Over on ZDnet this morning I wrote about a FAST ’14 paper modeling how a PCM SSD could be used in a hybrid – PCM SSD, flash SSD, HDD – storage system.
For an academic research paper, this one is refreshingly focussed on business case enabled by technology.
Any new NVRAM is going to cost way more than NAND flash on a per gigabyte basis, due to economies of scale, learning curves and short term market demand – NAND has a broad support ecosystem that the new tech won’t have. This is why the most likely successor will be built on current NAND fab lines.
Therefore the question: are there any leverage points in current hybrid arrays – such as Avere and Nimble – that would get an outsize benefit from a new NVRAM with new and improved, but costlier, characteristics?
From the current grab bag of new NVRAM – largely types of resistance RAM (ReRAM) – these characteristics could include
- Speed. Much faster writes.
- Endurance. What if we could write to NVRAM 10 million times instead of 10 thousand?
- Stability. The more writes to flash the shorter the time the data will be held.
- Power. What if we didn’t have to pump 20 volts to write NVRAM?
- Density. Soon NAND will be bumping up against the physics and further shrinkage will be difficult or impossible.
This isn’t about storing some parameters in NVRAM, but about using dozens or hundreds of gigabytes to produce a faster, better, cheaper storage system.
The StorageMojo take
Any ideas, architects?
In the not-yet-publicly-available FAST ’14 paper Evaluating Phase Change Memory for Enterprise Storage Systems: A Study of Caching and Tiering Approaches, testing a 64GiB PCM SSD, authors Hyojun Kim, Sangeetha Seshadri, Clement L. Dickey and Lawrence Chiu of IBM Almaden Research conclude
Based on the results above, we observe that PCM can increase IOPS/$ value by 12% (bank) to 66% (telecommunication company) even assuming that PCM is 4× more expensive than flash.
Such numbers are not likely to drive commercial adoption, especially given the rate at which flash SSD prices are dropping. But the concept of leveraging the advantages of ReRAM in a hybrid system is intriguing.
Assuming you had complete system control – write a new file system; implement unique tiering algorithmg; optimize dedup or snapshots – where are the leverage points that would justify a costlier NVRAM?
Courteous comments welcome, of course. This post is all about the comments, so wax eloquent!
Robin, there is another characteristic of at least some of the new RAM types, with potentially far reaching architectural consequences: addressability. We may find ourselves with a modern (fast, frugal & cheap) version of magnetic core memory. It may be possible to change the contents of single bytes or words, doing away with erase blocks, write amplification, garbage collection, wear leveling, etc. If the new technology could be used for static and dynamic RAM, cache memory, stable storage and backup, from mobile devices to supercomputers, the market dynamics would change completely.
I’ve seen some figures that suggest that if PCM gets economies of scale driven by adoption of mobile devices (where it could replace the DRAM component entirely) that the price pr GB would drop well below SLC and be extremely competitive on a $/IOP basis. We’re already seeing SLC+TLC hybrid drives, so if the scenario above pans out, I imagine that we’ll be seeing PCM+TLC based devices, which would drive further economies of scale for PCM.
Now imagine that you disaggregate the SLC/PCM and the the TLC Components and treat the TLC portion as a consumable (like printer ink). If you made it cheap enough it could replace tape (I’ve seen projections for that, I thought they were silly at first, now I’m not so sure). Then assume that a large portion of the PCM (or more likely Re-RAM) lives up in the server close to the CPU, and that you have software which extends into the host that looks after cache-coherency in the face of snapshots and restores on the slower/cheaper media further down the hierarchy
This goes well beyond the existing hybrid array designs, and speaks to a different way of thinking about storage as it all just becomes a multi-layer memory hierarchy that stretches from the CPU L1 cache all the way to a networked memory device on the other side of the planet.