Design and Evaluation of Reward Functions for Load Balancing in Cloud Datacenters

Task scheduling and virtual machine (VM) consolidation techniques optimize resource utilization and distribute the load evenly across physical machines and virtual machines in a virtualized data center. Users’ requests to applications hosted on the virtual machines are dynamic. Though schedulers assign tasks to VMs based on the current state of the VMs and the expected resource requirements of the tasks, the requirements of the tasks typically tend to deviate from the expected requirements. This eventually leads to load imbalance and skewed energy consumption across virtual machines. Load imbalance and high energy consumption occur due to dynamic nature of workloads across the data center. An overloaded VM experiences degradation in performance thereby violating Quality of Service (QoS) and Service Level Agreement (SLA) requirements. In this work, a reinforcement learning (RL)-based task scheduling approach is utilized to minimize load imbalance, wherein we compare the effectiveness of using different reward functions. The RL agent learns a policy to minimize the number of overloaded VMs and to reduce deviation from a load-balanced condition by efficiently scheduling tasks using Q-learning algorithm. It takes scheduling decisions with the help of previous experiences when it encounters similar workload patterns. Experiments have been performed on the real time Google cluster traces and the proposed reward functions were assessed for their performance.