The SNIA Persistent Memory Summit held in late January 2019 provided a good view into the current state of industry. Some key technologies and standards related to persistent memory are moving forward more slowly than expected. Others are finally transitioning from promise to products. This article summarizes a few key takeaways from the event as they relate to enterprise storage systems.
Great Performance Gains Possible Without Modifying Software
One point the presenters at this SNIA-sponsored event took pains to make clear is that great performance gains from storage class memory are possible without making any changes to the software that uses the storage. For example, a machine learning test using Optane to extend server memory capacity allowed a standard host to complete 3x more analytics models.
These results are being obtained due to the efforts of SNIA and its member organizations. They developed the SNIA NVM Programming Model and a set of persistent memory libraries. Both Microsoft Windows and multiple Linux variants take advantage of these libraries to enable any application running on those operating systems to benefit from persistent memory.
Optane is a Gap Filler in the Storage Hierarchy, Not a DRAM Replacement
One fact made clear across multiple presentations is that Optane (Intel’s brand name for 3D XPoint persistent memory) fills an important gap in the storage hierarchy, but falls short as a non-volatile replacement for DRAM. Every storage medium has strengths and weaknesses. Optane has excellent read latency and bandwidth, so deploying it as a persistent read-cache as HPE is doing may be its primary use case in enterprise storage systems.
MRAM is Shipping Now and Being Embedded Into Many Products
The main surprise for me from the event was the extent to which MRAM has become a real product. In addition to Everspin and Avalanche, both Intel and Samsung have announced that they are ready to ship STT-MRAM (spin-transfer torque magnetic RAM) in commercial production volumes.
MRAM offers read/write speeds similar to DRAM, and enough endurance to be used as a DRAM replacement in many scenarios. The initial focus of MRAM shipments is embedded devices, where the necessary surrounding standards are already in place. MRAM’s capacity, endurance and low power draw make it a great fit with the requirements of next-generation embedded edge devices.
Kevin Conley, CEO of Everspin Technologies, gave an especially helpful presentation describing the characteristics of MRAM and how it fits into the memory technology landscape. He stated that MRAM is currently being used in enterprise SSDs, RAID controllers and storage accelerator cards. His 10-minute presentation begins approximately 13 minutes into this video recording.
Persistent Memory Moving Onto the NIC
One new use case for persistent memory is to place it on network interface cards. The idea is to persist writes on the NIC before the data leaves the host server, eliminating the network and back-end storage system from the write-latency equation. It will be interesting to see how providers will integrate this capability into their storage solutions.
MRAM Memory Sticks Waiting on DDR5 and NVDIMM-P Standards
One factor holding back MRAM and other storage-class memories from being used in the familiar DIMM format is the lack of critical standards. The NVDIMM-P is the standard for placing non-volatile memory on DIMMs. The DDR5 standard will permit large capacity DIMMs. Both standards were originally expected to be completed in 2018, but that did not happen. No firm date for their completion was provided at the Summit.
Not all are waiting for the standards to be finalized. Intel is shipping its Optane DC Persistent Memory in DDR4-compatible DIMM format without waiting for the NVDIMM-P standard. The modules are available in capacities of 128, 256 and 512GB–a foretaste of what NVDIMM-P will do for memory capacities. While it is good to see some pre-standard NVDIMM products being introduced, the NVDIMM-P and DDR5 standards will be key to the broad adoption of persistent memory, just as the CCITT Group 3 and IEEE 802.3 standards were to fax and networking.
NVDIMM-N Remains the Predominant Non-Volatile Memory Technology for 2019 and 2020
The predominant technology for providing non-volatile memory on the memory bus is based on NVDIMM-N standard. These NVDIMMs pair DRAM with flash memory and a battery or capacitor. The DRAM handles I/O until a shutdown or power loss triggers the contents of DRAM to be copied to the flash memory.
NVDIMM-N modules provide the performance of DRAM and the persistence of flash memory. This makes them excellent for use as a write-cache, as iXsystems and Western Digital do in their respective TrueNAS and IntelliFlash enterprise storage arrays.
NVMe-oF Delivers in 2019 and 2020
If the DDR5 and NVDIMM-P standards are published by the end of 2019, we may see MRAM and other storage class memory technologies in enterprise storage systems by 2021. In the meantime, enterprise storage providers will focus on integrating NVMe and NVMe-oF into their products to provide advances in storage performance. Multiple vendors are already shipping NVMe-oF compliant products. These include E8 Storage, Pavilion Data Systems, Kaminario, and Pure Storage.
Learn More About Persistent Memory
DCIG focuses most of its efforts on enterprise technology that is currently available in the marketplace. Nevertheless, we believe that persistent memory will have significant implications for servers, storage and data center designs within the technology planning horizons of most enterprises. As such, it is important for anyone involved in enterprise information technology to understand those implications.
You can learn more about persistent memory from the people and organizations that are driving the industry forward. SNIA is making all the presentations from the Persistent Memory Summit available for viewing at https://www.snia.org/pm-summit.
DCIG will continue to cover developments in persistent memory, especially as it makes its way into enterprise technology products. If you haven’t already done so, please signup for the weekly DCIG Newsletter so that we can keep you informed of these developments.
This is the second in a series of articles about Persistent Memory and its use in enterprise storage. The first article in the series is Caching vs Tiering with Storage Class Memory and NVMe – A Tale of Two Systems. The third article is Ways Persistent Memory is Showing Up in Enterprise Storage in 2019.
This article was updated on 4/1/2019 to add more detail about MRAM and NVDIMM-P, and on 4/5/2019 to add links to the other articles in the series.