HPE aims superfast Memory-Driven Flash storage at real-time workloads

HPE is targeting a new class of storage it calls Memory-Driven Flash at enterprise data centers that are increasingly being called on to handle real-time analytics, high-speed transactions, big data and AI workloads that demand more storage performance than ever before.

To stay competitive, enterprises are moving to business models that require rapid analysis of an ever-widening stream of data. By 2020, as much as 70 percent of Fortune 200 companies will have at least one real-time big data and analytics application that will be viewed as mission-critical, according to IDC.

HPE’s Memory-Driven Flash is built around NVMe and storage-class memory (SCM) and designed to meet performance requirements of these new workloads.

NVMe is an extremely fast communications protocol and controller designed to move data to and from SSDs via the PCIe bus standard. 

Storage-class memory for high performance

SCM, meanwhile, is a hybrid memory/storage technology that, like flash, is nonvolatile (or persistent) but offers a level of performance somewhere between today’s NAND flash SSDs and DRAM. SCM offers up to 10 times lower latency than NAND flash technology, HPE says, and lower latency is crucial to meeting the performance requirement of real-time workloads.

“A number of storage vendors have started to ship enterprise-class storage platforms built around NVMe, but their continued use of NAND-flash-based storage devices leaves a significant amount of potential NVMe performance on the table,” according to IDC analyst Eric Burgener in a white paper on the new HPE technology. Memory-Driven Flash “uses Intel Optane technology as a caching tier in the shared storage array controllers, delivering acceleration for potentially all workloads running on that platform,” writes Burgener.

HPE has already engineered its high-end 3Par storage arrays to accept Optane-based expansion cards, made by HPE around Intel Optane media. The cards – block devices connected over an NVMe bus – will be available for 3Par system next month, HPE announced at its Discover conference in Madrid Tuesday. HPE’s mid-tier Nimble arrays will work with the new class of storage early next year, the company said.

HPE leverages Intel Optane for storage

Up to now, Intel Optane cards were being used in PCIe slots on X86 servers, Burgener said in an interview. “What HPE did, simply put, was put PCIe slots onto their storage controllers so you could just take that card and stick it into a storage controller,” Burgener said. “It’s acting as cache capacity on the storage array itself.”

HPE tests show that 3Par arrays with the cards are able to deliver sub-200-microsecond latencies on average, as measured from the application perspective using a Fibre Channel transport to the array, and indicate that the offering can deliver extremely low latencies at scale, Burgener notes.

“This is a true enterprise class solution that leverages Optane, so you get efficient sharing capacity and utilization, massive capacity, and enterprise class data services like snapshot replication, encryption, deduplication and all that stuff you didn’t have access to when you were putting the cards in servers,” he said.

For HPE, the announcement of its Memory-Driven Flash is the first step in a fundamental shift.

“We see Storage-Class Memory as revolutionizing the way software architectures are built,” said Ivan Iannaccone, general manager of HPE’s 3Par line. “It’s not NAND flash – in time we are really going to change the architecture of the 3Par OS and the Nimble OS to leverage huge pools of shared memory.”

InfoSight moves beyond predictive analysis

At Discover Madrid, HPE also announced improvements to the AI-based recommendation engine in its InfoSight predictive analytics platform.  

InfoSight has been tweaked for the Nimble product line to move beyond predictive analytics and generic system recommendations to offer specific suggestions on how to improve system performance, Iannaccone said.

InfoSight collects infrastructure information from sensors and runs analytics against the massive amount of usage data it has accumulated over the years to detect patterns in order to predict, for example, when a user might run out of storage.

“We collect data from the hypervisor, we model the data and we do what’s called cross-stack analysis – we look at what is happening on the array, what is happening on the network what is happening on the hypervisor layer and we can build a number of reports,” Iannaccone said. “We’ve been able to do this for a while, but now we can tell you what to do about it – add more memoryor scale out your computing infrastructure or maybe you need to scale out your storage infrastructure depending on where the bottleneck is.” 

HPE this week is also announcing that HPE Cloud Volumes, an enterprise storage service designed for hybrid and multicloud mobility, will expand into the U.K. and Ireland in 2019 to service UK and European customers who require local cloud-data access.