Fast Local VM Migration Against Hypervisor Corruption

Abstract

Virtual machine migration is an established technique for tolerating hardware faults affecting the virtualization infrastructure. Normally migration is performed between different physical hosts and hypervisors, which requires the memory state to be eventually sent over the network, thereby causing performance degradation in the migrated and co-located virtual machines, particularly when the migrated VMs are running IO-and memory-heavy workloads. Since most of the hardware faults are transient and can be recovered from by refreshing the affected component, we propose and evaluate a technique for migrating virtual machines over the same physical host almost instantly and with no overhead, by avoiding memory copy and taking advantage of Intel EPT's inner workings. This technique can be employed for refreshing the VMs' state held by the hypervisor with a lower VM downtime and performance overhead than what would be possible using traditional live migration.