VMSA: a performance preserving online VM splitting and placement algorithm in dynamic cloud environments

Server consolidation schemes whereby each server is replaced with a virtual machine (VM) and multiple such VMs are run on a single physical server can reduce the number of physical servers needed, and in turn, both the cost and energy consumption in data centers. However, existing schemes have not fully exploited the flexibility in the usage and allocation of virtualization resources, so as to allow one application originally deployed on a single large VM (LVM) to be split and hosted by multiple smaller VMs (SVM). Using multiple SVMs instead of an LVM enables resource allocation at a smaller granularity and thus may further increase the utilization and reduce the number of physical servers. However, a major challenge to overcome when deploying multiple SVMs for one application is to preserve the performance of the application in terms of response delay. In this paper, we show through theoretical analysis and experiments that in order to preserve the performance of the application, one needs to allocate sufficient resources to each SVM, and the total amount of resources required by all the SVMs will exceed that required by the LVM. Nevertheless, we also show that by using the proposed heuristic algorithm called VM splitting and assignment (VMSA), we can substantially improve the utilization and reduce the number of physical servers.