kentoh - Fotolia
NVMe can bring significant benefits in performance and power consumption today, but the new flash storage protocol is set for real take off in 2018.
Those are the views of the solid state drive product marketing executive for Samsung Semiconductor Europe, Kristian Vättö, who told Computer Weekly NVMe will take off as a storage medium in enterprise arrays with the arrival of dual port drives, next generation central processing units (CPUs) and NVMe-over-fabrics (NVMf).
The key problem highlighted by Vättö is that currently NAS and SAN have to switch between different protocols when they access storage, transitioning between small computer system interface (SCSI) and Ethernet and Fibre Channel.
“Because they are engaged in protocol switching, latency is added,” said Vättö. “That wasn’t a problem with hard disk drives when their latency was higher than that of the network.”
NVMe is an emerging flash storage protocol that can potentially speed access times by orders of magnitude compared to the SCSI and SATA protocols designed for spinning disk, but still in use for many flash drives.
But, hardware makers need to eliminate SCSI from the input/output (I/O) path and a bottleneck in storage processing to allow NVMe off the leash.
According to Vättö, NVMe’s most successful early adopters are hyperscale users. These are the likes of Google, Facebook etc that design their own hardware and software for use in their datacentres and don’t use proprietary storage platforms.
Read more about NVMe
- NVMe could boost flash storage performance, but controller-based storage architectures are a bottleneck. Does hyper-converged infrastructure give a clue to the solution?
- NVMe offers to unleash performance potential of flash storage that is held back by spinning disk-era SAS and SATA protocols. We run through the key NVMe deployment options.
Instead, they run huge clusters of servers with direct-attached storage (DAS) and typically, the server is not built with redundant components. In the event of failure, its workload is failed over to another server and the faulty device replaced. NVMe in these scenarios is slotted directly into server motherboards, thus avoiding the latency added by the storage controller.
“In hyperscale scenarios the cost of NVMe and SATA drives is similar but they can get 3x to 4x better performance from NVMe,” said Vättö.
“So, NVMe can reduce the numbers of servers needed, and this reduces power consumption. The Samsung PM963 NVMe drive, for example, has 2x the power consumption of a comparable SATA drive but 4x the performance.”
“With hyperscale customers that have hundreds of thousands of servers, they can reduce power use and make huge savings,” he said.
Enterprise customers not choosing hyperscale
But hyperscale isn’t in use by the bulk of enterprise customers. They generally buy their storage from established suppliers as preconfigured storage array products.
And so far there aren’t really any NVMe-based storage array products, at least not ones that have advanced features enterprise customers want such as data protection, encryption, data reduction and thin provisioning.
But, according to Vättö, 2018 will see a transformation in the market, with developments that will enable the productisation of NVMe-based storage arrays.
“So far, NVMe is in use by early adopters, like the hyperscale users, but next year three things will allow NVMe to take off,” he said.
These centre on the widespread availability of dual port NVMe drives, in other words, ones that can be built into the dual redundant paths used by enterprise storage architectures; features in Intel Skylake and AMD Epic CPUs, as well as NVMe-over-fabrics.
“The availability of dual port NVMe drives will allow traditional high-performance storage customers – such as banks – to use NVMe because it will be highly available,” said Vättö.
Optimised for NVMe
Regarding increased central processing unit capabilities in forthcoming upgrades, he said: “Previous platforms have not fully supported NVMe. They’ve had limited PCIe [peripheral component interconnect express] lanes, but these will be optimised for NVMe, with RAID for NVMe built in and allow NVMe to be more like SAS and SATA,” he said.
NVMe-over-fabrics will unleash the performance of NVMe in shared storage, freeing them from protocol switching. “NVMe products are sparse on the features side,” said Vättö. “They’re all aimed at specific niches which can do with fewer features. It needs NVMf to be mature on the features side, so I expect 12 to 18 months.”
“The big storage companies will be unwilling to release NVMf solutions until most features are available,” he said.