Energy-Aware Policies in Ubiquitous Computing Facilities - Cloud Computing with e-Science Applications - page 273

to understand the energy demand characteristics of the different
analysis applications or the power consumption characteristics of
different resources.
• Power model: Complex power models are not adequate for online
optimization, as different alternatives should be quickly evaluated
against the power model to proactively configure the whole system
for minimum energy consumption. These power models can be
trained with actual data from sensors to improve the quality and to
adapt to variations in the heterogeneous application support network.
• Optimization: Based on the current state of the system, the historic
data, and the energy characteristics of application and resources,
many optimization algorithms can be executed to enhance one or
more aspects of the population-monitoring system. Heterogeneity
can be analyzed to always assign tasks to the most adequate
resources; resources not being used can be turned off; cooling
energy can be taken into account when assigning tasks to resources;
and, at the same time, when a group of nodes is detected to behave
anomalously, they can be discarded to provide some kind of
self-healing mechanism.
• Decision-making system: There are so many aspects that can be opti-
mized (at different levels of abstraction and in different scopes), that
it would not be feasible to consider all of them in a single optimiza-
tion algorithm. Many partial optimization algorithms may propose
actions in the network; some of them could even be incompatible
with other decisions. We propose the use of a reputation system [18]
to compose the decisions provided by multiple optimization algo-
rithms and to adapt to changes in the system by changing the weight
of different optimizations.
• Actuation support: Finally, decisions should be executed. Software
agents in all levels of the application support network are in charge
of reconfiguring their behavior whenever an optimization decision
is made.
12.4 Energy and Power Models
To apply energy optimization techniques at all levels, but most importantly to
the cloud computing facility, we need to develop power and energy models of
the resources of the data center that can be applied to predict the energy con-
sumption of the workload to be executed. In this section, we describe the most
important contributors to the energy consumption in data centers, and we
present some of the most relevant energy- and power-modeling techniques.