Evaluating the Impact of Memory Allocation and Swap for Vertical Memory Elasticity in VMs

Abstract

Typically, virtual machine (VM) allocation is based on the host server's ability to meet the VM's maximum CPU, I/O and memory requirements. However, given that the requirements of applications within the VM may vary during execution, it might be more efficient to also vary over time the amount of resources dedicated to the VM. In cloud systems, vertical elasticity is the dynamic adjustment of the amount of a physical resource, such as memory, CPU cores, etc., that is allocated to a VM. With technology pushing up core counts and speeds of modern servers, and given the growing trend towards server consolidation, making the most of the available memory is crucial for good application performance. This paper investigates the impact of memory allocation and swap usage on VM performance. Through an experimental evaluation, hyper visor independent metrics and policies are identified for consideration by tools that claim to offer vertical memory elasticity. Based on the conclusions, the paper goes on to present a framework of a tool to dynamically manage memory allocations of VMs. Preliminary results with the proposed vertical Memory Elasticity Controller, MEC, highlight some of the benefits to both resource providers and applications through improved efficiency, throughput and performance. Ongoing work will continue to expand on the current evaluation and refine the scheduling policies to further improve the tool.