Developing a resource management scheme to keep the Cloud afloat

‘Dynamic Virtual Machine Consolidation for Energy Efficient Cloud Data Centers’  by Dong-Ki Kang, Fawaz Alhazemi, Seong-Hwan Kim, and Chan-Hyun Youn (School of Electrical Engineering, KAIST, Daejeon, Korea)
Best Paper Award at CLOUDCOMP 2015, 6th EAI International Conference on Cloud Computing

It doesn’t need saying that globally, we do a lot of computing every single day. What does need saying is that data centers and server infrastructures consume crazy amounts of energy. And the trend isn’t showing signs of stopping. On the contrary, cloud computing has officially stepped into the mainstream and global bandwidth demands increase steadily. Sustainability and evironmental aspects aside, energy consumption makes a severe dent in the budget of any operation. As stated in the paper, some reports estimate that the “cost of power and cooling has increased 400% over 10 years, and 59% of data centers identify them as key factors limiting server deployments.” Authors of this paper have set out to tackle this spreading issue by consolidating resource allocation dynamically with a Virtual Machine (VM).

Firstly though, let’s introduce the most popular method for achieving highest possible energy efficiency in data centers – Dynamic Right Sizing (DRS). DRS essentialy turns off idle servers, but there is more to it than that. To maximize the efficiency via DRS, one of primary adaptive resource management strategies is a Virtual Machine consolidation, in which running VM instances can be dynamically integrated into the minimal number of cloud servers in accordance with their resource utilization collected by hypervisor monitoring module. That is, running VM instances on under-utilized servers which are supposed to be turned off could be migrated to power-sustainable servers.

However, cloud users are many, and and their resource demands differ. Reckless switching of servers in high volume could lead to undesirable performance degradation. To achieve the demanded Quality of Service, authors of this paper have devised a careful resource management scheme which considers switching overheads. They propose a management system called Self Adjusting Workload Prediction to increase the prediction accuracy of users’ future demands even under irregular workload patterns. The method adaptively scales the history window size up or down according to extracted workload’s autocorrelations and sample entropies which measure periodicity and burstiness of workloads.
To learn more about how the system works and to see how the tests went, we recommend getting the full paper on EUDL.

CLOUDCOMP 2016 is still accepting papers! Learn more on the conference website.