Modern block storage across Ethernet and RDMA fabrics
NVMe-oF extends the parallel NVMe command model beyond a local PCIe bus. Build shared storage for virtualisation, databases and latency-sensitive applications using NVMe/TCP or NVMe/RDMA—with the availability, multipathing and command support the host actually requires.
TCP or RDMA
Independent path
NVMe is the command set; the fabric carries it
NVMe-oF is networked block storage. The client discovers an NQN, connects to one or more NVMe controllers and receives a namespace that behaves like a local block device.
NVMe/TCP
Transports NVMe over routable IP networks and familiar Ethernet infrastructure. It is usually the most straightforward entry point.
NVMe/RDMA
Reduces data-path overhead on qualified RoCEv2 or InfiniBand fabrics. The complete network must be designed and validated end to end.
Block—not file
The host owns VMFS, another file system or the application data structure. SMB and NFS remain the correct choice when clients need shared files and directories.
Fast media alone does not define application performance. Queue design, network paths, target implementation, data protection, CPU placement and the host I/O stack all influence the result.
NVMe-oF complements rather than automatically replaces every SAN
Choose the protocol by workload, operations, availability requirements and existing investment—not by the newest name.
| Protocol | Command model | Network | Operational profile | Typical fit |
|---|---|---|---|---|
| NVMe/TCP | Native NVMe | Ethernet / IP | Familiar routing and monitoring | Modern shared block storage |
| NVMe/RDMA | Native NVMe | RoCEv2 or InfiniBand | Qualified low-latency fabric | Latency-sensitive I/O |
| iSCSI | SCSI | Ethernet / IP | Mature and widely understood | Established IP SANs |
| Fibre Channel | SCSI or NVMe/FC | Dedicated FC fabric | Specialised SAN operations | Existing deterministic enterprise fabrics |
Workloads benefit for different reasons
VMware and shared datastores
Parallel queues, shared namespaces and multipath connectivity support consolidated VM storage when host and target capabilities match.
eEVOS clusters
The integrated initiator discovers and consumes remote NVMe/TCP or NVMe/RDMA namespaces as shared storage for virtualisation hosts.
Databases
Low-latency block access can reduce storage overhead, while database consistency, queue depth and failover behaviour still require application-level validation.
Analytics and HPC
High concurrency and RDMA-capable fabrics can help data-intensive applications when the processing pipeline is not limited elsewhere.
Host support is only half of compatibility
VMware introduced NVMe/TCP storage adapters with ESXi 7.0 Update 3. New projects should use a currently supported ESXi release and validate the exact NIC, driver, firmware, switch and storage configuration.
Unlike iSCSI, NVMe-oF uses subsystems, controllers and namespaces through the native NVMe stack. Path state and load balancing must therefore be verified with the NVMe-aware host tools and the target’s availability model.
VMware version reference: Broadcom Knowledge Base.
Compare and Write must remain one atomic operation
NVMe defines Fused Compare and Write as two adjacent commands executed as an atomic unit. The write proceeds only when the compare succeeds. This optional controller capability is important to host workflows that depend on atomic storage updates.
+
Broadcom documents VMware snapshot processing that sends fused Compare and Write commands over NVMe/TCP. A target that does not implement the required command behaviour can therefore be unsuitable even when basic discovery and read/write I/O work.
euroNAS implementation: the NVMe-oF target implementations in euroNAS Premium, HA Cluster and eEKAS support Fused Commands. eEVOS provides the initiator side for consuming compatible NVMe-oF storage. Not every storage target implements this optional NVMe capability.
Technical references: NVM Express command-set specification and Broadcom VMware fused-command example.
Multipathing and storage availability solve different failures
Two network paths protect connectivity. They do not automatically make a standalone target highly available.
Availability must extend beyond the cable.
For node-level continuity, use an HA or scale-out target architecture and test the complete sequence: storage ownership, target movement, address activation, controller recovery and application response.
One protocol, four different deployment models
The products are not interchangeable. Select the availability and scaling model first, then qualify the transport.
| Platform | NVMe-oF role | Transports | Multipathing | Fused-command compatibility | Availability model |
|---|---|---|---|---|---|
| euroNAS Premium | Target and initiator | TCP and RDMA | Supported | Target support | Standalone storage server |
| euroNAS HA Cluster | Target and initiator | TCP and RDMA | Supported | Target support | Two-node mirrored or shared-storage failover |
| eEKAS | Target and initiator | TCP and RDMA where qualified | Supported | Target support | Ceph-backed scale-out storage and gateways |
| eEVOS | Built-in initiator; no target | TCP and RDMA | Supported | Initiator-side use of compatible targets | Virtualisation host or cluster consuming shared storage |
euroNAS Premium
Present NVMe-oF alongside file, iSCSI and Fibre Channel services on standard server hardware or as a virtual storage appliance.
HA Cluster
Coordinate protected storage, service addressing and NVMe-oF targets across two nodes for automatic failover.
eEKAS
Expose Ceph RBD-backed namespaces through scalable gateway services rather than a two-node storage mirror.
eEVOS
Discover and connect remote NVMe-oF namespaces through the integrated initiator for shared virtualisation storage.
Use RDMA when the complete fabric justifies it
RDMA is not a checkbox applied to ordinary Ethernet. Adapter, switch, MTU, VLAN, congestion handling, firmware and host configuration must work as one qualified system.
RoCEv2
Routable RDMA over converged Ethernet. It combines Ethernet hardware with an RDMA data path but needs deliberate loss and congestion design.
InfiniBand
A native RDMA fabric with its own link and subnet-management model, commonly used where low latency and fabric specialisation are accepted.
Software RoCE
Useful for validation and selected workloads, but it consumes host CPU and should not be presented as equivalent to hardware RDMA offload.
A dedicated RDMA technology page will cover RoCEv2, InfiniBand, software RoCE, switch design, NFS over RDMA and operational validation in greater depth.
Configure transports and verify target state without routine command-line work
These examples show transport selection and active target visibility in euroNAS Premium and HA Cluster.
Select TCP or RDMAThe target workflow presents the two transport options explicitly.
Premium target overviewReview namespace, NQN, transport, portal and controller connections.
HA Cluster target placementSee the serving node, access address, transport and active target state in one view.Specifications and platform guidance
NVM Express specifications
Current NVMe base, command-set and transport specifications.
VMware NVMe/TCP introduction
Broadcom confirms support was introduced with ESXi 7.0 Update 3.
VMware fused-command example
Broadcom documents snapshot I/O using fused Compare and Write over NVMe/TCP.
Define the workload and failure model before selecting the fabric
Tell us which hosts, applications, networks and availability requirements must work together. We can review whether TCP, RDMA, HA or scale-out architecture fits the project.
NVMe, NVMe-oF and related marks belong to their respective owners. VMware is referenced as a technical interoperability example; this page does not imply a VMware partnership or certification. Supported configurations depend on the current euroNAS release, licence, qualified hardware and validated client platform.