Modeling workloads and devices for IO load balancing in virtualized environments

Abstract

Virtualization has been effective in providing performance isolation and proportional allocation of resources, such as CPU and memory between VMs by using automated distributed resource schedulers and VM migration. Storage VMotion allows users to migrate virtual hard disks from one data store to another without stopping the virtual machine. There is a dire need for an automated tool to manage storage resources more effectively by doing virtual disk placement and load balancing of workloads across multiple data stores. Applicable beyond virtualization, this problem is challenging because it requires modeling both workloads and characterizing underlying devices. Furthermore, device characteristics such as number of disks backing a LUN, disk types etc. are hidden from the hosts by the virtualization layer at the array. In this paper, we propose a storage resource scheduler (SRS) to manage virtual disk placement and automatic load balancing using Storage VMotion. Our initial results lead us to believe that we can effectively model workloads and devices to improve overall storage resource utilization in practice.