Professional Documents
Culture Documents
ABSTRACT
This paper addresses the problem of run-time management of a service centre
energy efficiency by using a context aware self-adapting algorithm. The algorithm
considers the service centre current context situation and predefined Green/Key
Performance Indicators (GPI/KPI) to adapt and optimize the service centre energy
consumption to the incoming workload. The service centre energy performance
context situation is obtained by collecting service centre current context data
related to the business applications scheduled to run on service centre resources,
service centre IT computing resources and the service centre facilities. A
reinforcement learning approach is used to decide the best sequence of actions to
be executed to bring the service centre resources as close as possible to a GPI and
KPI compliant state while minimizing the energy consumption.
1 INTRODUCTION AND RELATED WORK ambient data related to the service centre
environment. The self-adapting algorithm implements
Over the last years the energy efficiency the four phases of a MAPE control loop: Monitoring,
management of IT processes, systems and service Analyzing, Planning and Execution. In the
centres has emerged as one of the most critical monitoring phase, the service centre energy /
environmental challenges to be dealt with. Since performance context data is collected and represented
computing demand and electricity price is in a programmatic manner by using our RAP context
continuously rising, energy consumption of IT model [4]. In the analyzing phase, the service centre
systems and service centres is expected to become a Green and Key Performance Indicators (GPIs/KPIs)
priority for the industry. Nowadays, service centre are evaluated on the service centre energy
electricity consumption accounts for almost 2% of performance context instance and compared against
the world electricity production and their overall their predefined values. If the GPIs/KPIs are not
carbon emissions [1], [2]. The GAMES (Green fulfilled in the planning phase a reinforcement
Active Management of Energy in IT Service learning approach is used to generate the best
Centres) [3] research project aims at developing a set sequence of actions to be taken in the execution phase
of innovative methodologies, metrics, services and for bringing the service centre as close as possible to
tools for the active management of energy efficiency a GPIs/KPIs compliant state. The adaptation
of IT service centres. The GAMES vision is to create algorithm decides on two types of adaptation actions:
a new generation of Green and Energy Aware IT consolidation actions and DPM (Dynamic Power
service centres by designing energy efficient service Management) actions. The consolidation actions are
centres and also by managing at run-time the service executed for balancing the service centre workload in
centre energy efficiency. an energy efficient manner. DPM actions identify the
This paper addresses the problem of run-time service centre over-provisioned resources targeting to
managing the service centre energy efficiency by set them into low power states.
proposing a context aware self-adapting algorithm for In the literature, the development of energy-aware
adjusting the resources usage to the incoming computing systems is proposed in [5]. Different
workload. The service centre context consists of the energy aware techniques based on integer
following categories of energy / performance data: (i) programming, genetic algorithms or heuristics have
data related to business applications to be executed on been proposed in the last years [6], [7], [8], [9]. A
the service centre IT computing resources, (ii) data systematic analysis of cost issues in the design of
related to the workload and energy consumption of hardware and network systems is presented in [10]
the service centre IT computing resources and (iii) while a cost-oriented design of data centres in which
applications are allocated with a fixed number of tiers conditions in a specific context situation. We define
for satisfying response time requirements is proposed the self-adapting property for a context aware system
in [11]. as a function, which associates to each context
To reduce energy consumption, autonomic situation that fails to fulfil the prescribed set of
techniques have been developed to manage workload conditions, an adaptation action plan, that should be
fluctuations [12] and to determine the optimum executed in order to re-enforce the set of conditions.
trade-offs between performance and energy costs, Fig. 1 presents the proposed self-adapting
mainly by switching servers on and off. It has been algorithm for context aware systems. The algorithm
shown that by implementing Dynamic Voltage is based on a control feedback loop with the
Scaling (DVS) mechanisms (i.e., reduction of CPU following MAPE phases: (i) the Monitoring Phase –
frequency), significant energy reduction can be collects raw data about the system’s self and
obtained [13], [14]. Self-* algorithms for ad-hoc execution environment and uses it to construct and
networks, in which the connections between nodes represent the system context situation in a
are added / removed for energy efficiency purposes programmatic manner (lines 7-9), (ii) the Analysis
are proposed in [15], [16]. An agent based solution Phase – analysis the context situation in order to
for power management in complex service centres detect significant abnormal changes (lines 10-11),
comes from the IBM researchers R. Das and J. O. (iii) the Planning Phase – decides and selects the
Kephart [17]. The authors use virtualization and appropriate adaptation action plan to be executed
load-balancing techniques for efficiently and energy (lines 12-24), (iv) the Execution Phase – executes
aware management of service centre’s resources. In the selected action plan (lines 25-26).
[18], a self-adaptive solution for optimizing and
balancing performance and energy consumption for
RAID storage is proposed. This technique groups the
available data into hot (frequently accessed data) and
cold data. Hot data is kept on an active disk group
while the cold data is kept in an inactive disk group
which spins at low speed in order to save energy.
Energy–aware provisioning and load dispatching
algorithms that take into consideration the energy /
performance models and user experience are
designed for servers that host a large number of long-
lived TCP connections [19]. An architecture which
supports the design and programming of self-
optimizing algorithms for data locality and access in
distributed systems is considered in [20]. Based on
application's run-time behaviour, the data access and
distribution is re-modelled with respect to a set of
predefined energy performance objectives.
The rest of the paper is organized as follows:
Section 2 details each MAPE phase of our context
aware self-adapting algorithm; Section 3 shows a
how the context aware self-adapting algorithm can
be used to manage the energy efficiency of a service
centre while Section 4 concludes the paper.
Figure 1: The self-adapting algorithm.
2 THE CONTEXT AWARE SELF-
The following sub-sections detail the first three
ADAPTING ALGORITHM
phases of the self-adapting algorithm. The execution
phase is trivial and is not further detailed.
To define the self-adapting property for a
context aware system we have used three main
2.1 The Monitoring Phase
situation calculus concepts: (1) the situation concept
The goal of the self-adapting algorithm
also referred as system situation or context situation,
monitoring phase is to collect the system’s self and
represents the state of a context aware system
execution environment raw context data and to
(system’s self and execution environment) at a
construct / represent the system context situation in a
moment of time, (2) the set of conditions concept,
programmatic manner. The system’s self raw context
represents a set of rules that defines the possible
data refer to the system internal state and is used to
system-context interactions and are used to guide and
assure the system’s self-awareness. This kind of
control the system execution (3) the action concept,
context data is usually gathered using low level
represents the adaptation actions or actions plans that
system calls, logging or life cycle events. The
are executed by the system to enforce the set of
system’s execution environment raw context data are example). Each Reference has a unique name that
used to ensure the system environmental awareness. can be referred from all the other XML elements of
This kind of context data is captured using sensor the policy. The reference resources collection is
networks or intelligent devices with the goal of defined by applying three types of restrictions:
assuring the system context awareness. domain restriction (location as criteria), type
The system’s self and execution environment restriction (context resource class as criteria) and
context data collected from various sources is property restriction (context resource properties other
represented in heterogeneous formats and difficult to than location as criteria).
interpret and analyze. There is an evident need for The XML Subject element uses an Reference
providing an integrated, uniform and semantically name to identify the context resources collection on
enhanced context data representation model that can which the policy imposes restrictions. The Subject
be automatically processed and interpreted at run- element has two children: (i) EvaluationTrigger
time. To solve these problems we have used our XML element that contains a set of possible events
RAP (Resources, Actions and Policies) context that may trigger the policy evaluation and (ii)
model defined in [4] to represent the system’s EvaluationCondition XML element containing a set
context situation in a programmatic manner. The of conditions for which the policy is enforced.
RAP model defines two types of context information The XML Target element represents a collection
representations: (i) set based, used to determine the of context resources for which the goals specified by
context changes and (ii) ontology based, used to infer the PolicyGoals XML element applies. The
new context information by means of reasoning and PolicyGoals element is used to determine the action
learning algorithms. plan that has to be executed to enforce the policy.
In the set based approach, the context information To be evaluated, the policies are automatically
is modelled as a triple, C = <R, A, P >, where R is converted into SWRL (Semantic Web Rule
the set of context resources that provide raw data Language) rules and injected into the RAP specific
about the system’s self and execution environment, context model instance ontology. The SWRL rules
A is the set of adaptation actions which are executed are used to reason about RAP context model instance
to enforce system execution guiding rules and P is a ontology individuals in terms of ontology classes and
set of policies which define the rules that guide the properties. Rules are written in the form of an
system execution. In the ontology-based implication between an antecedent (body) and a
representation, the relationships between the context consequent (head).
model sets are modelled in a general purpose context The process of obtaining the SWRL rule
ontology core. The domain specific concepts are antecedent from the XML policy description consists
represented as sub trees of the core ontology by from two phases: (i) the generation of the SWRL
using is-a type relations. A RAP context model atoms used to identify the specific context model
instance represents a context situation and contains instance ontology individuals involved in the
the values of the RAP context model sets elements in reasoning process and (ii) the transformation of the
a specific moment of time. XML policy evaluation condition into SWRL rules.
The SWRL rule consequent contains a SWRL
2.2 The Analysis Phase atom that sets the EvaluationResultProp attribute to
In the analysis the context policies that guide the true for all RAP context model Policy class
system execution are evaluated for the current individuals, which correspond to the XML broken
context situation with the goal of identifying those policies.
context situations in which the rules are broken. The
context policies are represented in XML and 2.2.2 The context situation entropy
automatically converted in SWRL for evaluation The system context situation entropy measures
using our policy representation / evaluation model the level of system’s self and execution environment
proposed in [22]. For measuring the degree of disorder by evaluating the degree of fulfilling the
fulfilling the context policies in a context situation, context policies. If the evaluated system context
we have defined the concept of context situation situation entropy is below a predefined threshold T E,
entropy (ES) and its associated threshold (T E). The then all the policies are fulfilled and adaptation is not
next sub-sections present a short overview of our the required. Otherwise the system must execute
policy representation / evaluation model and adaptation actions to bring the entropy below TE. The
introduce the context situation entropy concept. entropy for a context situation is:
(5)