Applications-aware virtual machine provisioning

Abstract

Virtualization is widely used in large-scale computing environments, such as clouds, data centers, and grids, to provide multi-tenancy while retaining application isolation and portability. The consumers and providers of such virtualized infrastructures typically have distinct, often conflicting, requirements and optimization metrics. For instance, the user of a virtualized infrastructure typically cares about achieving high performance for their applications while minimizing costs. From the infrastructure provider's perspective, the desired objective is to achieve high system throughput and resource utilization, while minimizing energy consumption. We argue that providers can significantly lower operational costs and improve hosted application performance through application-aware virtual machine provisioning: using application characteristics, such as their data dependencies and resource usage patterns, in VM placement, migration, and resource management decisions. In this talk, Prof. Chandra will illustrate the notion of application-aware virtual machine provisioning through two systems that have been developed as part of his research. Starling is an affinity-aware VM provisioning framework that incorporates intra-application communication patterns to allocate virtual machines on the available physical resources. It dynamically adjusts VM placement to achieve both faster application runtimes as well as lower network communication overheads. STEAMEngine is a provisioning framework for virtualized MapReduce clusters that incorporates MapReduce job characteristics to optimize both cloud-level (e.g., energy) and user-level (e.g., performance) metrics. STEAMEngine exploits spatio-temporal tradeoffs in VM provisioning: efficient spatial fitting of VMs on servers based on their resource requirements, as well as balanced temporal fitting based on the estimated job runtimes