The Memory1 hardware was announced last year, but now, as it hits production release, it is accompanied by Diablo Memory Expansion (DMX) software.

Ottawa-based Diablo Technologies has announced the production release and volume availability of its Memory1 128GB DDR4 system memory module. It is also announcing the accompanying Diablo Memory Expansion (DMX) software, which unlike the hardware had not been publicly revealed previously.

“A year ago, when Memory1 was first announced, we were talking it about relative to the DIMM we created, said Kevin Wagner, Diablo’s Vice President of Marketing. “It’s now the DIMM plus the software.”

Diablo’s Memory1 technology is designed to provide very high capacity system memory, with its all-flash DDR4 system memory modules being able to provide the same performance as high end DRAM [dynamic random-access memory], but at a much lower cost. Leveraging its unique memory-channel architecture, which lets flash interface directly onto the DDR bus, Diablo uses NAND flash to get standard technologies to behave like DRAM.

“It is fully DD4 compliant, but has a different architecture,” Wagner said. “No changes are required to the motherboard, operating system or applications. It goes in just like a standard DD4 DIMM.”

The result is large increases in server and application capability that make Memory1 highly suitable for memory-intensive workloads and customers. Their tagline for the product is ‘Big Data applications need Big Memory.

“Hyperscale, Webscale and large enterprise customers are our target markets,” Wagner said. “Cloud, database and Big Data processing are our main areas of focus right now. We expect to see use in HPC [High Performance Computing] later.”

Both 1 TB and 2 TB application acceleration appliances are available from OEM partners.

“We are seeing more demand for the 2 TB systems than the 1 TB,” Wagner said. “We expect to have a 4 TB next year.”

While the release of the Memory1 module to general production is important, even more interesting news is the availability of the accompanying Diablo Memory Expansion (DMX) software, because it is being announced for the first time. It is a combination of firmware and software, which leverages the CPU hardware to intelligently manage application memory access, performance and endurance.

“It loads just like a driver, with no application changes,” Wagner said. “As memory operations come in, we can see them and decide what to do with them.”

The DMX Software has two layers, for data management and media management.

“The data management layer uses DRAM in the system as a cache, with high priority and hot data going to the DRAM cache,” Wagner said. “This initiates data tiering – to DRAM, flash, or default to storage. The user can decide what tier to place the data in.”

Memory Fencing and Barriers, which are used by compilers, are enhanced by the layer’s sequentialization of data.

“This supports the fences which have been created,” Wagner said.

A Quality of Service feature – Priority Associated Data Placement – allows data in DRAM to be identified with response time requirements.

“A customer asked for this, to tag parts of data as high priority or low priority,” Wagner said. “This may not be a ‘hot’ feature but it is important.”

Other data management features include a Learning Engine for application profiling and analytics.

“We decided to build a learning engine to profile the application and determine access patterns before they happen,” Wagner indicated. “It monitors application data access behaviors, and predicts the next or additional pages required. Smart Data pre-fetch intelligently pre-fetches pages to DRAM based on profiling, history, and data access patterns. Clustered pages capability pre-fetches grouped data from surrounding pages, knowing there’s a high likelihood they will also be accessed.”

Movement between DRAM and Memory1 ensures data will be local to the associated node, for improved performance.

“We can ensure data locality because we own the memory manager and the interface to flash,” Wagner stated. “We can keep data local to the thread, which gives us lower data and faster access times.”

Finally, Amortized Page Faults groups page requests together, fully leveraging page faults.

“This minimizes the number of page faults we have to deal with,” Wagner said.

On the Media Management level, the most notable features relate to flash management and intelligent traffic management. Low-level media management is handled in the firmware. Other flash management is provided by SoftFTL Adaptation for 4K pages, a tuneable cache ratio, and device striping.

“For Intelligent Traffic Management, Dirty Page Writes, which avoids premature writes to flash for frequently written pages, increases endurance considerably,” Wagner said. We also have traffic sequentialization, where pages evicted are written sequentially to flash. Because we own the interface to the flash, we can decide how it’s written for the flash. So we have created this queue so we can write sequentially to the flash and get better performance.”

In addition to the Memory1 announcements, Wagner emphasized other areas of company momentum.

“We are now at around 75 people, growing especially in software and application engineering,” he said. “Our recent Series C round came in at around 37 million, from both current and new investors.

“We also just announced qualification with Inspur Systems, which has a big server business in China, and has recently been investing heavily in the U.S. Supermicro and another very large server vendor are in the process of qualification as well.”