Some Problems Translating application QoS into VM configurations Quantifying the effects of interference and affinity Placement strategies
Generating VM Configs Input: Application performance characterization leading to: Building and solving predictive performance models for the application QoS operating ranges and estimated load patterns Output: A set of VM configurations along with a mapping of application components to VMs.
Challenges: Automating characterizing appl perf. For virtualized env. Extending standard performance prediction techniques (Queuing) to include the effects of virtualization
Interference/Affinity VMs dont provide performance isolation VMM takes up some percentage of resources. Given an application components performance on a single VM, can we estimate the effect the colocating other VMs running different types of workloads (CPU intensive, I/O intensive etc.). Further, can we characterize this effect with changing parameters of the interfering component?
Results - 2 Ping latency of a VM doubled when when it was deployed with a mixture of CPU-intensive and I/O bandwidth intensive VMs, as compared to when it was deployed with only I/O bandwidth intensive VMs.
VM Placement Given a current deployment and the set of VMs that need to be deployed in a data center, output a plan of placing the VMs on the existing physical machines to optimize number of physical servers used and other application specified constraints (for fault tolerance etc.) Migration costs A multi dimensional bin packing problem subject to various constraints: Interference conflicts App driven conflicts