It sounds unlikely: technology that makes MLC flash as reliable and long-lived as more expensive SLC? Right!
Or even better: make 3 bit per cell flash equivalent to 2 bit MLC? What are these guys smoking?
Whatever it is, I’d like to try some myself. Because it looks like they have the goods.
Flash vendors are close-mouthed about flash problems, not wanting to kill the goose that’s laying billion$ in revenue. But the problem most people worry about – life span – isn’t much of a problem.
So what is? Errors. The good news: most NAND flash errors have predictable causes. And what can be predicted can be corrected.
That’s Avraham Meir’s story and he’s the CTO of Anobit. We had a chat at SNW.
According to their web site:
Mr. Avraham Meir is an internationally recognized authority in NAND Flash technology and products. Prior to joining Anobit, Mr. Meir was VP Corporate Engineering at SanDisk (NASDAQ: SNDK), and the CTO at M-Systems (NASDAQ: FLSH, acquired by SanDisk).
Flash errors are commonly due to cross coupling between adjacent cells, read disturbs and program disturbs, and reading or programing the wrong cell. Retention impairments and endurance impairments are more common as geometries shrink.
Data retention phenomena and endurance effects are not random: normally high levels go to lower levels. In adjacent cells, one high and the other low, voltage will leak across to reduce the higher voltage.
So the first stage is reducing errors is knowing how flash impairments behave to predict an error. Then working with flash vendors they can repair the impairment in the wild.
By reducing the errors first and then applying signal processing and error correction Anobit makes less-reliabile flash look like more reliable flash.
Extending endurance has another effect: with flash as the number of writes increases so does the error rate. And as the endurance increases the data retention time drops to as little as a few months. Write once and your data may last for years. But don’t trust that old thumb drive.
The StorageMojo take
Anobit has an impressive IP portfolio: go to the USPTO patent application search engine, type in anobit and you’ll see what I mean. And they have about a dozen granted already.
What is evident is that flash as a medium is still in its infancy. If it turns out, as some predict, that flash will hit a geometry wall in a generation or two, Anobit’s technology looks to be transferrable to whatever next-gen non-volatile device wins. In storage, reliability and endurance are problems that will never go away.
Courteous comments welcome, of course.
Ahh, good, i can be only a bit OT with this flash related link, which harks back to a discussion here about in-mem db’s.
Back on topic, for a second, and i’m writing you from the fairly, somewhat, almost litigously conservative British Isles*, i’ve been very privately debating recently whether even transient data needs to be streamed and timestamped properly to a WORM.
This is not because of some vendor pitch. This is an internal debate, about how things like indicative offers on a trade might affect contrcatual intent, and extend chain of evidence inside a pricing algo. Not your usual discovery / audit stuff. The current counter is, how do you do that, if like LMAX you’re running the process inside the CPU? Synchronous processing with a duplexed bus, mainframe style?
(Dear me, i just found the Doomsday Scenario for x86! 🙂
So it seems, having massive IOPS is a total pain, because how on earth can you offload that neatly, even at a fraction of the data being relevant? Well, not at that level, anyhow.
Point of my illustration, is at what point do we need relplicated streams wherever critical processes touch SSD, and how is that to be engineered to cope with the inherent performance mismatch?
That paper above is starting to look much more interesting now. Just like clustering, do the darned thing as high up as you can.
. . .
This, about COW vs B-tree is also very pertinent:
I do not want to buy someone’s storage kit because of their algo. I want to talk native FS, and be able to pick and choose the underlying layout. I think what we have now is many potential licensors looking for licensees, well in the kind of way a pre-teen might get a crush, but not actually go talk, and the potential licensees, who have the scale to keep serious companies cool under the collar, are circling wagons self defeatingly. So, not for the first time, instead of cross licensing, we get holdouts who hope to sell enough vertical to sell at sweet multiples. That just doesn’t work in tech- IPO poor times.
*we “lead” you by miles on some things, like libel and gagging orders, tho’!
Well, that sounds like Sandisk’s “secret sauce” that they add to their TLC chips that make them emulate MLC and, Eli Harari claimed last year, even SLC. The fact that Meir used to work for Sandisk makes it seem even more likely that this is the case. I am a little surprised, and wonder if you can find out more about it. (Apologies for how late this response is–I only just found your blog.)