NVMe SSD virtualization reaches locally attached PCIe drives
NVMe’s updated specifications extend SSD virtualization and VM migration to local PCIe drives, but product support is still unconfirmed.
By Dominic Okoye · Staff Writer
· 3 min read
The NVM Express consortium has updated all 11 specifications in its NVMe suite, adding a mechanism for NVMe SSD virtualization on locally attached PCIe drives. The feature, PCIe Exported NVM Subsystem Migration, aims to let virtual machines move between servers while retaining a consistent view of their storage, without reconfiguring storage inside the VM.
NVMe is the protocol used by flash storage and SSDs for parallel data transfer with lower per-I/O overhead. The update is an architectural change for virtualized servers, not evidence that existing SSD fleets or hypervisors can use the capability today. The materials do not identify supported drive products, operating systems, hypervisors or cloud platforms, nor give an availability timetable.
How does NVMe SSD virtualization work on local PCIe drives?
Exported NVM subsystems were already part of NVMe over Fabrics, which connects storage across a network. The new specification extends that abstraction to PCIe-attached SSDs installed locally in a server, according to NVM Express.
Rather than exposing physical SSD controllers and namespaces to a VM, a drive can present exported controllers and namespaces. The host defines that virtual storage view and controls its exported identifiers. On a destination server, those identifiers can be recreated even if the replacement hardware uses different internal identifiers, NVM Express said in a technical blog by Samsung executive and NVM Express board member Mike Allison.
In practical terms, a migration flow described by NVM Express calls for an administrator or host to create an exported subsystem on the destination, reproduce the controllers and namespaces with the same identifiers, apply matching templates, and copy runtime state before moving the VM. The source-side exported resources can then be removed.
What moves from the hypervisor into the SSD?
Samsung says conventional storage virtualization is commonly handled in the hypervisor or host software. That work can include namespace mapping, isolation, migration and interception or emulation of administrative commands. Direct assignment may also reveal the underlying drive’s identity and other subsystem information to the guest, Samsung said.
Under the new model, templates tell the SSD which capabilities are visible, which settings are fixed or configurable while running, and the format of migration data. NVM Express says the device can use those templates to enforce the VM-visible storage behavior, allowing the VM to send administrative commands to the exported subsystem rather than relying on host interception.
NVM Express and Samsung say that division of labor could reduce hypervisor complexity, improve isolation and support more consistent migrations. Those are vendor and standards-body claims; the available material includes no independent latency, scalability or interoperability results.
The wider specification refresh also adds support for post-quantum cryptography, controller-level bandwidth and IOPS rate limiting, voltage monitoring, and a factory-default restore function. It includes NVMe 2.4, NVM Express Management Interface 2.2, NVMe Boot 1.4, and updated PCIe, RDMA and TCP transport specifications.
For infrastructure teams, the immediate significance is standardization: locally attached SSDs now have a specified path to participate in storage-aware VM migration. Whether that path becomes operational depends on implementation across drives and the virtualization stack.
This story draws on original reporting from The Register.
