VMware to OpenStack migration with minutes of downtime

Broadcom completed its acquisition of VMware in November 2023 and moved vSphere to subscription bundles. Since then many teams have planned an exit to open infrastructure. OpenStack gives VMs the standard APIs: Nova for compute, Cinder for volumes, Neutron for networks and Glance for images, whether the cloud is RHOSO 18.0, with the control plane on OpenShift and Red Hat's life cycle, or a community release deployed with Kolla-Ansible or another tool.

The hard part of the exit is the move itself: hundreds of VMs, each with an owner who wants to know when it will be down and for how long. Seamless Migrate turns that into a plan: waves with a pilot first, a downtime estimate for every VM before anything moves, an approval and a change window for every cutover, and a rollback until you finalize. Weighing it against MigrateKit, vJailbreak or Red Hat's toolkit? See the migration tools compared.

Warm downtime follows the last CBT delta and the in-place conversion, so it barely grows with the disk. A cold copy keeps the VM stopped for the whole transfer.

Try your own VM, with its disks, used space, link speed and write rate, in the downtime calculator.

1. 01

   Connect vCenter and the destination

   Add both providers with their presets: vCenter as the source, RHOSO 18.0 or a community OpenStack cloud as the destination. Credentials are write-only and go to the platform's secret store. The connection check reports what each side can do, CBT support included.

2. 02

   Discover and plan

   The inventory lists every VM with its disks, NICs, guest OS and its CBT, snapshot and independent-disk state. Map networks, flavors and volume types, then let the planner order waves by risk with a pilot first.

3. 03

   Validate

   Pre-flight checks flag what would fail, each with its fix: CBT disabled, independent disks, a guest virt-v2v can't convert. Every VM gets an estimate for both strategies and an automatic choice.

4. 04

   Pre-copy while the VM runs

   Warm VMs copy in CBT passes while they keep serving traffic, until their change window opens.

5. 05

   Cut over with an approval

   Inside the change window, with a second approver's approval and a free cutover slot, the VM stops, the final delta moves, virt-v2v converts the guest in place and the VM boots on the destination.

6. 06

   Verify, roll back or finalize

   Seamless checks that the server is ACTIVE, its ports and its TCP probes. A failed check deletes the destination server and powers the VMware VM on again. Finalize, with an approver and the VM's name typed in full, closes the migration.

VMs with an NVIDIA vGPU, time-sliced or MIG-backed, or with a GPU passed through with DirectPath I/O migrate as a Technology Preview, validated in simulation. Seamless reads the vGPU profile, such as grid_a100-4c, or the device's ids, such as 10de:20b5, from vCenter, and the plan maps each to a destination GPU flavor in mappings.gpus; an unmapped GPU VM stays blocked. Validation counts the GPUs against the destination's free capacity in Placement, and the cutover gate reads it again before the VM stops: with no room, the VMware VM keeps running and the migration waits.

A vGPU VM keeps both VMware strategies. A VM with a passthrough device moves only with vmware_cold, powered off, because vSphere takes no snapshot of a VM with a DirectPath I/O device. No GPU state moves: the VM boots on the destination as after a shutdown, and GUEST_GPU_DRIVER and GUEST_GPU_LICENSE name the NVIDIA driver and the license server its guest needs there.

When a VM's address changes on the destination network, a DNS swap, also a Technology Preview, moves its record in Designate or an RFC 2136 server by compare-and-swap once the server is created, and swaps it back on a rollback. Hooks an admin registers, Ansible playbooks or HTTPS webhooks, run at fixed points of each migration for the rest. Neither has run on GPU hosts or real clouds in the lab yet.