Using Empirical Estimates of Effective Bandwidth in Network-Aware Placement of Virtual Machines in Datacenters

Abstract

Datacenter operators are increasingly deploying virtualization platforms to improve resource usage efficiency and to simplify the management of tenant applications. Although there are significant efficiency gains to be made, predicting performance becomes a major challenge, especially given the difficulty of allocating datacenter network bandwidth to multitier applications, which generate highly variable traffic flows between their constituent software components. Static bandwidth allocation based on peak traffic rates ensures SLA compliance at the cost of significant overprovisioning, while allocation based on mean traffic rates ensures efficient usage of bandwidth at the cost of QoS violations. We describe MAPLE, a network-aware VM ensemble placement system that uses empirical estimations of the effective bandwidth required between servers to ensure that QoS violations are within targets specified in the SLAs for tenant applications. Moreover, we describe an extended version of MAPLE, termed MAPLEx, which allows the specification of anticolocation constraints relating to the placement of application VMs. Experimental results, obtained using an emulated datacenter, show that, in contrast to the Oktopus network-aware VM placement system, MAPLE can allocate computing and network resources in a manner that balances efficiency of resource utilization with performance predictability.