This Thesis aims to evaluate an approach to reduce data centre energy consumption using reinforcement learning algorithms to optimize the virtual machine (VM) selection process.VM selection is the process of selecting an VM from a overloaded host and moving it to another host.An optimized selection of VMs can lead to a few overloaded hosts and this leads to a reduction in energy usage.
Two reinforcement learning algorithms, Q-learning and SARSA were implemented in the cloudsim toolkit. Subsequent experiments employed two distinct policies—epsilon greedy and SoftMax—to determine the most effective hyperparameters, specifically the learning rate (alpha) and the discount factor (gamma). The proposed algorithm results in a energy saving of 18% compared to the Lr-Mmt approach. The results of this thesis conclude that the RL algorithm can intelligently optimize the VM selection process and thereby reducing the energy consumption in the data center.
Language:Java
RL Algorithms: Q learning and SARSA
The code is uploaded as a Zip file 'cloudsim-3.0.3.zip'.Its based on the cloudsim toolkit Tag 3.0.3 from CLOUDS Labortaory in Melbourne.Few of the files are updated to implement Reinforcement Learning.
The experimental Logs are updated in Logs . The text logs from the Lrmmt and Lr-RL experimental runs are logged and converted to csvs and saved in this folder.
The Analysis_Notebooks has all the files used to analyze the experiments and create statistial Reports using the experimental Logs.