Has NetApp solved the TLC for biz puzzle? We look at its MYSTERY SSD Mars sauce

NetApp is pinning FlashRay's future on TLC flash, which everyone else says is not fit for enterprise use. We talked to NetApp to find out how and why it's apparently got hold of secret TLC sauce that no one else has. Let's reiterate that TLC NAND is a crap enterprise storage technology. It stores 3 bits per cell instead of MLC …

[Anonymous Coward]

Log structured writes will eliminate write amplification

But you can't go below 1.0, at which point you're still bound by the same restrictions as everyone else. Unless they think they have a magic way of erasing each block 5000 times that everyone else can only erase 500 times.

They're probably banking on most customers not actually writing at anything close to the rates they probably believe (once DBs are out of the picture) especially when write cache hits reduce the amount of write data actually saved to the underlying storage.

I'll bet their strategy is as follows: Their SSDs will still have blocks that can only be erased/rewritten 500 times, but that's fine because they'll know when blocks start failing (cutting into the overprovisioning) and can easily do an advanced replacement of the SSDs before they actually fail. For that small fraction of customers who really do have a write rate far too high for TLC, they can silently swap in some higher priced MLC drives and thereby avoid having to swap their drives 3x a year.

When I got SSDs at home I was bit worried about the lifetime, but seeing as SMART thinks I've still got over 90% of the life left after several years it looks like I greatly overestimated my own "full disk write" rate :)

[Big Ed]

Re: Log structured writes will eliminate write amplification

Former NetApp'er here...

Maybe these NetApp'ers ain't so crazy.

One thing that NetApp did very well in the past was predictive analytics.

A lot of perople think of write-cycle limits in terms of full-block writes. But what if you know what bits in a record actually change and only write the updated bits. Micron/Intel boffins claim their new spinless silirust can address individual bits, why not the other boys? And if NetApp has the partnership they claim it's in the realm of possibility that they could get a custom/nextgen controller that does bit addressing.

Think about it, past the initial write of a record, generally speaking only a few bytes change in a DB record update, and at the bit level there is a 50/50 probability that a bit actually changes in a given write, and for text fields; the printable ASCII character set reduces the odds across a whole byte.

NetApp may be on to something!!!