NVMe: The End of Storage Networking as We Know It

If you have flash storage in your data center, you’ve no doubt wondered what the point is to having a high-speed medium on one end tied to high-speed processing on the other while in-between is a low-speed SATA or SCSI interface.

For a while, the reason behind this was that there was no viable alternative. Once something did come along, it was very expensive. But now that costs for the Non-Volatile Memory Express (NVMe) interface are close to parity with earlier solutions, it seems the floodgates are open for a faster and more efficient data center network infrastructure.

NVMe is the best way to unlock the true potential of solid state storage, particularly the 3D NAND flash devices that are making their way into the data center, says George Chacko, tech consultant with Brocade India. NVMe was designed for the low latency and parallelism of solid state storage and the corresponding CPU on the processing side. As such, it reduces much of the I/O overhead of existing storage and memory solutions and removes many of the performance-hampering artifacts of previous interfaces, such as lengthy command queues. The result is a dramatic acceleration of overall performance and tighter, more modular hardware configurations.

Expect NVMe deployments to kick into high gear once native support on the processor level hits the channel, which IBM’s Eric Herzog says should happen later this year. The technology provide not only better I/O performance, but also better utilization of processor cores by allowing direct access to on-server memory from remote sources. And if more streamlined networking is not enough of an incentive to deploy NVMe, with the CPU no longer the gateway to stored data, application licensing costs should go down because fewer cores are required to support a given app.

Perhaps even more interesting, though, is the fact that NVMe won’t replace existing storage network solutions like Ethernet and Fibre Channel, but will in fact help convert them to advanced fabrics. Cavium Networks just announced that its QLogic Gen 6 Fibre Channel and FastLinQ Ethernet adapters will support the NVMe over Fabrics (NVMe-oF) protocol for scale-out storage environments. In this way, the company says it can implement fabric topologies without ripping and replacing legacy infrastructure and provide support for concurrent implementations of networking solutions like RDMA, RDMA over Converged Ethernet (RoCE) and iWARP. NVMe-oF will initially appear on the 2690 Gen 5 and 2700 Series Gen 6 32G FC adapters, as well as the 100 GbE FastLinQ 45000 series.

In addition, Micron just released its new SolidScale Platform Architecture that uses NVMe over Fabrics to provide shared, accelerated storage for both data center and cloud-native applications. The system connects multiple NVMe storage nodes using RoCE, which the company says provides the performance of direct attached storage across converged infrastructure footprints. At the same time, Micron has added an intuitive Web-based management interface that provides graphical setup and configuration for advanced data services.

The prevalence of NVMe and fabric technology in general, coupled with the rise of on-server memory and modular infrastructure, is slowly but surely bringing an end to specialized server and storage network architectures. In the future, it will just be known as plain vanilla networking as data is shuttled between integrated computing nodes situated across local and cloud-based data facilities. This will certainly make things a lot easier, because it reduces the multitude of systems, protocols, formats and connection points between one data source and another.

NVMe might not be the final answer to the enterprise’s networking woes, but it certainly is a step in the right direction.

Arthur Cole is a freelance journalist with more than 25 years’ experience covering enterprise IT, telecommunications and other high-tech industries.

Latest Articles

Follow Us On Social Media

Explore More