Getty Images/iStockphoto
Look for PCIe 5.0, NVMe 2.0 to boost storage performance
Enterprises that need high-performance storage for AI and ML workloads can keep an eye out for new products based on the latest PCIe 5.0 and NVMe 2.0 specifications.
Enterprise buyers in need of higher performance storage might want to start checking with vendors about their plans to support PCIe 5.0 and NVMe 2.0 technologies.
The NVM Express Inc. organization recently finalized the 2.0 version of the non-volatile memory specification that can help to lower latency and boost storage performance. New NVMe 2.0 features include command sets for key-value (KV) and Zoned Namespace (ZNS) technologies that can run on PCIe, remote direct memory access (RDMA) and TCP transports.
Updating from the PCIe 4.0 to PCIe 5.0 interconnect doubles the data rate from 16 gigatransfers per second to 32 GT/s, and the aggregate bandwidth from close to 64 GBps up to 128 GBps with a 16-lane connection. The extra throughput could be especially helpful with compute-intensive workloads such as artificial intelligence and machine learning (ML).
Products in '22 and '23
But enterprise IT shops will have to wait for products that support the technologies. For instance, the first PCIe 5.0 SSD controllers from Marvell Technology Inc. did not begin sampling to customers until this year, well after the PCI-SIG released PCIe 5.0 in 2019. Marvell, headquartered in Wilmington, Del., does not expect SSDs that use its PCIe 5.0-based Bravera SC5 controllers to hit until PCIe 5.0-based servers reach volume production in the second half of 2022 and 2023, a company spokesperson said.
Intel disclosed in June that its next-generation Xeon Scalable processor, code-named "Sapphire Rapids" -- which is due to support PCIe 5.0 -- will reach production in the first quarter of 2022 and ramp in volume beginning in the second quarter of 2022. AMD has yet to confirm details on its PCIe 5.0 plans.
That stands in contrast to what happened with PCIe 4.0. After PCIe 4.0 was finalized in 2017, AMD got a substantial jump on Intel with support for PCIe 4.0 in its Ryzen desktop and Epyc server processors in 2019. Intel didn't ship processors that support PCIe 4.0 until this year.
Tom Coughlin, president of Coughlin Associates, said, even though the initial PCIe 5.0 products may emerge in 2022 and 2023, they might not become mass market until 2025 or 2026. Coughlin said he expects NVMe 2.0 products could start to emerge late this year, with more in 2022.
Jim Handy, semiconductor analyst and general director of Objective Analysis, noted that NVMe 2.0 goes beyond the normal evolutionary change, with the addition of support for ZNS and key-value store. He cautioned that NVMe 2.0-based SSDs will require application software changes to support the ZNS and KV features in order to provide a substantial benefit to users.
NVMe 2.0 changes
On the plus side, SSD manufacturers will be able to develop a single device that IT pros can use with one namespace for ZNS access, another namespace for KV access, and yet another namespace for NVMe-based block access to the device, noted officials of the nonprofit NVM Express Inc., a consortium that defines and manages NVMe technology.
With the 2.0 release, the NVM Express organization broke the specification into the NVMe Base specification, Command Set specifications and Transport specifications to enable the development of new command sets and new transports, without impacting those that have already been developed, officials said.
The multiple transport specifications will also allow devices to connect to the host server in a variety of ways. For example, an NVMe-over-Fabrics SSD could use the NVM command set and TCP transport, or an NVMe SSD could combine the NVM command set and PCIe transport. Another SSD could provide the ZNS and the KV command sets on different namespaces over PCIe.
Potential ZNS benefits
Western Digital and Microsoft took the lead in bringing the ZNS proposal to NVMe Express. The ZNS command set targets applications that write sequentially, with the host operating system placing data into independent zones on an SSD. The main benefits include reducing overprovisioning to enable higher capacity, lessening garbage collection to facilitate more predictable performance and, potentially, using less DRAM and power-loss protection capacitors to drive down costs, noted Dave Landsman, an NVM Express board member who is director of industry standards at Western Digital.
ZNS SSDs are already in the pipeline, notably from Western Digital and Samsung. But Eric Pike, senior director of cloud enterprise flash marketing at Western Digital, said ZNS would be a "long rolling implementation," as vendors work to marry software stack development with ZNS drives.
"Those two pieces need to come together to get the capabilities," Pike emphasized, noting that ZNS is "all about an efficient way of handling data" with structured data sets for transactional workloads. KV, by contrast, focuses on unstructured data storage, Pike said.
Key-value advantages
The KV proposal came to NVM Express Inc. from Samsung and NetApp. One of the main benefits of KV is faster data retrieval based on a group of keys to identify the data. Prior to KV, objects used a host-based translation layer to link the object identifier, or key, to one or more logical blocks. The storage device needed another translation to get from the logical blocks to the physical locations on the storage media. The KV protocol removed the need for the translation layers, improving performance and reducing storage needs for the translation tables, noted Bill Martin, who represents Samsung on various SSD/IO standards committees and serves on the NVM Express board.
Martin said KV is most applicable for database applications, such as RocksDB and Ceph, and enables shards of objects to be stored and retrieved from the KV device without translations in the host.
Another key addition with NVMe 2.0 is support for HDDs using the same command sets as NVMe-based SSDs in a consolidated I/O stack. The specification adds a bit to identify a device as rotational media, supports device statistics and log pages with HDD-specific info, and enables the host server and host applications to determine the optimal media for specific applications, according to NVM Express board members.
"Having one set of storage interfaces will simplify system integration and management," Coughlin said.
But Handy said he sees little reason for HDDs to go to PCIe, which he said would restrict the way the spinning disk can connect to the computer. The existing HDD interfaces already support the highest level of performance that a hard drive can provide, he said.
When, or if, NVMe-based HDDs materialize remains an open question. For instance, Toshiba has no official plans on its roadmap to offer NVMe HDDs, according to Raghu Gururangan, the company's vice president of engineering and strategic planning. However, he added, it will participate in the standards process and closely monitor market adoption of NVMe HDDs.
More NVMe 2.0 features
Additional NVMe 2.0 enhancements that the NVM Express organization spotlighted include:
- Endurance Group Management, enabling the creation of NVMe storage domains within an SSD or across SSDs based on endurance and quality of service characteristics. The capability could be helpful for hyperscale data centers to apply domains to specific applications.
- Simple Copy Command, for copying multiple logical blocks into a larger, contiguous block to potentially reduce write amplification and increase drive life.
- Command Group Control, to protect SSDs after provisioning and prevent unauthorized access to the drive.
- Key per I/O, to provide a mechanism for an application to set the security key on each I/O that it performs.
- Namespace Types, to allow the NVM, KV and ZNS command sets to run on different namespaces -- potentially, on the same drive -- to enable different workloads to run on the same NVMe SSD.
Carol Sliwa is a former TechTarget senior writer covering storage arrays and drives, flash and memory technologies, and enterprise architecture.
