Energy-Aware Policies in Ubiquitous Computing Facilities - Cloud Computing with e-Science Applications - page 282

(based on different operating system environments) that can be dynami-
cally started and stopped according to the system workload and that share
physical resources.
Some research work has tried to address the VM provisioning challenge
by predicting the workload profile with neural networks and using heuris-
tics to assign the workload [37]. However, to obtain the most energy-efficient
setup, the MILP and GA previously described can be used to dynamically
assign VMs to physical servers, also deciding the amount of VMs needed to
execute a certain workload.
12.6.2 Consolidation
Historically, data centers have been oversized, using a small fraction of
their computing resources. Consolidation uses virtualization to share
resources and reduces energy consumption by increasing resource utiliza-
tion. This technique allows multiple instances of operating systems to run
concurrently on a single physical node, avoiding wasted physical resources.
Consolidation allows reducing the number of operating servers to process
the same workload, minimizing the static consumption, which leads us to
operating server set and turn-off policies.
Workload allocation algorithms should also take into account the possi-
bility of consolidation. As the number of decision variables and the design
space grow larger, GA-based solutions become more suitable for the purpose
of efficient VM assignment and consolidation.
12.6.3 Operating Server Set and Turn-off Policies
This technique consists of modifying the active server set by switching off
idle hosts when occupancy decreases. Another advantage of cloud comput-
ing is that in many applications, such as data mining and web searching,
using MapReduce provides outsourcing of the workload. MapReduce, pop-
ularized by Google [38], is widely used in application-level energy-aware
strategies due to simplified data processing for massive data sets to increase
data center productivity [39]. When a MapReduce application is submitted,
it is separated into multiple Map and Reduce operations so its allocation may
influence task performance. This factor allows leveraging server resources
by distributing the workload to achieve the optimal minimum consumption.
One of the issues to consider when implementing this type of policy is the
characterization of the use of the data center by customers. The demand for
resources reaching the data center is variable and usually follows seasonal
patterns depending on the time of day or certain periods of the year. In addi-
tion, the data center must be prepared to support peak demand.
Also, the quality of service (QoS) contracted by customers must be satisfied
in matters of availability and both execution and response time constraints.