Will persistent memory become an important storage medium in your data center? Here are the two storage class memory use cases that illustrate why it's most likely inevitable.
Storage class memory, also known as persistent memory, is one of the most exciting storage technologies to appear within the past few years. SCM is slightly slower than traditional dynamic RAM, but it is persistent. This means SCM-stored content is preserved throughout a power cycle, making it a potential breakthrough technology for a wide range of demanding storage applications.
"The benefits of the SCM are higher density, lower power and economical affordability compared to DRAMs," said Sylvain Dubois, vice president of business development and marketing for solid-state storage technology provider Crossbar Inc. "Today, most critical data is stored and accessed from DRAM, since it's the only viable solution to feed the computing cores in data centers for data processing, deep analytics and artificial intelligence," he noted.
DRAM, however, is power-hungry and expensive compared with the very dense and cheap NAND technology. "This is why a big portion of the data in data centers has to be stored in NAND-based SSDs or even cheaper disk drives," Dubois said. SCM products can use NAND as a backing store and DRAM as a cache for active data.
Promise of storage class memory use cases
Storage class memory use cases have a lot of promise in terms of reducing the system-level latency of storage arrays and the number of servers used in data centers, said Rajesh Vijayaraghavan, director of product management at data storage systems provider Hitachi Vantara.
The most exciting use case is persistent DRAM, where the contents behave as storage in that it's persistent, but at performance levels closer to DRAM.
Robert HormuthVice president and CTO of Dell EMC server and infrastructure systems
There are at least two key implementations that promise to unlock SCM's potential, he added. SCM can be used in the storage array back end where it could act as the caching tier for hot data with ultra-low latencies, he said. Or it can be deployed on the server-controller side to go lockstep with the Intel CPUs that support SCMs as a replacement for DRAM, he said.
Perhaps the more promising of those two storage class memory use cases is as a DRAM replacement for read-intensive applications, such as AI and deep data analytics. "From a single source, niche market opportunity in 2018, SCM adoption is widening to multiple sources and scaling up in performance and volume in the next couple of years," Dubois said. Standards bodies are already working on directing and unifying the various varieties of SCM and other emerging memory technologies to NVDIMM-P, GenZ, Cache Coherent Interconnect for Accelerators and Open Coherent Accelerator Processor Interface.
Size and savings
SCM brings the capacity and economics of storage to memory, said Kais Belgaied, CTO of Viking Enterprise Solutions, a server and storage systems technology provider. One of the main data center benefits is server consolidation. "Fewer nodes are required to run a memory-bound, distributed in-memory database or an in-memory compute engine, such as Apache Spark, which, in turn, reduces cost," he said.
Where storage class memory fits in
As a DRAM replacement, SCM can offer a lower dollar-per-GB rate than traditional DRAM, but only at the cost of potentially reduced performance and non-deterministic behavior, said Robert Hormuth, vice president and CTO of Dell EMC server and infrastructure systems. "The most exciting use case is persistent DRAM, where the contents behave as storage in that it's persistent, but at performance levels closer to DRAM," he noted. "In this use case, applications need to be optimized to understand and deal with the persistent characteristic."
SCM is likely to be adopted as a back-end media in almost all the key storage arrays by 2020, since the implementation is far simpler than using SCM as a DRAM replacement on the server side, Vijayaraghavan predicted. "The adoption of SCM on the server side may take until 2021 to mature and get adopted across all server-storage OEMs."
The advantages of using storage class memory for server consolidation are becoming too obvious to pass up, so SCM will likely be found on many upcoming server refresh cycle lists, Belgaied said. On the software side, multiple trends are bound to materialize, including more memory-hungry applications using SCMs as an alternative to horizontal scaling. Hypervisors, such as VMware's ESXi, will evolve to make SCM available to virtual machines, which will result in a jump in workload density for only a little extra latency cost, Belgaied said. "Orchestrators for cloud-native applications, such as Mesos or Kubernetes, will also evolve to recognize this different type of cheap and deep memory."
The road to adoption
The key to large-scale adoption of any technology lies in the availability of multiple sources and economies of scale, Vijayaraghavan said. "Hyperscale cloud vendors might be the first to adopt SCMs, both as a media and DRAM replacement, since they have tighter control on their system architectures and supply chain strategies."
