3.5 RDM, PMem, NVMe, NVMe-oF & iSER Use Cases
Key Takeaways
A raw device mapping (RDM) is a mapping file on a VMFS datastore that points a VM at a whole LUN; physical compatibility mode passes SCSI commands through and does not allow VM snapshots.
Physical-compatibility RDMs suit SAN-aware applications and Windows Server Failover Clustering across hosts, while vSphere 7.0 and later also offer clustered VMDKs on VMFS-6 for WSFC.
ESXi supports NVMe over Fabrics using Fibre Channel, RDMA (RoCE v2), and TCP transports, and claims NVMe devices with the VMware High-Performance Plug-in (HPP).
iSER (iSCSI Extensions for RDMA) runs iSCSI over RDMA-capable NICs to cut CPU use and latency, and is enabled with esxcli rdma iser add.
vSphere 8 supports Intel Optane PMem as vPMem (a virtual NVDIMM for PMem-aware guests on hardware version 14+) or vPMemDisk; PMem support ends with VMware Cloud Foundation 9.0.
3.5 RDM, PMem, NVMe, NVMe-oF & iSER Use Cases
Most VMs store their disks as VMDK files on VMFS, NFS, vSAN, or vVols. Objective 1.3.6 covers the special-purpose options you pick when a workload needs direct LUN access, persistent memory, or the lowest-latency transports.
Raw Device Mapping (RDM)
An RDM is a mapping file (a small .vmdk pointer) on a VMFS datastore that gives a VM direct access to a whole LUN. The VM's I/O goes to the LUN, while the mapping file lets vCenter manage it like a disk.
| Virtual Compatibility Mode (vRDM) | Physical Compatibility Mode (pRDM) | |
|---|---|---|
| SCSI commands | Virtualized like a VMDK | Passed through to the device (except REPORT LUNs) |
| VM snapshots | Supported | Not supported |
| Can be converted to VMDK with Storage vMotion | Yes | No (only the mapping file moves) |
| Typical use | Features of a VMDK plus a dedicated LUN | SAN-aware applications and clustering |
Use cases for RDMs:
- Windows Server Failover Clustering (WSFC) across hosts with shared disks that use SCSI-3 persistent reservations (pRDM).
- SAN management or array-aware software in the guest that must send commands directly to the LUN, such as array snapshot agents.
- Physical-to-virtual flexibility, where a LUN may need to be presented to a physical server again later.
Since vSphere 7.0, clustered VMDKs on VMFS-6 datastores (with eager-zeroed thick disks) are an alternative to pRDMs for WSFC, which simplifies management. NPIV, which gives a VM its own Fibre Channel WWPNs on an RDM, is being retired: the vSphere 8.0 release notes announce its deprecation in a future release.
NVMe and NVMe over Fabrics
NVMe is a storage protocol designed for flash, with deep parallel queues. vSphere uses it in three places:
- Local NVMe devices: PCIe SSDs in the host, often used for vSAN (ESA requires NVMe) or local VMFS.
- NVMe over Fabrics (NVMe-oF): NVMe commands carried over a network to an external array.
- Virtual NVMe controller: a VM-side controller type (hardware version 13 and later) that lowers guest I/O overhead. It is independent of the physical transport.
ESXi's NVMe-oF transports:
| Transport | Adapter Type | Notes |
|---|---|---|
| NVMe over Fibre Channel (FC-NVMe) | Hardware FC HBA that supports NVMe | Appears as a Fibre Channel adapter (vmhba) whose protocol is NVMe; no extra adapter configuration |
| NVMe over RDMA (RoCE v2) | Software NVMe over RDMA adapter on an RDMA-capable NIC | Needs a lossless or well-tuned RoCE network; add controllers after creating the adapter |
| NVMe over TCP | Software NVMe over TCP adapter on a standard NIC | Works on ordinary Ethernet with no special NIC; add controllers after creating the adapter |
Key points:
- NVMe devices present namespaces rather than LUNs, and you connect to array controllers through discovery.
- ESXi claims NVMe devices with the VMware High-Performance Plug-in (HPP) instead of the Native Multipathing Plug-in used for SCSI. HPP has its own path selection schemes, such as load balancing by round robin or by latency.
- Use cases: latency-sensitive databases and analytics on all-flash arrays, or moving off FC SCSI to NVMe without changing the fabric (FC-NVMe), or using existing Ethernet (NVMe/TCP).
iSER: iSCSI Extensions for RDMA
iSER keeps the iSCSI model (targets, IQNs, port binding) but moves the data transfer onto RDMA, using RDMA-capable NICs (RoCE). This cuts the host CPU work and latency of software iSCSI.
- Enable the adapter from the ESXi shell with
esxcli rdma iser add, which creates a VMware iSER adapter (a vmhba). - Bind it to a VMkernel adapter on an RDMA-capable NIC, then configure discovery and targets as with iSCSI.
- Use case: an iSCSI array that supports iSER, where you want lower latency and CPU cost without moving to a new protocol.
Persistent Memory (PMem)
PMem (Intel Optane persistent memory) sits between DRAM and flash: byte-addressable like memory, and persistent like storage. vSphere 8 supports it in two ways:
| Mode | How the VM Sees It | Requirements |
|---|---|---|
| vPMem | A virtual NVDIMM device, bypassing the storage stack | Hardware version 14 or later and a PMem-aware guest OS |
| vPMemDisk | A virtual disk stored on the host-local PMem datastore | Works with legacy guests that are not PMem-aware |
PMem can also run in Memory Mode, where it transparently expands host memory capacity.
- Use cases: in-memory databases and applications that need very fast restarts or persistent in-memory structures.
- Constraints: PMem is local to each host, so VMs using it need destination hosts with PMem for vMotion, and HA restarts need spare PMem.
- Status: Intel discontinued Optane, and Broadcom ends PMem support starting with VMware Cloud Foundation 9.0. It is still in the vSphere 8 blueprint, so know the modes.
Choosing Among Them
| Requirement | Best Fit |
|---|---|
| Shared-disk WSFC across hosts | pRDM, or clustered VMDK on VMFS-6 (vSphere 7.0+) |
| Guest software must issue SCSI commands to the array | pRDM |
| Lowest-latency external flash over existing Ethernet | NVMe over TCP |
| NVMe on an existing Fibre Channel fabric | FC-NVMe |
| Lower CPU cost for an iSCSI array that supports RDMA | iSER |
| PMem-aware in-memory database | vPMem |
Exam Traps
- pRDM means no VM snapshots. Backup tools that rely on VM snapshots cannot protect a physical-mode RDM, so plan array-level protection.
- NVMe devices are not claimed by NMP. If a question asks which plug-in manages NVMe paths, the answer is the High-Performance Plug-in (HPP).
- NVMe/TCP needs no special NIC; NVMe/RDMA and iSER do. Match the transport to the hardware in the question.
- PMem is host-local. Plan vMotion and HA capacity accordingly, and remember that support ends with VMware Cloud Foundation 9.0.
A database team needs to run a Windows Server Failover Cluster whose nodes are on different ESXi hosts and share a disk that uses SCSI-3 persistent reservations. The array admin will present a dedicated LUN. Which disk configuration fits?
A thin-provisioned VMDK on an NFS 3 datastore
A virtual-compatibility RDM with VM snapshots enabled
A physical-compatibility RDM to the shared LUN
A vPMemDisk on each host's PMem datastore
An organization wants NVMe performance from a new all-flash array but has only standard Ethernet NICs and switches, with no RDMA or Fibre Channel. Which ESXi configuration should be used?
A software NVMe over TCP adapter
A software NVMe over RDMA adapter
An iSER adapter created with esxcli rdma iser add
A hardware FC-NVMe HBA
Which requirements apply when presenting Intel Optane persistent memory to a VM as vPMem (a virtual NVDIMM)?
A physical-compatibility RDM and SCSI-3 reservations
An iSER adapter bound to an RDMA-capable NIC
Any guest operating system and hardware version 10
Virtual hardware version 14 or later and a PMem-aware guest operating system
Sections you finish are checked off in the contents.