Hacker News Reader

Comment by ac29 | Hacker News Reader

Interesting! I never would have thought of a spinning drive as needing a TRIM command, but it makes sense for drives where data is literally overlapping.

But, why would failure to expose TRIM cause errors in in rebuilding a RAID, in which case old drives should be more or less read-only, and replacement drives write only? My understanding is that TRIM should affect re-writing sectors only.

View on HN

TRIM isn't the cause. The problem with SMR is that disk failure in a RAID array causes read/write patterns during a rebuild that SMR simply was not designed for. I'm sure SMR friendly RAID software will be available in the future for but this requires the software to know that the drive uses SMR. SMR also tends to have really awful performance for regular desktop usage.

Yah, by blocking foreground read/write activity until an entire SMR region has been filled. That will work well...

Sorry what I meant was the difference between host-managed and DMSMR. Having TRIM doesn't make any difference in the rebuild, it just means you can detect it's SMR and you can help the drive in normal usage. Either one will have trouble in a RAID rebuild. Being able to detect it's an SMR drive could probably let some RAID software handle it in a better way, but I don't think any existing software does (since nobody is crazy enough to use SMR in a RAID)

In theory, SMR drives shouldn't have issues with reading and they should also write fine if you write entire stripes. With HMSMR it should be possible. With DMSMR - I don't know. Do DMSMR drives partially skip staging area on large writes and fully if you align them?

I don't see why RAID rebuilds should take notably longer on SMR drives, if only they were properly managed - by both the RAID driver and drive FW. Though it's hard to properly utilize a DMSMR black box.

RAID's are generally rebuilt online. Which means that the normal read and write activity is still occurring. Rewriting a block of data in the middle of a SMR region requires reading the entire SMR region, updating the changed data and then writing it back. So, normally a raid controller will just stop updating a strip (frequently much much smaller than a SMR region) and perform whatever forground activity is needed. The smarter ones rebuild the the strip of the new data along the way. This is going to really mess up rebuilding on SMR because suddenly what you assume is probably a linear operation isn't from the perspective of the drive.

TRIM is not involved when RAID is rebuilding. You are correct in that.

Reply on news.ycombinator.com