Information Technology Reference
In-Depth Information
for short-term variability in the grid as well as for contingency purposes, such
as recovering after faults in transmission lines or generators.
Many different techniques have been proposed to reduce the energy consump-
tion of data-centers [1], from low-level hardware solutions to high-level software
methods. Sustainable energy-aware techniques are in conflict with system per-
formance metrics that accounts for the QoS perceived by the user, because in-
creasing performance usually leads to increase the energy consumption. Thus,
multiobjective approaches are required to model the reality of current datacen-
ters operation when taking into account the energy eciency.
This paper presents a multiobjective approach for datacenter planning that
accounts for both server (IT) and cooling infrastructures: free cooling and air
conditioning (AC) in order to provide appropriate levels of quality of service
(QoS) and temperature when following a specific power consumption profile.
We propose a two-phase approach for control and scheduling. In the upper
level, a multiobjective evolutionary algorithm (MOEA) is applied for power con-
trol according to a reference power profile and temperature, providing multi-
ple trade-off solutions to the problem. In the lower level, specific energy-aware
scheduling heuristics are applied to provide appropriate QoS according to Server
Level Agreements (SLA) between provider and user. The multiobjective ap-
proach helps the datacenter planner to explore different options for controlling
the system performance and energy consumption.
The experimental evaluation, performed considering a set of realistic work-
loads and hardware scenarios, indicate that the proposed approach is a useful op-
tion for power management in datacenters. When compared against a business-
as-usual (BAU) strategy, the proposed MOEA is able to compute solutions with
up to
78%
improvement on power tracking and
86%
on the temperature values,
with very low degradation in QoS-related metrics.
The paper is organized as follows. Section 2 reviews the related work about
power control and energy-aware scheduling in datacenters. The problem model
and the proposed control and planning strategies for energy-aware datacenters
is described in Section 3. The proposed MOEA for energy-aware datacenter
control and planning is described in Section 4. The experimental evaluation is
reported and discussed in Section 5. Finally, Section 6 presents the conclusions
and formulates the main lines for future work.
2 Related Work
This section reviews the main related work in literature in the areas of control
and energy-aware scheduling in datacenters.
Control and energy aware datacenters
. GreenSlot [9] considers job allocation
for HPC applications in a datacenter powered by solar panels, using job infor-
mation (nodes per job, deadline, estimated runtime) for scheduling when solar
energy is available. GreenHadoop [10] considers green generation and energy
prices to allocate MapReduce jobs using heuristics to predict the energy re-
quirements. GreenSwitch [8] also considers energy storage (batteries and net
