At the 14th USENIX Conference on File and Storage Technologies (FAST ’16) held in Santa Clara, CA from February 22, 2016 to February 25, 2016, researchers Bianca Schroeder, University of Toronto, Raghav Lagisetty and Arif Merchant, Google Inc. concluded that “SLC drives, which are targeted at the enterprise market and considered to be higher end, are not more reliable than the lower end MLC drives.”
The full report can be found here.
This simple, yet unexpected statement, vindicates what many flash vendors have argued for a long time that the use of MLC is viable in today’s enterprise environments.
That said, the report also concludes that “when it comes to SLC drives and their RBER, … they are orders of magnitude lower than for MLC and eMLC drives.” But “SLC drives do not perform better for those measures of reliability that matter most in practice: SLC drives don’t have lower repair or replacement rates, and don’t typically have lower rates of non-transparent errors.”
What does this mean for the industry?
In some ways, the difference between SLC and MLC drives are somewhat (albeit, a bit of a stretch) analogous to the high density (HD) and double density (DD) floppy disks of yesteryear. Research shows that the production of both disks were identical, and that the difference between them was merely whether a batch passed or failed certain final testing.
The impact of these findings, which Neuralytix demands further confirmation by other researchers, could mean that the price of high end flash arrays can drop significantly, since the primary cost of a flash array is the media (in this case the NAND) itself, can be halved. The cost of developing more intelligent media management within the firmware of software of the arrays is a worthy exercise if these arrays can become more competitive against other “lower end” arrays.
It also means that the viability of vendors such as Diablo Technologies, with their Memory1 product that enables NAND to be used in the DIMM slot normally occupied by DRAM, can have a greater acceptance in the marketplace at a reasonable and commercially viable price point.
These findings, again Neuralytix stresses that more research and confirmation of these findings are necessary, puts additional pressure on rotating magnetic hard disk drives (HDD) from a price comparison perspective. One thing the report does not discuss is the performance. While the findings bode well for MLC with respects to reliability, performance must also be taken into account before any wholesale changes are made to products and deployments.
Ultimately, Neuralytix believes that whatever the price of NAND, it will never gain parity with high capacity HDDs. We believe that while some hyperscale datacenters may become an all-flash datacenter, most customers will still follow the rule that roughly 5-10% of data should reside on flash memory. Big Data and analytics is may double these percentages, but not higher 25% for most customers.
We do not expect customers or vendors to rush to replace SLC with MLC (and nor should they). Performance is still a key criterion, and customers and vendors need to always look at the four dimensions of reliability, performance, capacity and cost. (Neuralytix stresses that NO MODERN INFRASTRUCTURE TECHNOLOGY can be measured on a two-dimensional basis.