Slow ZFS resilver performance - would this be a reason to go back to HW RAID?

[alpha754293]

@Exard3k

distributed spare kicks in immediately and (as far as NVMe-backed VM disks go) it was rather quick.

I surmise that this is true IFF you have a spare drive as a part of your dRaid pool (which in my deployment, I won’t have any hot spares).

@Exard3k

If I had >10 disks and a capacity-optimized approach on storage, I’d probably go with dRAID.

I’ll have to do a LOT more testing with it before I can deploy it into PROD.

@Exard3k

Which is good, because your spares have data and are always working, improving performance not only during a resilver. Imagine dRaid as spare capacity rather than spare disks. No cold useless spares ever again.

Yup. I watched Mark Maybee’s presentation.

But again, this assumes that you have hot spares in the pool, which may or may not necessarily be the case.

@Exard3k

of course not. The vdev is degraded, not the rest of the pool. Only if the pool is just this one dRAID vdev…which may be true and even more preferrable than with raidZ. You basically have multiple RaidZ “vdevs”, called groups within a dRaid, giving all the advantages from multiple vdevs.

But again, per Mark Maybee’s presentation – he states that it will use all of the drives for the rebuild.

But in his example, he only has one 53-wide dRaid vdev. (See slide 8.)

So I just experimented with this and you are correct – it only resilvers JUST that vdev.

It is my understanding that for a raidz(#) resilver, does that also resilvers only using the disks from that vdev or does it resilvers using the entire pool?

It’s a shame that I didn’t experiment with this until now.

Tech YouTuber ElectronicsWizardry did a video about dRaid about 10 months ago (by now), but I didn’t really have time to experiment with it then.

And the system was already built.

It would be a pain for me to try and rebuild the entire system now, as I would have to offload all of the data from the bulk storage pool (onto LTO-8) so that I can destroy the pool, and then re-create it using dRaid.

That’s going to be a longer term project because it is going to take quite a substantially longer amount of time, to be able to offload that much data as I don’t think that I can take down my bulk storage pool, one vdev at a time.

(But please correct me if I’m wrong.)