2.2. Power-aware resource provisioning in data centers In recent times, there has been a focus towards optimizing
the workloads in data centers by maximizing resource utiliza-
tion while minimizing the energy costs. Researchers [
15
,
28
,
37
–
43
] have focused on energy-efficiency-based strategies for efficient
utilization of data center resources.
The authors [
39
] provide an energy estimation model based on
VM resource utilization. Using this model, they develop a power-
aware scheduling algorithm for VM placement in a cloud cluster.
Each VM is assigned a quota of maximum energy utilization. As
soon as a VM exceeds its energy quota, it is assigned no further
tasks until the VM quota is re-set. A major drawback in this work
is the reliance of quota computation based on CPU utilization. As
mentioned earlier, Storage I/O and network utilization are also a
factor in energy consumption, which was not addressed in this
work. Xiangming Dai [
28
] proposed algorithms to reduce energy
consumption in data centers by considering the intelligent place-
ment of VMs on the servers. They formulated the scheduling prob-
lem as an integer programming problem and explored two greedy
approximation algorithms, a minimum-energy VM scheduling al-
gorithm (MinES), and a minimum-communication VM scheduling
algorithm (MinCS) to reduce the energy while satisfying the tenant
service-level agreements. Their results demonstrate that MinES
and MinCS yield scheduling that is within 4.3% to 6.1% of energy
consumption of the optimal solution while being computationally
efficient.
Annan [
27
,
44
] described an optimization approach using the
dynamic placement of VMs in cloud computing environment. They
propose a model that addresses minimization of the cost of placing
VMs based on categorization of clients SLA requirements. VMs with
workflow belonging to a client are strategically placed on collo-
cated PM to reduce communication overheads, whereas a penalty
is applied for each VM migration to another PM. Limited simulation
studies with only 3 VMs compare the proposed approach. A seri-
ous flaw in this approach is that the complexity would increase
with the increase in workflows, which is also evident from the
results. Researchers [
37
] have introduced an uncertainty-aware
scheduling architecture to mitigate the effect of uncertain factors
on the workflow scheduling quality. Based on the proposed archi-
tecture, the authors presented a scheduling algorithm by incor-
porating event-driven and periodic-rolling strategies for dynamic
workflows. Extensive simulation studies show that the proposed
algorithm performs better compared to the traditional scheduling
algorithms.
Chen et al. [
40
] presented an energy-efficient workload aware
(EEWA) task scheduler using online profiling that collects work-
load information of tasks for CPU-bound parallel applications. A
frequency adjuster tunes the core frequencies based on workload
information. The tasks are assigned to the cores in the processor by
the scheduler. The EEWA task scheduler preserves a tradeoff be-
tween energy utilization and CPU-bound application performance
in multi-core architectures. Authors in [
43
], improve on the disk
caching methodology used in Apache Spark for processing inter-
mediate results in Machine Learning and Data mining algorithms.
They propose an optimization algorithm to adaptively determine
and cache the most valuable and reusable intermediate datasets.
Based on real-time experimentation study carried out on apache
Spark, the proposed algorithm improves the performance. This
work however does not address the impact on energy efficiency
of the proposed approach.
Yang et al. [
45
] address the challenge of energy cost optimiza-
tion in datacenters due to the bottleneck effect created by the mod-
est performance of the storage I/O. The parallel I/O to multiple disk
drives in Redundant Array of Independent Disks (RAID) provides a
limited performance improvement for big data applications. They
devise ‘‘AutoTiering’’ an optimization framework to provide the
best global migration and allocation solution over the data center.
An optimization algorithm for efficient placement of resources
based on historical and predicted performance factors is presented.
Researchers [
41
] have proposed an elastic distributed-resource
scaling framework called AGILE. This framework is capable of han-
dling dynamic workloads by predicting server-overload to avoid
SLA penalties. It does so by considering the workflow requirements
(CPU, IO profiles) in recent past using a sliding window mechanism.
If a server overload is predicted, the VMs on this server a marked
for migration. The VMs set for migration are cloned with minimum
penalty and migrated to vacant PMs. Although the model is light