Why Use NVMe-over-Fabrics for vSphere 7 Virtual Machines?

By democratizing virtualization, VMware has forever changed the data center. Long gone are the days of one application running on one physical server. In the modern data center, compute and networking resources are shared, with virtual resources being made available to the applications that need them when they are needed.

For all the benefits of virtualization in the data center, one of this technology's biggest challenges has been storage. Over the years, storage has both enabled and constrained what can be done with virtual machines. Storage area networks (SANs) became the standard for sharing block-level data because they solved many of the problems associated with direct-attached storage (DAS). Moving away from DAS, which delivered poor resource utilization, created islands of storage, and significantly constrained capacity, was an easy choice.

The rise of SANs was important to the development of VMs. The ubiquity of shared SAN storage made the adoption of vMotion and other virtualization technologies far easier. SANs have greater reliability than DAS, can be easier to scale, and provide multipathing to easily migrate VMs from one physical server to another. Yet for all the benefits of SANs, they were limited in the performance that they could deliver. Storage network latency far exceeded that of DAS, creating a barrier to using VMs for latency-sensitive applications. IT managers wanted to take advantage of the benefits VMware offered, yet struggled to do so for latency-sensitive applications such as high-performance databases.

To solve this, vendors introduced various forms of Hyperconverged Infrastructure (HCI). When combined with vSphere, this enables management of storage and compute from within a single platform to become much easier. HCI aggregates all of the DAS storage within a cluster of independent systems and to create a shared virtual pool across all the systems. This is a very effective way to deliver the performance of DAS while simultaneously leveraging some of the benefits of a SAN.

Still, for all the benefits that HCI offers, some of the challenges inherent to DAS remain, such as the difficulty of scaling storage and compute independently of each other. What is needed is a way to have shared storage separate from compute, but without the performance and latency issues of a legacy SAN.

To enhance the performance of all-flash arrays (AFAs), VMware has added support for NVMe-oF in the recently announced vSphere 7. Adding support for NVMe-oF enables VMware to offer its customers the best of both worlds. Data centers can still aggregate compute resources to maximize processing, while also getting the benefits SANs provide with the additional performance of all-flash via NVMe-oF.

NVMe-oF extends NVMe across a fabric to enable extremely high performance and low latency communication to an external array. Since NVMe and NVMe-oF were designed from the ground up for flash, there is no unnecessary overhead from legacy interfaces.

Now that VMware has support for NVMe-oF, customers are well-positioned to take advantage of the next generation of high-performance all-flash storage, such as the Pavilion Hyperparallel Flash Array (HFA).

The Pavilion HFA can deliver significantly better performance than a traditional AFA. A typical AFA will usually deliver about 1M IOPS at 1 ms latency. In contrast, the Pavilion HFA can provide 20M IOPS at only 40μs. Combined with support for RDMA over Converged Ethernet (RoCE), the Pavilion HFA can move data into system memory with almost no added latency, meeting the needs of even the most demanding applications.

Now there are no barriers to using VMware across all applications in the data center.