5 ms·

What would be interesting for SSD is percentage of advertized TBW when (or just before) the SSD failed, ie 100% if the SSD fails at exactly its advertized TBW,

by guerby 5y ago

What would be interesting for SSD is percentage of advertized TBW when (or just before) the SSD failed, ie 100% if the SSD fails at exactly its advertized TBW, 50% if it fails at half the TBW, and 200% if it lasts two times the advertized TBW. In case someone from backblaze read this :)

londons_explore 5y ago
A well written SSD firmware simply slows down with age. It will never get to failure because the slowdown gets so extreme it becomes unusable. The drive also gets slightly smaller (by passing write failures to the OS to mark sectors bad). Thats because a "worn out" flash sector is never fully worn out - it can still store some data, just less than the error correction can correct. It is possible to combine two sectors to have extra error correction data to still recover a sector. Now you have less than half the performance. Worn out flash also doesn't hold data long - perhaps only a few hours before too many bits have flipped and it is unreadable. To fix that, you need to rewrite the data, which slows everything down more. And now that you have a bunch of unreliable sectors, you also need "super-sectors" which can do sector-based hierarchical erasure coding to recover data from sectors where even the methods above have caused data to be lost. This slows down writes even more. In the worst case, reading a single sector requires reading every sector on the drive to reconstruct. Clearly thats going to be slow enough the drive will have stopped being used long before that. Sadly some drive firmware doesn't implement some or all of the above, so they appear to have "failed" and become unreadable, which IMO is inexcusable when it's very easy to design so that worn out drives become slow instead.
- bl4ckneon 5y ago
  Wow this was a really in depth comment!
- capableweb 5y ago
  Super interesting. Do you have any examples of drives that have the proper firmware in your experience? Sounds much better to own than something that suddenly fails.
  - londons_explore 5y ago
    Sadly not. Black-box testing an SSD for these kind of features takes months and hundreds of drives, and manufacturers will never talk about the inner workings of their firmware. Many big SSD users develop their own SSD hardware and firmware partly for this reason. SSD firmware is also a spectrum between "correct" and "performance", and I know of no SSD's that for example maintain all of the acknowledged data on a power failure. Sure many SSD's may typically do that, but that isn't a guarantee when the power fails in worst-case conditions. Old Apple SSD's for example have a special extra wire on the connector specifically for "impending power failure" to help them do that. PC SSD's don't even have a standard message to mean "power failure expected in 250 milliseconds".
- labawi 5y ago
  While I agree SSD's should either slow down or fail read-only, the rest seems like wishful thinking and/or extreme exaggeration. Do you have any examples or references that drives that implement your suggested algorithms? I wouldn't expect drives to have sector-splicing and super-sectors (though multi-level cells regularly store less bits/cell), infinitely degrade their size via write-failures or rewrite data every hour. Especially frequent rewriting would self-destruct the drive if it weren't already preceded by catastrophic data loss.

by guerby 5y ago

londons_explore 5y ago

A well written SSD firmware simply slows down with age. It will never get to failure because the slowdown gets so extreme it becomes unusable. The drive also gets slightly smaller (by passing write failures to the OS to mark sectors bad). Thats because a "worn out" flash sector is never fully worn out - it can still store some data, just less than the error correction can correct. It is possible to combine two sectors to have extra error correction data to still recover a sector. Now you have less than half the performance. Worn out flash also doesn't hold data long - perhaps only a few hours before too many bits have flipped and it is unreadable. To fix that, you need to rewrite the data, which slows everything down more. And now that you have a bunch of unreliable sectors, you also need "super-sectors" which can do sector-based hierarchical erasure coding to recover data from sectors where even the methods above have caused data to be lost. This slows down writes even more. In the worst case, reading a single sector requires reading every sector on the drive to reconstruct. Clearly thats going to be slow enough the drive will have stopped being used long before that. Sadly some drive firmware doesn't implement some or all of the above, so they appear to have "failed" and become unreadable, which IMO is inexcusable when it's very easy to design so that worn out drives become slow instead.

bl4ckneon 5y ago
Wow this was a really in depth comment!
capableweb 5y ago
Super interesting. Do you have any examples of drives that have the proper firmware in your experience? Sounds much better to own than something that suddenly fails.
- londons_explore 5y ago
  Sadly not. Black-box testing an SSD for these kind of features takes months and hundreds of drives, and manufacturers will never talk about the inner workings of their firmware. Many big SSD users develop their own SSD hardware and firmware partly for this reason. SSD firmware is also a spectrum between "correct" and "performance", and I know of no SSD's that for example maintain all of the acknowledged data on a power failure. Sure many SSD's may typically do that, but that isn't a guarantee when the power fails in worst-case conditions. Old Apple SSD's for example have a special extra wire on the connector specifically for "impending power failure" to help them do that. PC SSD's don't even have a standard message to mean "power failure expected in 250 milliseconds".
labawi 5y ago
While I agree SSD's should either slow down or fail read-only, the rest seems like wishful thinking and/or extreme exaggeration. Do you have any examples or references that drives that implement your suggested algorithms? I wouldn't expect drives to have sector-splicing and super-sectors (though multi-level cells regularly store less bits/cell), infinitely degrade their size via write-failures or rewrite data every hour. Especially frequent rewriting would self-destruct the drive if it weren't already preceded by catastrophic data loss.