KVM Forum 2018

The public cloud users, especially enterprise users, are increasingly concerned about the stability of the infrastructure provided, therefore the performance impact of the live migration itself can no longer be ignored. In order to perform migration, the following needs to be considered: 1)How to classify the workload and evaluate the performance impact. 2)When should we migrate to reduce the impact to a minimum. 3)How to select params of the migration to balance the resource consumption and migrate speed. The ClouAtlas we build rely on KVM hypervisor to collect data like CPU usage, memory/TLB access and IO. Based on a statistical model and Deep Learning Algorithm, we classify the workload of the VM, and predict proper time to perform migration. After tens of thousands of tests, our system shows an obvious alleviation of the migration impact and a boost of the migration performance.