Reserve policy‑aware VM positioning based on prediction in multi‑cloud environment

Abstract

The proper positioning of Virtual Machines (VMs) on the hosts in a cloud environment reduces the need for the VM migration and its consequences. The positioning becomes more signifcant when there exists a multi-cloud environment where the hosts exist on multi-site datacenters. Based on user’s requests, VMs should be dynamically positioned; however, if the users’ future demands can be predicted, the positioning can be adaptively done in advance, which is both more cost-efective for users and more requests are met. To this end, at the request of their users, VMs’ providers can reserve VMs for the users’ future needs. However, if some users would not like to reserve VMs or if the number of reserved VMs is less than users’ needs, VMs should be allocated on demand. However, the reserve or on-demand policy cannot be applied freely if users have constraints and objectives. Among others, cost of using resources and response time are the most important users’ objectives, and load balancing hosts and datacenters for the proper resource utilization is the most important providers’ objective. To consider the reserve policy, a multi-layered model is presented in this paper where a multi-objective optimization is used to meet the objectives. The proposed model was applied to Google, Wikipedia, and NASA datasets. The results show: (1) The number of predicted VMs for reserve is closer to the real VMs requested in datasets NASA, Wikipedia, and Google than the related work. This was due to the use of a dynamic neural network, called NARX; (2) objective cost is regarded more than the related work, while it respects more trade-of between the user’s objectives and provider’s one; (3) placement of VMs on hosts is done in a balanced way, leading to the reduction of overloaded hosts and response time.