Smart VM co-scheduling with the precise prediction of performance characteristics

Abstract Traditional virtualization systems cannot effectively isolate the shared micro-architectural resources among VMs. Different types of CPU and memory-intensive VMs contending for these shared resources will lead to different levels of performance degradation, which decreases the system efficiency and Quality of Service (QoS) in the cloud. To address these problems, we design and implement a smart VM co-scheduling system with precise prediction of performance characteristics. First, we identify the performance interference factors and design synthetic micro-benchmarks. By co-running these micro-benchmarks with VMs, we decouple two kinds of VM performance characteristics: VM contention sensitivity and contention intensity. Second, based on the characteristics, we build VM performance prediction model using machine learning techniques to quantify the precise levels of performance degradation. By co-running large numbers of different VMs and collecting their performance scores, we train a robust performance prediction model. Finally, based on the prediction model, we design contention aware VM scheduling algorithms to improve system efficiency and guarantee the QoS of VMs in the cloud. Our experimental results show that the performance prediction model achieves high accuracy and the smart VM scheduling algorithms based on the prediction improves system efficiency and VM performance stability.