TORONTO — Following on the heels of a major specification update and its eighth annual plug fest, NVM Express is poised to have a busy year as it continues to develop the base NVMe specification while expanding the NVMe Management Interface (NVMe-MI) specification and one for accessing SSDs on a PCIe bus over fabrics.
In June, the NVMe specification got its first major update in nearly three years, putting it on the cusp of becoming the defacto standard for SSD interfaces. Version 1.3 added a significant number of new features, something that hasn't been done since November 2014, encompassing 24 technical proposals spread across three major buckets that address client, enterprise and cloud features. Most significant was improved support for virtualization so developers can more flexibly assign SSD resources to specific virtual machines, thereby addressing latency.
Meanwhile, the eighth NVMe Plugfest at the University of New Hampshire Interoperability Laboratory last fall offered the first official NVMe Over Fabrics (NVMe-oF) compliance and interoperability transport layer testing for RoCE, Remote Direct Memory Access (RDMA) over Converged Ethernet, and the Fibre Channel. UNH-IOL fills the role of independent testing provider of standards conformance solutions and multi-vendor interoperability, and the latest plugfest generated 14 new certified products for the base NVMe integrators list and one for the NVMe-MI integrators list. Eight inaugural products were also approved for the newly launched NVMe-oF integrators list, which accepts RoCE initiators and targets, Ethernet switches, as well as Fibre Channel initiator, targets and switches and software.
The NVMe-oF specification was first published in 2016 to extend NVMe onto fabrics such as Ethernet, Fibre Channel and InfiniBand, and provide access to individual NVMe devices and storage systems. It's set to get a further update by mid-year, according to David Allen, NVMe board member and Seagate's senior director of marketing. NVMe-oF 1.1 will include enhanced discovery, in-band authentication and TCP transport binding. He said many of the high availability features that are part of SCSI and Fibre Channel have become integrated into current SSDs.
Updates to the NVMe-oF this year will support all of the NVMe-oF and NVMe Architecture features, and layered over standard IETF TCP transport to allow software-only and/or hardware (accelerated/offloaded) implementations.
Micron, however, opted in May to announce its own NVMe over Fabric architecture ahead of standards development, introducing its SolidScale architecture for low-latency, high-performance access to compute and storage resource. Despite the standard being immature, Micron said at the time that the SolidScale software-defined storage architecture will provide the benefits of shared storage with performance of server-side flash and address CPU underutilization. It connects multiple nodes using high-speed RDMA over Converged Ethernet (RoCE) Mellanox fabric with low-latency software to deliver a converged infrastructure that performs like local direct attached storage.
As of December, TCP transport binding was being defined with the NVMe.org Technical Working group, while Linux Host and Target Drivers are being developed within the Fabric Driver Working Group. The plan is to co-release the specific and tested Linux drivers as part of the next NVMe-oF release. The NVMe-TCP workgroup project will look at enabling the use of NVMe-oF over existing datacenter IP networks. Updates will support all the NVMe-oF and NVMe architecture features, and layered over standard IETF TCP transport to allow software-only and/or hardware (accelerated/offloaded) implementations.
Meanwhile, NVMe 1.4 is slated for release sometime in 2019 and will support IO determinism, said Peter Onufryk, NVMe board member and NVMe-MI workgroup chair, which is gaining a considerable amount of industry attention. It enables hosts to treat an SSD as many small sub-SSDs and process IO in parallel in each small sub-SSD. This means host threads can process IO independently in small sub-SSDs without blocking from other thread IOs. He said IO determinism can reduce average read latency significantly for higher performance and for better quality of service (QoS) due to the parallel execution of IOs without any conflict to the media.
NVMe 1.4, slated for release sometime in 2019, will support IO determinism and enables hosts to treat an SSD as many small sub-SSDs and process IO in parallel in each small sub-SSD.
As NVMe is now fronting major storage arrays, Onufryk said NVMe 1.4 will support NVMe Multipathing and Namespace Sharing, where two or more hosts access a common shared namespace using different NVM Express controllers .
Given all of the different updates and features, Allen said the goal of NVM Express is to maintain original benefits NVMe while scaling and adding management capability, such as the enclosure management in NVMe MI 1.1 that includes out-of-band management support to operate hardware resources and components that are independent of the operation system control.
NVM Express will host a webcast on January 17 at 9:00 a.m. Pacific Time outlining the new features to all three specifications.
NVM Express first began as a workgroup with its first specification published in 2011. In 2014 the group incorporated as NVM ExpressInc. and is governed by an elected board of directors comprised of representatives from vendors including Facebook, Dell, Intel, Micron and Western Digital. Its work is split into four workgroups — the technical workgroup handles NVMe Base and NVMe Over Fabrics; the Management Interface Workgroup covers out-of-band-management over SMBus and PCIe VDM; and the Interop Group handles interop and conformance testing in collaboration with UNH-IOL. A fourth group handles of all its marketing initiatives.
— Gary Hilson is a general contributing editor with a focus on memory and flash technologies for EE Times.