Diablo Technologies Memory1 Briefing Note
Flash is typically thought of as a fast alternative to storage, the ultimate hard drive upgrade. But it is memory, so there is no reason IT can’t use it as an alternative to DRAM. The challenge is how to deliver a flash memory device whose latency is close enough to DRAM as to make it a viable alternative to DRAM. It has to get close enough in performance to make the cost savings and density gains worthwhile.
Overcoming Flash Latency Challenges
Flash introduces latency in two critical areas. First at the speed at which it writes data, thanks to its well documented program-erase cycle. It also suffers latency in its connection to the application. Typically flash connects to a server via SCSI, SATA or PCIe. Each of these connections incur communication latency and they incur overhead as data has to traverse the storage protocol stack.
Diablo’s Memory1 blends the benefits of flash and DRAM. It overcomes the first problem, write latency, by coupling a large amount of flash with a small amount of DRAM acting as buffer giving flash the time it needs to go to through its program-erase cycle. Diablo overcomes the latency induced from traditional latency by installing in the memory bus. The result is a flash memory DIMM that blends with traditional DRAM to massively increase memory per server while decreasing costs.
Data and Media Management
Other than assembling the right hardware components, Diablo is delivering two important features in Memory1. First, the solution provides a data management layer that pre-fetches data based on a learning engine. The design enables the right data to be in DRAM at the right time. The second feature is a media management feature that manages how data is written to flash. The goal is to maximize performance and maximize flash endurance.
The Road to Flash as Memory
Diablo has the hardware, the flash DIMM and software done. Diablo has also completed the second piece of the puzzle, a software driver that allows the operating system to see the Flash DIMM as memory instead as storage. The final piece of the puzzle requires server manufacturers to update their ROM Bios to support the product. Diablo is making progress here, but as one would expect it is slow going.
Today’s applications demand more memory, with an increasing number of them having a memory only design. The challenge that Cassandra, Couchbase, Splunk and Hadoop environments face is managing node proliferation due to by a shortage of memory not CPU performance. A Memory1 type of solution should enable each node to maximize its CPU potential and RAM potential at the same time, which should eliminate the number of nodes each data center requires.