Evaluating the need for complexity in energy-aware management for cloud platforms

• Authors: Pooja Ghumre;Junwei Li;Mukil Kesavan;Ada Gavrilovska;Karsten Schwan
• Venue: ACM SIGMETRICS Performance Evaluation Review
• Year: 2012

In order to curtail the continuous increase in power consumption of modern datacenters, researchers are responding with sophisticated energy-aware workload management methods. This increases the complexity and cost of the management operation, and may lead to increases in failure rates. The goal of this paper is to illustrate that there exists considerable diversity in the effectiveness of different, potentially 'smarter' workload management methods depending on the target metric or the characteristics of the workload being managed. We conduct experiments on a datacenter prototype platform, virtualized with the VMware vSphere software, and using representative cloud applications -- a distributed key-value store and a map-reduce computation. We observe that, on our testbed, different workload placement decisions may be quite effective for some metrics, but may lead to only marginal impact on others. In particular, we are considering the impact on energy-related metrics, such as power or temperature, as corresponding energy-aware management methods typically come with greater complexity due to fact that they must consider the complex energy consumption trends of various components in the cloud infrastructure. We show that for certain applications, such costs can be avoided, as different management policies and placement decisions have marginal impact on the target metric. The objective is to understand whether for certain classes of applications, and/or application configurations, it is necessary to incur, or if it is benefitial to avoid, the use of complex management methods.