Colocation-Aware Modeling of CPU Usage for P2V Transitioning Applications

Traditional data-centers are giving way to virtualization based shared hosting platforms. This requires knowledge of how much resources are required to host a set of virtualized services. Due to the resource overhead incurred by virtualization, it is essential to estimate the virtual resource usage correctly, in order to avoid inefficiency due to excessive provisioning as well as prevent performance degredation due to over-aggressive multiplexing. Thus, the first and foremost issue that needs to be addressed to provision applications in virtual execution environments, is the mapping of resource requirements from physical to virtual environments. In a previous work, we have demonstrated the CPU savings when communicating VMs are placed in co-located manner, as opposed to dispersed. In this work, we apply this knowledge to the P2V (physical to virtual) transition for virtualizing applications and services. We present an automated system that can account for virtualization CPU overheads and predict the virtualized CPU requirement, given the physical resource usage usage levels. We develop prediction models for two virtualization technologies — Xen and KVM — to demonstrate the generality of the modeling approach. We further conduct extensive experiments with synthetic and application benchmarks to validate the models. Our experiments show that with synthetic benchmarks, the 90 percentile error is around 5% absolute CPU, for all workload types. With application benchmark testing, the 90 percentile is within 6%. Keywords-platform virtualization, application virtualization, capacity planning