Brownout-oriented and energy efficient management of cloud data centers

Abstract

Cloud computing paradigm supports dynamic provisioning of resources for delivering computing for applications as utility services as a pay-as-you-go basis. However, the energy consumption of cloud data centers has become a major concern as a typical data center can consume as much energy as 25,000 households. The dominant energy efficient approaches, like Dynamic Voltage Frequency Scaling and VM consolidation, cannot function well when the whole data center is overloaded. Therefore, a novel paradigm called brownout has been proposed, which can dynamically activate/deactivate the optional parts of the application system. Brownout has successfully shown it can avoid overloads due to changes in the workload and achieve better load balancing and energy saving effects.

In this thesis, we propose brownout-based approaches to address energy efficiency and cost-aware problem, and to facilitate resource management in cloud data centers. They are able to reduce data center energy consumption while ensuring Service Level Agreement defined by service providers. Specifically, the thesis advances the state-of-art by making the following key contributions:

1) An approach for scheduling cloud application components with brownout. The approach models the brownout enabled system by considering application components, which are either mandatory or optional. It also contains brownout-based algorithm to determine when to use brownout and how much utilization can be reduced.

2) A resource scheduling algorithm based on brownout and approximate Markov Decision Process approach. The approach considers the trade-offs between saved energy and the discount that is given to the user if components or microservices are deactivated.

3) A framework that enables brownout paradigm to manage the container-based environment, and provides fine-grained control on containers, which also contains several scheduling policies for managing containers to achieve power saving and QoS constraints.

4) The design and development of a software prototype based on Docker Swarm to reduce energy consumption while ensuring QoS in Clouds, and evaluations of different container scheduling policies under real testbeds to help service provider deploying services in a more energy-efficient manner while ensuring QoS constraint.

5) A perspective model for multi-level resource scheduling and a self-adaptive approach for interactive workloads and batch workloads to ensure their QoS by considering the renewable energy at Melbourne based on support vector machine. The proposed approach is evaluated under our developed prototype system.