CPU-Feature-Aware Virtual Machine Migration in Heterogeneous Environments

Migrating virtual machines (VMs) across physically heterogeneous hosts remains brittle when source and destination processors expose different CPU feature sets. We propose a per-application, dynamic approach: a C-based runtime profiler executes on the source Xen dom0, instruments the running guest (domU), and computes the minimal set of CPU features actually exercised by the target workload. A companion shell script validates this set by iteratively disabling each feature and verifying that the workload remains functional. Validated on two Intel hosts with differing micro-architectures using a PostgreSQL 12 OLTP benchmark, reducing the exposed flag count from 47 raw CPUID flags to 12 essential flags, demonstrating successful migration with zero application-level errors.

Open-source implementation: github.com/hashirama21/xen-cpu-aware-migration

Recommended citation: Vincess Dongmo, Tchudjing Ruben, Nguekeng Marc Donald, Ngom Mbock Michel, Tsanga Zo Onyaba, Nnange Akume. (2023). "CPU-Feature-Aware Virtual Machine Migration in Heterogeneous Environments." Preprint, Polytechnique Yaoundé.
Download Paper