VMware distributed resource Management : design , Implementation , and lessons learned

Automated management of physical resources is critical for reducing the operational costs of virtualized environments. An effective resource-management solution must provide performance isolation among virtual machines (VMs), handle resource fragmentation across physical hosts and optimize scheduling for multiple resources. It must also utilize the underlying hardware infrastructure efficiently. In this paper, we present the design and implementation of two such management solutions: DRS and DPM. We also highlight some key lessons learned from production customer deployments over a period of more than five years. VMware's Distributed Resource Scheduler (DRS) manages the allocation of physical resources to a set of virtual machines deployed in a cluster of hosts, each running the VMware ESX hypervisor. DRS maps VMs to hosts and performs intelligent load balancing in order to improve performance and to enforce both user-specified policies and system-level constraints. Using a variety of experiments, augmented with simulation results, we show that DRS significantly improves the overall performance of VMs running in a cluster. DRS also supports a " what-if " mode, making it possible to evaluate the impact of changes in workloads or cluster configuration. VMware's Distributed Power Management (DPM) extends DRS with the ability to reduce power consumption by consolidating VMs onto fewer hosts. DPM recommends evacuating and powering off hosts when CPU and memory resources are lightly utilized. It recommends powering on hosts appropriately as demand increases, or as required to satisfy resource-management policies and constraints. Our extensive evaluation shows that in clusters with non-trivial periods of lowered demand, DPM reduces server power consumption significantly.