Reply to post: Everything eventually will go NVMe

Everything eventually will go NVMe

The biggest takeaway at Flash Memory Summit was the new form factor proposed by Samsung. Note it was all about cramming density into an enclosure, but also note the connection was NVMe (a modified M.2 interface).

All SSDs, regardless of media type (NAND TLC, NAND QLC, 3d-Xpoint, SCM, etc.) will use NVMe. Of course NVMe will be available for SCM non-volatile memory, but more traditional NAND will go that direction as well. The U.2 and M.2 connectors make it straightforward, the new CPUs have plenty of PCIe lanes, and what they do not have can be made up with PCIe switch ASICs.

Laptops and servers are already there (all NVMe).

All-Flash storage arrays will move to using NVMe based drives. AFAs will offer NVMe over Fabrics (NFMf) connectivity options alongside legacy SCSI based protocols. Users will be able to carve out a SCSI LUN or an NVMf Namespace as needed.

NVMe over Fabrics using FC (FC-NVMe) will likely take over Fibre Channel connectivity over the next 5 years. It is evolutionary, and can reside alongside SCSI over FC. As new OSs support FC-NVMe, they will connect using FC-NVMe. Legacy systems connect using SCSI.

Optane and other Storage Class Memory technologies will be used as a caching tier, perhaps as both a read and write caching tier. This may be necessary with QLC NAND to allow efficient use of the technology. I am guessing 99% of application storage requirements should be able to be met with SCM cached NAND based arrays. Pure SCM based external storage will remain an esoteric tier.

The only big question is the Ethernet world. NVMf over ROCEv2 seems the most likely solution. As I understand it, ROCEv2 supports IP addressing. But will it be as easy to use as iSCSI? That needs to happen. The industry needs a simple, straightforward way to connect to NVMf storage arrays.

FCoE is likely to finally die if NVMf over Ethernet gains traction. It may be available as a legacy connectivity technology, alongside iSCSI, but I can see no advantage to FCoE over NVMf over ROCEv2.