An ELM-based Load Balancing Algorithm for Cloud Computing Platforms

Since the workload of the end users and the provisioned cloud resources are dynamically changed over time, the workload is not evenly distributed over the cloud. Therefore, designing appropriate mechanisms to detect the status of the cloud and properly distribute the load on each host can play an effective role in improving system performance and energy consumption in cloud data centers. Reactive load balancing approaches don’t prevent load-imbalance in cloud and make virtual machines (VM) migrate after load imbalance and increase energy consumption and job response time. Also, in proactive load balancing methods, some problems, such as host state detection with insufficient accuracy and fixed threshold of cpu utilization without considering the host current and future states in VM migrations, prevent the optimal number of balanced hosts and energy consumption in datacenters. In this paper, a proactive approach to the early detection of host states is presented which is based on Extreme Learning Machine (ELM). The proposed approach predict the CPU utilization of each host over time and applies an adaptive threshold to determine the future status of each host (i.e., overload, underload, secure and normal state). Then, a subset of VMs are migrated to hosts with minimum overload probability in future to avoid overloaded hosts. Implementation of the proposed method and its evaluation on the real data sets in Cloudsim show that the proposed method improves energy consumption, response time, the number of VM migrations and non-violation of the Service Level Agreement (SLA) in comparison to competitive algorithms including RF-LB [7] and ANN-LB [13].