Briefing Note: Diablo Delivers Flash As Memory
Since its introduction to the enterprise, flash has been viewed as an expensive but fast alternative to the hard disk drive, but not as inexpensive as RAM. Flash implementation options like PCIe SSDs and Flash DIMM that look like memory are still like storage. Diablo is set to change that view with the announcement of its Memory1 solution; Flash DIMM technology that can replace DRAM.
The Need For Cheap Memory
The need for a plentiful, inexpensive alternative to traditional DRAM is rising thanks to in-memory databases and highly dense virtual environments. But memory is expensive to purchase, especially high-capacity DIMMs. It is also limited; the high capacity DIMM that can be purchased today is 64GB, and those are in limited supply. The result is that the IT planner is limited as to how much memory they can install into a server that is supporting an in-memory database or virtual environment.
The Diablo Memory1 will have 256GB of flash storage on a module that looks identical to a DRAM DIMM and is designed to go into a memory slot. Diablo’s software will take over memory management and manage both DRAM. A small amount of DRAM (10:1 ratio) will be required to accompany the Memory1 module. While flash performance will not be as fast as DRAM, it will benefit from being on the dedicated memory bus and not having to emulate a storage I/O stack. Latency should be very low and performance very acceptable for most use cases.
Support of the Diablo Memory1 will require that server vendors make changes to their BIOS to support it, something that Diablo claims is relatively straightforward. But more importantly, IT professionals will need to make no changes to their operating systems or applications. The Memory1 module should look just like RAM to those environments.
There are many environments that can leverage a Memory1 type of technology. At the top of the list are hyper-scale cloud data centers that are under constant pressure to personalize their user’s experience. That personalization has to be done seamlessly and in real-time. In-memory seems to be the method of choice, but the limitation on how much memory this method can access internally in the server causes a problem. The moment these applications have to get memory externally, the seamless, real-time experience is threatened. Memory1 significantly expands the memory capacity of the server and eliminates external fetching of data.
While the initial demand for this technology will clearly come from hyper-scale cloud data centers, we expect this type of technology to be broadly applicable to data centers of all sizes. As pricing decreases and more server vendors adopt the technology, a solution like Memory1 could be used to build a very dense virtual server and desktop infrastructure and bring new levels of scale to traditional database applications.