My colleague Henry Baltazar and I have been watching the development of new systems and storage technology for years now, and each of us has been trumpeting in our own way the future potential of new non-volatile memory technology (NVM) to not only provide a major leap for current flash-based storage technology but to trigger a major transformation in how servers and storage are architected and deployed and eventually in how software looks at persistent versus nonpersistent storage.
All well and good, but up until very recently we were limited to vague prognostications about which flavor of NVM would finally belly up to the bar for mass production, and how the resultant systems could be architected. In the last 30 days, two major technology developments, Intel’s further disclosure of its future joint-venture NVM technology, now known as 3D XPoint™ Technology, and Diablo Technologies introduction of Memory1, have allowed us to sharpen the focus on the potential outcomes and routes to market for this next wave of infrastructure transformation.
Intel/Micron Technology 3D XPoint™ Technology
Still short of a product announcement, this was more of a (partial) clarification of Intel’s next generation NVM plans, which they have coyly hinted at with no detail for about a year. This announcement, while still not really specifying exact product shipping dates, pricing, delivered density or even exactly what kind of device it is (although the collective betting is that it is a variant of phase change memory, into which Intel has reportedly invested a lot of money), at least gave us some significant hints:
Higher density than current flash and performance, much higher endurance (number of write cycles), and much faster (by a claimed factor of 1000, but still well short of DRAM speeds)
CMOS compatible design and process with the ability to stack multiple layers of devices, hence the “3D” nomenclature, with the implication that they will be able to manufacture this stuff efficiently
Some nice pictures ships and diagrams of the devices showing its basic crossbar switching and access architecture (itself not a unique concept)
Strong hints that the devices could be sampling in 2016 and maybe begin to show up in products in 2017
The potential impact of this technology on current storage, even absent any fundamental changes in architecture is immense. Think about a 4U storage slice with a Petabyte of storage with latencies reduced by a very conservative decimal order of magnitude or more. Or for more modular and closely coupled storage, consider a standard 2U server with 10s of TB of embedded flash
There are multiple other horses in the next-gen NVM race, and it will be interesting to see if the Intel/Micron announcement triggers any public announcements or realignments of the players.
Diablo Technologies Memory1
Memory1 technology is a different spin on alternative memory architectures, in which current (and presumably future) generations of NAND flash are used in lieu of DRAM thanks to a combination of Diablo's clever hardware that allows flash to be inserted in the system DIMM sockets and software that turns part of the server memory into a transparent cache allowing umnmodified software to access the flash as standard DRAM. Diablo claims that Memory1 can enable capacities in the order of 1-2 TB in a standard 2S server with performance very close to all DRAM (a claim which we think is probably workload-dependent), and at a fraction of the price, DIMM count and power. For a large number of analytics and other workloads requiring large in-memory processing the advantages of flash in terms of increased density, lower power and cost can be transformative, allowing for a rapid expansion of very large memory configuration nodes for massive analytics problems that are currently not economically viable.
Does it really work? Since production shipments are a year or more away, it's a bit early to tell exactly how well, but we can make two reasonably compelling assertions:
- It does work – Diablo has already shown that it can produce the hardware side of the product since it has been shipping flash in DIMM products for a couple of years now, albeit with limited commercial success due to a number of factors, mostly related to partner dynamics as opposed to fundamental technology issues. Since it is currently in trials with multiple vendors, we can also assume that the integrated hardware/software stack works as well.
- It will probably be sensitive to workloads – While it is claimed to work with unmodified software, Diablo does allow that software that has been modifed to recognize that it is running out of what amounts to a really large cache will probably run better. This implies that workloads dominated by random reads and particularly writes will probably fare poorly and those where the caching layer can intelligently prefectch larger sequential blocks (which appears to include a lot of in-memory analytics workloads) will probably do well. As noted, Memory1 is currently in trials with multiple vendors, and we will track any further developments as we get data.
Interestingly, Memory1 is not persistent memory desipte its use of flash as its underlying technology. Diablo certainly has the ability to eventually make it persistent, but that precludes current software from using it without extensive modifications to be aware of which parts of its memory address range is persistent.
Long Term Transformation
The initial deployment of both technologies will be as augmented capabilities for conventional architectures. 3D XPoint™ technology will allow vastly expanded capacity and performance of current storage architectures, and Memory1 will allow software to run with much larger addressable memory spaces at significant cost and power savings. Longer term, as software is rewritten to understand that part of its memory space is now persistent, we will see fundamental changes in application architecture as technologies like Memory1 with persistent memory are integrated into servers and the boundaries between DRAM and persistent storage begin to blurr and eventually dissolve.
Looking into our crystal ball, the crossover/integration of Memory1 and other forms of NVM are extremely attractive, and it seems a pretty low-risk prediction that either Diablo Technologies will evolve to support 3D XPoint™ Technology or competing emerging NVM technologies as an alternative to flash, and/or that Intel will itself introduce the ability to address its NVM as memory. My prediction is that this transition will take place within the next five years, and that when the software community discovers the advantages of being able to assume that parts of your address space are persistent, there will be an avalanche of software written for integrated memory/persistence architectures.
Look for additional content from us on Forrester's website about this architectural evolution.