬁllment of a given goal.

The adoption of a CEP engine is required as the

information continuously ﬂows from the Data Moni-

tor and the Data Predictor. A query in a CEP, indeed,

once it is issued it logically continuously runs over the

incoming data, in contrast to traditional DBMS where

the queries are run once over the current data set. Usu-

ally, queries for a CEP are based on Event Processing

Language (EPL): an SQL-like language extended to

handle event streams and patterns. Sample queries

are reported in the following:

// Current Response time

SELECT * FROM RT(responseTime >5)

// Predicted CPU Usage

SELECT * FROM CPUUsagePrediction

(probability >60.0,CPUUsage <80) OR

CPUPrediction (probability >60.0,

CPUUsage >90)

Thus, the CEP has to verify if one of the events de-

ﬁned in the model is either occurring or, using the pre-

dicted data, will occur in the future. Since the occur-

rence of an event corresponds to a violation, the CEP

will inform the Adaptation actions ranker about the

critical situation to identify the most suitable adapta-

tion strategy. Considering both the current and future

data allows our system to support both reactive and

pro-active adaptations.

In our implementation we adopted Esper (http:// es-

per.codehaus.org/) as CEP engine

4.4 Adaptation Actions Ranker

Once a notiﬁcation for a violation arrives, this module

suggests to the user the best repairing action, among

those implemented, to resolve the current or immi-

nent violation. As a consequence, this module is also

in charge of deciding if an adaptation should be en-

acted to solve the detected problem or not. Indeed,

it depends on the severity of the violation and on the

nature of the analyzed time series.

For time series with a signiﬁcant stochastic com-

ponent, any time-isolated violation is considered as

an outlier, i.e., a value that is not logical to expect fre-

quently. For this reason, only if several violations for

a given indicator occur in a limited period of time, the

module begins its adaptation selection procedure. In

case of critical indicators, or time series with a limited

stochastic component, the strategy selection will start

as soon as the violation is detected.

Starting from the goal-based model, the Adapta-

tion actions ranker looks for the treatments that can

have a positive impact on the goal under violation.

As different degree of impact can be deﬁned in the

model, this module sorts the suggested actions ac-

cording to a rating that is calculated by the difference

between the number of positive and negative impacts

a repairing action has on the system. These impacts

are determined as relationships that each treatment

has with the other elements of the goal-based model:

the ones that introduce an alleviation of the effect an

event has on the fulﬁllment of a goal is considered a

positive impact, whilst the ones that promote a viola-

tion have negative impact.

The repairing actions suggested by the tool always

have a score equal to zero if they do not modify the

situation, or greater than zero if they improve the cur-

rent condition of the system.

We assume that the ﬁnal choice to activate or not

a repairing action always depends on the user. Such

a request for a human being interaction is intended

to reduce false positives problem. According to the

results of the simulations, false positives represent

about 5% of a sample of 300 measurements. We be-

lieved that a human control, together with that of the

machine, may further reduce this percentage.

5 VALIDATION

With this validation we verify the ability of the system

to detect a violation and to produce the list of avail-

able adaptation actions. This requires the correctness

of the prediction offered by the Data predictor and the

ability of the CEP Engine to capture the events of in-

SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems

206

(a) Response time (b) Process power consumption

Figure 5: Centralized deployment.

(a) Response time (b) Process power consumption

Figure 6: Distributed deployment.

terest. The adopted testbed is composed by 4 servers

where Glassﬁsh

is installed. The services compos-

ing the process introduced in Section 2 can run on any

of these servers. A ﬁfth server hosts the application

that orchestrates the process: i.e., calling the services

in the proper order (see Figure 1). A set of Nagios plu-

gins are installed to be able to gather all the data that

make possible the computation of the KPIs and GPIs,

as deﬁned in Table 1. In particular, we focused on:

(i) the response time, as a typical KPI, calculated at

the end of each invocation; and (ii) the process power

consumption, representing GPIs class sampled every

180 sec.

As a ﬁrst trial, we assume that the services com-

posing the process are running only on Server1 (i.e.,

centralized deployment). With such a conﬁguration,

the experiment lasts one hour, during which a new

request to the process is sent every 1.0 sec in the av-

erage. Figure 5 shows the curves obtained for the re-

sponse time and the process power consumption. The

power consumption is computed summing the power

consumption of all the ﬁve servers even if some of

them are not used in that moment for the execution of

the process. Indeed, we assume that these servers con-

sume power as they remain in stand-by mode, ready

to run applications. Given these curves, proving the

correctness of a prediction means that we need to en-

http://glassﬁsh.java.net/

Figure 7: Process power saving.

sure that the distance between the predicted event at

time t − 1 and the real occurrence of the event at time

t is minimum, in term of numerical value. Such a

property is strongly linked to the goodness-of-ﬁt of

the model that describes the time series of the indi-

cator. On these bases, we can notice how the pre-

diction curves, after an initial set-up time, is able to

follow the same trend of the actual ones. Especially

in case of the power consumption this means that we

are able to know the power consumption, in case of

one-step prediction, three minutes before sensing the

real value. This also means that we can anticipate the

violation and having a proactive adaptation.

Considering the process power consumption, the

CEP module informs about the upcoming violation

the Adaptation actions ranker every time the value is

beyond the threshold (i.e., the horizontal line). As

IncreasingEnergySavingwithService-basedProcessAdaptation

207

a consequence, after ﬁve to ten samples (i.e, ﬁfteen

to thirty minutes), the list of adaptation actions will

be produced as the process power consumption sys-

tematically overcomes the limit value. This requires

the goal-based model instance (see Figure 2), that is

used by the Adaptation actions ranker, to realize that

the violation of the constraint on this indicator can be

solved in two ways: either redeﬁning the deployment

with a higher degree of distribution or reducing the

CPU scaling frequency. These two actions compose

the list that represents the result of our approach.

Figure 6 shows the effects of the process workload

split among the various servers installed and running

on the testbed. Despite a slight increase of the average

response time, the beneﬁts introduced with regard to

the process power consumption are evident.

Figure 7 highlights the power saving that can

be achieved applying one of the suggested adapta-

tion strategies, such as a redistribution of the ser-

vices composing the process among the servers. The

dashed line represents the power consumption if no

adaptation strategy was enabled after any detected vi-

olation; instead, the solid line depicts a situation in

which an adaptation action was performed in order to

solve the encountered problems and, at the same time,

to reduce the energy consumption. An estimation of

the value of such a power saving is around 0.085 kWh,

equal to 6.75% of total energy consumption.

6 RELATED WORK

Focusing on the use of prediction in service-based

systems, (Metzger et al., 2012) tries to work with

online quality prediction. However, there are many

challenging issues, as the limited control and visibil-

ity on the third-party services that call for concrete

solutions. In our approach, instead, prediction is cal-

culated based on KPIs and GPIs, that represent the

quality of the system, both in terms of performance

and for what concerns greenness. We deﬁned these

indicators and, with them, we are able to access ev-

ery signiﬁcant aspect of the system that we want to

monitor. (Wetzstein et al., 2012) provides a detailed

integrated monitoring, prediction and adaptation ap-

proach for preventing KPI violations of business pro-

cess instances. Here, KPIs are used to map the pro-

cess features and prediction is calculated starting from

checkpoint, dislocated through the process itself, via

a classiﬁcation learning algorithm based on decision

tree, called J48. This is the most closely related work

to our approach. The differences stay in how to calcu-

late the prediction and in the selection of the adapta-

tion actions. We use an AR model to obtain a reliable

prediction, while (Wetzstein et al., 2012) opts for de-

cision trees, that means it needs a signiﬁcant initial-

ization step in order to built meaningful trees. Even

the selection of the repairing action is computation-

ally much more expensive of our: in (Wetzstein et al.,

2012) there are no predeﬁned actions, but only indi-

vidual adaptation strategies that have to be combined

to solve the ongoing or future violation. Instead, our

tool relies on a goal-based model that contains all the

possible actions that can solve the detected problem.

In (Zeng et al., 2008) the main concern is Qual-

ity of Service Management (QoSM), which is a new

task in IT-enabled enterprises that supports monitor-

ing, collecting and predicting QoS data. (Zeng et al.,

2008) presents an event-driven QoS prediction system

that contains a real-time metric and KPIs prediction

mechanism. The differences with our work stand in

the way of deﬁning the KPIs and in the instruments

they used to examine the events and to estimate the

predictions. KPIs play a marginal role in (Zeng et al.,

2008), as they are intended only as the result of the

aggregation of some metrics, while they are on the

basis of the whole architecture in this paper. We used

Complex Event Processing and AR models to deal re-

spectively with events and prediction; on the contrary,

Event-Condiction-Action (ECA) rules and Exponen-

tial Smoothing are used by (Zeng et al., 2008) to pro-

cess the events and to compute the predictions.

Artiﬁcial neural networks theory holds all the

elements to be elected as the foremost prediction

method. It resolves the problem of the identiﬁcation

of the model easily in respect of both ARMA struc-

tures and Box-Jenkins methodology. However, we

preferred an autoregressive model because the accu-

racy of a neural network model may be seriously com-

promised when it is used recursively for multi-step

prediction purposes. (Nguyen and Chan, 2004) and

(Adya and Collopy, 1998) are both trying to inves-

tigate a hybrid methodology that combines artiﬁcial

neural networks and ARMA models.

In the area of process monitoring and adaptation,

using CEP techniques, several approaches have been

proposed. CEVICHE (Hermosillo et al., 2010) is a

framework that combines the strength of Complex

Event Processing, a dynamic business process adap-

tation method via an AOP (Aspect-Oriented Program-

ming) paradigm, which complements the more used

object-oriented one, and an extension of the BPEL

language for communication. It differs from our work

just because is only able to activate a single adapta-

tion action, that consists in skipping some optional

services, and also it has only a reactive behavior.

(Sen, 2008) discusses how the existing state-of-the-art

BAM solutions, of which CEP is an essential compo-

SMARTGREENS2013-2ndInternationalConferenceonSmartGridsandGreenITSystems

208

nent, have not come up to the expectations of provid-

ing real-time information across business processes

and supporting business users decision. Our work,

on the contrary, is not limited to monitoring and vi-

sualization of events, but rather deals with analyz-

ing and providing appropriate decision support for

the users. (Leitner et al., 2010) proposes a frame-

work, named PREvent, which is a system that in-

tegrates event-based monitoring via CEP, prediction

of Service Level Agreement (SLA) violations, using

machine learning techiniques, and automated runtime

prevention of those violations by triggering adapta-

tion actions. It differs from our work, ﬁrst of all, in

some aspects related to the prediction: in fact, it com-

putes a prediction only when the esecution of the pro-

cess reaches a checkpoint and the prediction targets

are Service Level Objectives (SLOs), not indicators.

Secondly, the predicted value is calculated by mul-

tilayer perceptrons, a variant of artiﬁcial neural net-

work, while our approach relies on AR models.

7 CONCLUSIONS AND FUTURE

WORK

This paper has introduced an approach for support-

ing the selection of adaptation actions in case of

service-based processes to reduce the energy con-

sumption. This approach combines the use of a con-

ceptual model for deﬁning the relationships between

the system goals, in terms of KPIs and GPIs, and the

adaptation actions, with a prediction system. The re-

sulting architecture is able to support the reactive and

proactive adaptation of a service-based process.

Future extensions of the proposed approach will

involve the use of a n-step predictor to improve the

proactiveness of the system. At the same time, to

close the loop, an approach to automatically veriﬁes

the positive or negative effects of the selected adapta-

tion actions is required.

ACKNOWLEDGEMENTS

This work has been partially funded by Italian project

“SeNSori” (Industria 2015 - Bando Nuove Tecnolo-

gie per il Made in Italy) - Grant agreeement n.

00029MI01/2011.

REFERENCES

Adya, M. and Collopy, F. (1998). How Effective are Neural

Networks at Forecasting and Prediction? A Review

and Evaluation. J. of Forecasting, 17(5-6):481–495.

Asnar, Y., Giorgini, P., and Mylopoulos, J. (2011). Goal-

driven risk assessment in requirements engineering.

Requir. Eng., 16(2):101–116.

Cheng, H., Tan, P.-N., Gao, J., and Scripps, J. (2006).

Multistep-Ahead time series prediction. In Proc. of

the 10th Paciﬁc-Asia conf on Advances in Knowledge

Discovery and Data Mining, PAKDD’06, pages 765–

774, Berlin, Heidelberg. Springer-Verlag.

Hamilton, J. D. (1994). Time Series Analysis. Princeton

University Press.

Hermosillo, G., Seinturier, L., and Duchien, L. (2010). Us-

ing Complex Event Processing for Dynamic Business

Process Adaptation. In Proc. of the 7th IEEE 2010

Int’l Conf. on Services Computing, SCC ’10, pages

466–473, Washington, DC, USA. IEEE.

Kephart, J. and Chess, D. (2003). The Vision of Autonomic

Computing. Computer, 36(1):41–50.

Leitner, P., Michlmayr, A., Rosenberg, F., and Dustdar, S.

(2010). Monitoring, Prediction and Prevention of SLA

Violations in Composite Services. 2012 IEEE 19th

Int’l Conference on Web Services, pages 369–376.

Luckham, D. C. (2001). The Power of Events: An Intro-

duction to Complex Event Processing in Distributed

Enterprise Systems. Addison-Wesley Longman Pub-

lishing Co., Inc., Boston, MA, USA.

Metzger, A., Chi, C.-H., Engel, Y., and Marconi, A. (2012).

Research Challenges on Online Service Quality Pre-

diction for Proactive Adaptation. In Software Services

and Systems Research - Results and Challenges (S-

Cube), 2012 Workshop on European, pages 51 –57.

Nguyen, H. and Chan, W. (2004). Multiple neural networks

for a long term time series forecast. Neural Comput.

Appl., 13(1):90–98.

Raab, F., Kohler, W., and Shah, A. (2001). Overview of the

TPC Benchmark C: The Order-Entry Benchmark.

Sen, S. (2008). Business Activity Monitoring Based on

Action-Ready Dashboards And Response Loop. In

Proceedings of the 1st International Workshop on

Complex Event Processing for Future Internet.

Wetzstein, B., Zengin, A., Kazhamiakin, R., Marconi, A.,

Pistore, M., Karastoyanova, D., and Leymann, F.

(2012). Preventing KPI Violations in Business Pro-

cesses based on Decision Tree Learning and Proactive

Runtime Adaptation. J. of Sys. Integration, 3(1):3–18.

Zeng, L., Lingenfelder, C., Lei, H., and Chang, H.

(2008). Event-Driven Quality of Service Prediction.

In Bouguettaya, A., Krueger, I., and Margaria, T.,

editors, Service-Oriented Computing - ICSOC 2008,

volume 5364 of Lecture Notes in Computer Science,

pages 147–161. Springer Berlin Heidelberg.

IncreasingEnergySavingwithService-basedProcessAdaptation

209