DAWN: Durable Array of Wimpy Nodes (2011) [pdf] | Hacker News Reader

DAWN: Durable Array of Wimpy Nodes (2011) [pdf]

ssrc.ucsc.edu

19 pointssp3329 comments

There's already at least one startup building a flash controller intended to enable a DAWN-like architecture using TLC (triple-level-cell) flash, NxGnData http://www.nxgndata.com/ - more details on my blog at http://blog.dshr.org/2014/08/is-this-dawn-of-dawn.html

Full disclosure - I am a co-author of the DAWN paper. The key point of the DAWN paper is the same as the FAWN paper, that conventional architectures with large amounts of storage paired with powerful CPUs can be out-performed on many axes by large numbers of small nodes each with a small amount of both storage and computation. Flash is the available memory to do this with now, it will be replaced eventually by better solid-state memories (e.g. Phase-Change Memory) but the architecture will remain. See also Seamicro http://www.seamicro.com/ for a current example in the FAWN-like space.

The point of this paper rests on an unfair cost comparison of local flash vs. S3. We know that S3 is much more expensive than local disk. If they compared local flash against local disk (e.g. Pergamum), disk would still win.

They have a newer paper that sounds better: http://www.ssrc.ucsc.edu/pub/gupta25-mascots.html

Erm, yes and no. We focused more on the medium than the exact architecture. In point of fact, we directly say we draw heavily from the Pergamum design. The high-order bit, so to speak, is that disks suffer from a variety of electromechanical problems that flash and other SCM's don't suffer from, making it more appealing as a long term storage medium. Its (potentially) longer 'shelf-life' and lower power requirements mean it would likely have lower total cost of operation than a disk based approach.

They've buried the lede under cute terminology, like "wimpy nodes" and "storage class memory (SCM)".

The main thrust of the paper is that under certain assumptions, flash/solid-state/etc memory ("storage class memory") may already be competitive with spinning hard disks or tape for long-term archival storage of bulk data.

The "wimpy nodes" was a bit too cute for my tastes, but storage class memory is actually a fairly common term in the research community. The "wimpy nodes" was meant to be a coat-tail grab for the FAWN project, which we (disclaimer: I am one of the authors) heavily drew from in addition to the Pergamum project.

Thanks for the clarification. As an anecdotal data point, I didn't recognize the storage class memory term, even though I'm quite interested in these topics, sometimes reading related research papers.

The abstract would have been more helpful had it alluded to the definition, for example "…of storage class memories (SCMs), such as flash memory, has been…".

Similarly, the term isn't defined on first use, and even following the footnote which strongly suggests definition/discussion of the same topic elsewhere ([21]), that particular term doesn't appear in the referenced Rosenthal "Keeping Bits Safe" article.

(Now I know, but I'm still not a fan of the term. At these levels of abstract analysis, it's all 'memory', and it's all being considered for 'storage', so storage class memory lacks vividness/contrast.)

SCM tends to be an abbreviation for "whatever comes after flash, even though we don't know what that will be".

The industry term is NVM, which refers to all classes of non-volatile memory, typically high density ones.

How do you like Crystal Ridge?

NVM makes slightly more sense to me than SCM.

Crystal Ridge? (If you mean the nursing facility, still hoping for a more Rainbows End arrangement.)

Reply on news.ycombinator.com