Energy and Performance: Management of Virtual Machines: Provisioning, Placement, and Consolidation

Abstract

Cloud computing is a new computing paradigm that offers scalable storage and compute resources to users on demand through Internet. Public cloud providers operate large-scale data centers around the world to handle a large number of users request. However, data centers consume an immense amount of electrical energy that can lead to high operating costs and carbon emissions. One of the most common and effective method in order to reduce energy consumption is Dynamic Virtual Machines Consolidation (DVMC) enabled by the virtualization technology. DVMC dynamically consolidates Virtual Machines (VMs) into the minimum number of active servers and then switches the idle servers into a power-saving mode to save energy. However, maintaining the desired level of Quality-of-Service (QoS) between data centers and their users is critical for satisfying users’ expectations concerning performance. Therefore, the main challenge is to minimize the data center energy consumption while maintaining the required QoS.

This thesis address this challenge by presenting novel DVMC approaches to reduce the energy consumption of data centers and improve resource utilization under workload independent quality of service constraints. These approaches can be divided into three main categories: heuristic, meta-heuristic and machine learning.

Our first contribution is a heuristic algorithm for solving the DVMC problem. The algorithm uses a linear regression-based prediction model to detect over-loaded servers based on the historical utilization data. Then it migrates some VMs from the over-loaded servers to avoid further performance degradations. Moreover, our algorithm consolidates VMs on fewer number of server for energy saving. The second and third contributions are two novel DVMC algorithms based on the Reinforcement Learning (RL) approach. RL is interesting for highly adaptive and autonomous management in dynamic environments. For this reason, we use RL to solve two main sub-problems in VM consolidation. The first sub-problem is the server power mode detection (sleep or active). The second sub-problem is to find an effective solution for server status detection (overloaded or non-overloaded). The fourth contribution of this thesis is an online optimization meta-heuristic algorithm called Ant Colony System-based Placement Optimization (ACS-PO). ACS is a suitable approach for VM consolidation due to the ease of parallelization, that it is close to the optimal solution, and its polynomial worst-case time complexity. The simulation results show that ACS-PO provides substantial improvement over other heuristic algorithms in reducing energy consumption, the number of VM migrations, and performance degradations.

Our fifth contribution is a Hierarchical VM management (HiVM) architecture based on a three-tier data center topology which is very common use in data centers. HiVM has the ability to scale across many thousands of servers with energy efficiency. Our sixth contribution is a Utilization Prediction-aware Best Fit Decreasing (UP-BFD) algorithm. UP-BFD can avoid SLA violations and needless migrations by taking into consideration the current and predicted future resource requirements for allocation, consolidation, and placement of VMs.

Finally, the seventh and the last contribution is a novel Self-Adaptive Resource Management System (SARMS) in data centers. To achieve scalability, SARMS uses a hierarchical architecture that is partially inspired from HiVM. Moreover, SARMS provides self-adaptive ability for resource management by dynamically adjusting the utilization thresholds for each server in data centers.