Oracle engineers have published a technical account of how Exadata release 26.1 can live migrate database virtual machines between servers. The post appeared on the Oracle Linux blog on November 12, 2026. It describes the QEMU and kernel work needed to move a running VM that talks over RDMA.
Live migration has long been a requested Exadata feature. It helps with load balancing and with maintenance, because a VM can leave a server without being shut down.
Why Exadata Is Hard to Migrate
Exadata database servers and storage servers are linked by a network that supports Remote Direct Memory Access. On current systems this is RDMA over Converged Ethernet with Nvidia ConnectX-7 adapters. The VMs use virtual functions of those adapters, so they hold device state that must move with them.
Downtime is the main risk. If a VM stops answering for too long, a protocol timeout can fire, or Real Application Clusters can decide the VM has died and start cluster recovery. The RDMA virtual functions can carry 20 megabytes or more of runtime state to copy in the final stage.
Key Techniques in the Post
- Postponing timeouts: while a VM cannot handle RDMA work, the system uses the Linux FWCTL interface to tell adapter firmware to push back queue pair timeouts. If migration still fails, it is cancelled and the source VM resumes.
- Parallel device state transfer: a feature added to QEMU 10.0 by Oracle engineer Maciej Szmigiero sends each VFIO device's state on its own thread. With four ConnectX-7 functions and about 20 MB each, this cut that part of the final stage from 1783 ms to 1068 ms.
- Hardware dirty-bit tracking on AMD: instead of trapping writes, KVM scans nested page table bits. Oracle reports the older method topped out at 11 GB/s per CPU in its tests, short of its needs.
- IOMMUFD: Oracle engineers contributed to this upstream kernel API, which lets QEMU use IOMMU hardware to track pages written by RDMA devices.
- Hash-based page comparison: a 2MB page marked dirty is checked in 4KB pieces by comparing hashes, so only truly changed pieces are sent.
How Dirty Page Tracking Was Solved
Oracle first tried the adapter's own dirty page logging. It allowed only four virtual functions at once and slowed DMA performance, which hurt VMs that were not migrating. So the team dropped it in favor of the IOMMU approach.
Tracking 2MB pages means far fewer entries to scan, but one small write dirties the whole page. Sending it as 512 small pages made migrations fail to finish on busy database VMs. The hashing step fixes that, at the cost of CPU time, cache traffic and memory bandwidth.
The authors say this is a working solution, not a permanent one. They expect more efficient tracking from processor and adapter vendors and from newer kernels.
What It Means for DBAs and Engineers
For DBAs, the practical gain is operational: moving database VMs for maintenance without stopping them, with RAC protected from false failure signals. The post gives no migration timings for end users, so plan tests in your own environment.
For systems engineers, it is a rare look at real limits in virtualization with passthrough devices. The same ideas, such as parallel device state and IOMMU dirty tracking, apply to other KVM and QEMU deployments. For related Exadata news, see the fixed monthly Exadata update schedule.
Conclusion
Exadata 26.1 adds live migration of database VMs, backed by changes in QEMU, KVM and the Linux kernel. If you run Exadata, check the 26.1 documentation for supported configurations before planning migrations.
Sources
See my articles more often when you search Google.
