A Runtime Evaluation Methodology and Framework for Autonomic Systems

ACEEE Int. J. on Network Security , Vol. 03, No. 02, April 2012

A Runtime Evaluation Methodology and Framework
for Autonomic Systems
Chonghyun Lee1, Hyunsang Youn1, Ingeol Chun2 and Eunseok Lee1
1
Sungkyunkwan University /Department of Computer Engineering, Suwon, South Korea
Email: carve9142@skku.edu, wizehack@skku.edu, leees@skku.edu
2
Electronics and Telecommunications Research Institute(ETRI), Daejeon, South Korea
Email: igchun@etri.re.kr

Abstract - An autonomic system provides self-adaptive ability Furthermore, to reduce the message transmission overhead,
that enables system to dynamically adjust its behavior on our goal model is dynamically activated depending on the
environmental changes or system failure. Fundamental system state.
process of adaptive behavior in an autonomic system is consist This paper is organized as follows: the next section
of monitoring system or/and environment information, provides a brief introduction of autonomic computing and
analyzing monitored information, planning adaptation policy
other related researches. Section III presents our self-adaptive
and executing selected policy. Evaluating system utility is
one of a significant part among them. We propose a novel
autonomic system framework and autonomic systems runtime
approach on evaluating autonomic system at runtime. Our evaluation methodology. Section c! presen ts the
proposed method takes advantage of a goal model that has implementation details of our prototype autonomic system
been widely used at requirement elicitation phase to capture and evaluation results. In section d!, we conclude our work
system requirements. We suggest the state-based goal model with future research directives.
that is dynamically activated as the system state changes. In
addition, we defined type of constraints that can be used to II. BACKGROUNDS
evaluate goal satisfaction level. We implemented a prototype
of autonomic computing software engine to verity our proposed A. Autonomic Computing
method. We simulated the behavior of the autonomic Autonomic computing (AC) is a computing paradigm
computing engine with the home surveillance robot scenario proposed by IBM[5] to solve the problems associated with
and observed the validity of our proposed method
the complex computing systems and unpredictable
Index Terms—autonomic computing, adaptive systems, goal operational environment. In conventional computing
model, embedded system paradigm, software engineers design, implement and test the
software assuming that the system will be operated under
I. INTRODUCTION predictable runtime environment predefined at requirement
elicitation process. As system execution environment have
As computing systems runtime environment is becoming become extremely complex, it is impossible to predict every
extremely complex, unpredictable situations those were situation a system might undergo. Autonomic computing
unforeseen at system design phase frequently occurs at introduced the concept of the human autonomic nervous
runtime. To cope with such situations after system system, which controls vital bodily function without human
deployment, enormous cost might be required, especially for intervention. Autonomic computing allows such
the large scale embedded systems such as a cyber physical functionalities to the system by embedding additional
system(CPS). Autonomic computing has been proposed to infrastructure in the system. The core aspect of autonomic
deal with such issue in pursuit of developing software that computing systems is as below:
adapt to its own situation and minimize the occurrence of the Self-configuration – Automatically re-configuring system
system failure without human intervention. Such autonomic components or integrating new components as system policy.
systems usually consist of system properties that deliver a Self-optimization – Continuously attempt to improve
monitoring, analysis, planning and execution feedback loop. system performance and efficiency seamlessly.
Evaluation of the autonomic system is one of a crucial part Self-healing – Automatically detects, localize system failure
among various research area related to autonomic computing. and recovers software and hardware problems
If an autonomic system provides a wrong evaluation of the Self-protection – System defends itself from malicious
adaptation result, such system might continuously make a attacks or failures.
wrong decision. Despite such importance, there are not
enough leading research and actual practice. In this paper, B. Utility-based Self-adaptive System Evaluation
we propose a novel approach on evaluating autonomic system Self-adaptive systems change system behavior or
at runtime. Our proposed method exploits a goal model that structures without human operator involvement. Therefore,
has been widely used at requirement elicitation phase to deciding when to execute adaptive policies and evaluating
capture system requirements. To utilize a goal model on runtime how effectively applied policies actually improved system
evaluation of the autonomic system, we defined constraints utilities are crucial. Kephart[9] suggested a system utility
that can be regarded as a runtime attribute of the goal model. evaluation method for self-adaptive system. The author
© 2012 ACEEE 21
DOI: 01.IJNS.03.02.91


separated self-adaptive system utility as service level utility analysis module to check goal violation.
and resource level utility. The method calculates every Analysis module – The analysis module examines received
possible resource level utility to find the best amount of data based on the analysis model(goal model) in the knowledge
resource to be allocated. D. Garlan[10][11] also proposed base. Analysis model actually computes goal satisfaction
similar approach on calculating system utility in Rainbow level at runtime. The analysis module sends goal violation
self-adaptive software architecture. Such methods were signal to the diagnosis module when goal satisfaction level
applied to a large-scale data center and web server system is below the predefined threshold. A detailed process of
respectively to show validity. Both of aforementioned calculating goal satisfaction level is described in section III -
methods adopt limited number of metric to calculate overall B.
system utility such as response time. However, such method Diagnosis module – The diagnosis module receives goal
is limited to a system with multiple objectives such as a violation information from the analysis module. Failure
robotics system, since a robotics system generally offers information from the analysis module is then mapped into
various functions. the top event of the diagnosis model, which is represented
as a fault tree in our framework. Fault tree is a graph
C. Goal Oriented Requirement Engineering
representing the combination of the fault events that can
A goal model has been used to capture system derive specific failure of a system. Conventional fault tree is
requirements for traditional systems on requirement used at system design phase to analyze system reliability,
elicitation phase. Goals tend to reflect what the system is but we utilize a fault tree at runtime to localize the root cause
supposed to do, and it has been regarded as a natural way to of the failure. Since the detailed mechanism of the runtime
model system requirements. fault tree is out of this papers scope, we do not describe it in
The Knowledge Acquisition in automated this paper.
Specification(KAOS) framework is a well-known approach Planning Module – The planning module receives root cause
for modeling system requirements using a goal model[12]. information from the diagnosis module. The planning module
The KAOS model is used at requirement elicitation phase to choose corresponding plan among several alternative plans
capture system goals. System goals are specified at high- and pass the chosen policy id to the execution module. The
level goals, and each of them is then decomposed into sub- execution module then sends policy id to the managed
goals that elaborate how the goal is achieved. The i* frame- elements through the effecter. After policy execution, the
work is another goal-oriented requirement engineering ap- evaluation module receives constraint information again from
proach that focus on the interaction between actors. There the managed elements and sends calculated goal satisfaction
are several recent researches that extend existing approaches level to the planning module for feedback.
to deal with uncertainty or inconsistency. However, as far as
we know there is no approach that takes advantage of a goal
model to evaluate autonomic systems behavior and the re-
sult of adaptation at runtime.

III. PROPOSED APPROACH

In this section, we first give a brief overview of our
autonomic computing software framework. Then the detailed
explanation about how the autonomic computing engine’s
analysis and evaluation are performed will be followed.
A. Overall Self-adaptive System Framework
Figure 1 Overall autonomic computing engine architecture
Fig. 1 shows an overall structure of our autonomic
computing software engine. It is an improved version of our B. Runtime evaluation
previous work[3]. We originally exploited rule-based tables To evaluate system status at runtime, we exploit a goal
as knowledge models, but we replaced these with tree model. By using a goal model, a system can trace current
structured goal models, and a fault tree. system status so that deciding which data to monitor be-
The autonomic computing engine in Fig. 1 interacts with comes easier. In case of a simple system such as a web server,
the managed element, namely the actual running system, only limited metrics such as response time and server cost
through the sensor and the effecter. The autonomic are enough to monitor the system performance. On the other
computing engine continuously checks the system status at hand, a system that supports complex services such as ro-
runtime, and decides whether adaptation policies are required. botic systems, an evaluation metric should be changed con-
A detailed functionality of each module in the autonomic sidering current system status or objectives that the system
computing engine is as follows: is supposed to achieve. The basic structure of our goal model
Monitoring module – The monitoring module requests is similar with the conventional goal model such as the one
required data to the managed element and receives the data used in KAOS. Fig. 2(a) shows an example goal model of our
through the sensor. Monitored data are then reported to the approach. The model represents the basic requirements of a
© 2012 ACEEE 22
DOI: 01.IJNS.03.02. 91


home surveillance robot that is capable of move to the desti- by Baresi[8], but they only applied it to non-functional
nation, avoiding obstacles, taking a picture with embedded requirement in the goal model. In our perspective, providing
camera, reporting environment information such as humidity flexibility on evaluating functional requirements is also
and temperature, and constant battery checking. Upper level necessary in many cases, so we extended fuzzy concept to
goals(parent node) are decomposed into lower level goals(child every leaf goals. To measure the satisfaction level of operation
node) and each child has a weight depending on the contri- constraints, a membership function should be set. Fiq. 3 show
bution level to a parent goal. Each weight can be gained by examples of membership functions.
using weight elicitation methods in the value tree analysis The evaluation module in fig. 1 calculates the total goal
[14]. satisfaction level. Each activated leaf goal’s achievement
At runtime, an entire goal model shown in fig. 2(a) is not level(LGA) is measured with the equation below:
fully maintained. Instead, the goals those are related to current
system state are activated at runtime. For example, when a
robot’s current goal is to move to some location (Move state),
activated goals at this time will be G0, G1, G4, G5, G6 and G12,
as shown in fig. 2(b). The robot is supposed to avoid where Gk denotes k th goal in the goal model and MCk, OCki is
obstacles in the moving path, so when the robot finds an the satisfaction level of the maintenance constraint and the
obstacle, the system state will be changed from Move state operation constrain respectively. f ki is i th operation
to Avoid Obstacles state. Therefore, the goals corresponding constraint’s membership function, where n is the number of
to Avoid Obstacles state will be automatically activated as the operation constrains, so the operation constraints are
the system state changes (fig. 2(c)). Each leaf goal has one or weighted sum of each constraint. After calculating LGAs,
more constraints that can be calculated at runtime. There are overall achievement level of the current goal model is also
two kinds of constraints in our model; maintenance calculated with weighted sum method. If a goal node has one
constraints(MC) and operation constrains(OC). or more leaf goals, equation(2) is applied, otherwise equation
(3) is recursively calculated until it reach to the root goal.

IV. EVALUATION
A. Scenario
Due to the lack of freely available open source autonomic
systems, we implemented a prototype of an autonomic
computing engine to validate the feasibility of our proposed
method(fig. 4). Our prototype is based on the home
surveillance robot, which has six available states; start, idle,
move, avoid obstacle, recharge, environment information
report, shoot pictures and end. The autonomic computing
Figure 2 Example goal models engine was developed in Java using Eclipse SDK, along with
MC represents a desire to keep a condition true over entire SQLiteJDBC a Java driver for SQLite. In this prototype system,
operation time. For example, when a robot is not supposed to the autonomic computing engine is not embedded in the
stop during the Move state, a MC could be keep the speed managed element, so we took advantage of TCP/IP protocol
over 0.0(or some other threshold value). MCs are periodically for the communication between the managed element and
monitored during runtime at certain time interval, and the the autonomic computing engine. This protocol can be
analysis module captures the goal violation as soon as MCs replaced with inter-process communication(IPC) when the
are violated. OCs represents the constraints those can be autonomic engine is embedded into the managed element.
evaluated after execution of the goal. For example, when a To demonstrate the home surveillance robot, we implemented
requirement for a robot is to reach to a certain destination, a simulated home environment. The robot consists of eleven
OC becomes elapsed time t to reach to the destination. On sensors that can be monitored by autonomic computing
evaluating OCs, we adopted fuzzy concept to deal with the engine. Our first experiment was set to evaluate whether the
ambiguity around the constraint boundary. For instance, if system accurately evaluates the system utility and detects
aforementioned OC is set to return 0.0 of satisfaction level goal violation. We simulated the robots moving behavior
after time t, the result will be 0.0 no matter how close the that belongs to the move state. The robot continuously sent
robot accomplished the goal to the requirement, and constraint information(current velocity in this case) to the
unnecessary adaptation process will continuously take place. autonomic computing engine when moving from the starting
Adopting fuzzy concept to a goal model is initially proposed point to the destination point. When the robot reached to
© 2012 ACEEE 23
DOI: 01.IJNS.03.02. 91


the destination, it sent operation constraint information to B. Experimental Result
the autonomic computing engine to calculate goal Fig. 6 shows the operation log message of the robot and
satisfaction level. To validate our state-based goal model in the autonomic computing engine. In fig. 6(a), the surveillance
terms of efficiency, we also simulated each available state robot continuously sent maintenance constraint information
and counted the number of messages sent between the home to the autonomic computing engine during move operation.
surveillance robot and the autonomic computing engine. For As soon as the robot arrived at the destination, operation
the counterpart, we counted the number of messages when constraint information(constraint ID(CID2-elapsed time)) was
the full goal model is kept during operation. sent to the autonomic computing engine. In fig. 6(b), we
simulated that the robot was stalled on the middle of path,
and CID1(current velocity) became 0.0, so the autonomic
computing engine instantly detected the goal violation. In
fig. 6(c), the robots move behavior ended without any goal
violation, so the autonomic computing engine received the
operation constraint information from the robot and calculated
goal satisfaction level based on current active goal tree(fig.
2(b)). We set the elapsed time to 7.0 seconds, where the
designated constraint on goal model to finish move operation
was 6.0seconds, therefore the membership function shown
in fig. 3(a) returned 0.714 of satisfaction level.
In our second set of experiments, we calculated total
number of messages sent between the autonomic computing
engine and the home surveillance robot to measure the cost
of monitoring. We simulated the autonomic computing
engine’s move, obstacle avoid, environment information
report, shoot pictures and charge battery state with state-
based goal model that we propose.
Figure 4 Class diagram of the autonomic computing engine

Figure 5 Samples of runtime evaluation

TABLE 1 NUMBER OF MESSAGES BETWEEN AUTONOMIC COMPUTING ENGINE AND THE of messages. On average, the number of messages passed
HOME SURVEILLANCE ROBOT
was reduced 20% than its counterpart. This is due to our
goal model that dynamically activates goals as system status
changes. Since unnecessary goals in current state are
excluded, the number of messages could be reduced.

CONCLUSIONS
We have proposed a runtime evaluation method for the
autonomous system. Our evaluation method takes advantage
At the same time, we simulated the autonomic behavior with
of stated-based goal model that dynamically activates the
a goal model that fully maintains state information. As
goal model depending on the system state. Through a
expected, the result shown in table.2 indicates that for any
simulated prototype system, we have observed that our
possible states, our proposed method reduced the number
© 2012 ACEEE 24
DOI: 01.IJNS.03.02. 91


proposed method can properly evaluate system utility and [4] Jae Sun Kim, Sooyong Park, VijayanSugumaran, “Contextual
detect system failures. In addition, our experiment has shown problem detection and management during software execution in
that the monitoring overhead from message passing was complex environments”, Industrial Management & Data Systems,
reduced by 20% with our state-based goal model. vol. 106, issue. 4, pp. 540-561. 2006
[5] J. O. Kephart, “Achieving self-management via utility
For the future work, we are investigating ways to
functions”, IEEE Internet Computing – INTERNET, vol. 11, no.1,
dynamically changing weights in the goal model, depending pp.40-48, 2007
on the contextual situation. This would provide the autonomic [6] J. O. Kephart, D. M. Chess, “The vision of autonomic
system a more robust way to evaluate system utility in various computing”, IEEE Computer – COMPUTER, vol. 36, no. 1, pp.41-
environments. 50, 2003
[7] Luciano Baresi, Liliana Pasquale, “Adaptive goals for self-
ACKNOWLEDGMENT adaptive service compositions”, IEEE International Conference on
Web Services(ICWS 2010), pp. 353-360, 2010
This work was supported by the IT R&D program of MKE/ [8] Luciano Baresi, Liliana Pasquale, Paola Spoletini, “Fuzzy goals
KEIT [10035708, “The Development of CPS(Cyber-Physical for requirements-driven adaptation”, In Proceedings of the 2010
Systems) Core Technologies for High Confidential Autonomic 18th IEEE International Requirements Engineering Conference (RE
Control Software”]. 2010), pp. 125-134, 2010
[9] S. A. DeLoach, M. Miller, “A goal model for adaptive complex
REFERENCES systems”, International Journal of Computational Intelligence:
Theory and Practice, vol. 5, no. 2, 2010
[1] C. Zhong, S. A. DeLoach, “Runtime models for automatic [10] Shang-Wen Cheng, David Garlan, Bradley R. Schmerl,
reorganization of multi-robot systems”, In Proceeding of the 6th “Evaluating the effectiveness of the Rainbow self-adaptive system”,
international symposium on Software engineering for adaptive and International Conference on Software Engineering – ICSE, pp. 132-
self-managing systems (SEAMS 2011). ACM, New York, NY, USA, 144, 2009
20-29. [11] Shang-Wen, An-Cheng Huang, David Garlan, Bradley R.
[2] Emmanuel Letier, Axel van Lamsweerde, “Reasoning about Schmerl, PerterSteenkiste, “Rainbow: Architecture-based Self-
partial goal satisfaction for requirements and design engineering”, adaptation with Reusable Infrastructure”, IEEEComputer –
In Proceedings of the 12th ACM SIGSOFT twelfth international COMPUTER, vol. 37, no. 10, pp.276-277, 2004
symposium on Foundations of software engineering (SIGSOFT [12] Van Lamsweerde, A. and Letier, “handling obstacles in goal-
’04/FSE-12), pp. 53-62, 2004 oriented requirements engineering”, IEEE Trans. On Software
[3] GiljongYoo, HyunsangYoun, Eunseok Lee, “Design Engineering. 26(10): 978-1005
methodology for runtime self-adaptiveness of CPS”, International [13] W. E. Walsh, G. Tesauro, J. O. Kephart, R. Das, “Utility
Conference on Computers, Communications, Control and functions in autonomic systems”, International Conference on
Automation, vol.2, pp. 82-85, 2011 Autonomic Computing – ICAC, pp. 70-77, 2004
[14] www.mcda.hut.fi/value_tree/theory/theory.pdf

© 2012 ACEEE 25
DOI: 01.IJNS.03.02. 91

A Runtime Evaluation Methodology and Framework for Autonomic Systems

More Related Content

What's hot (16)

Viewers also liked (8)

Similar to A Runtime Evaluation Methodology and Framework for Autonomic Systems (20)

More from IDES Editor (20)

Recently uploaded (20)

A Runtime Evaluation Methodology and Framework for Autonomic Systems