SMR hard drives... MDRAID or DMRAID (instead of zfs)?

[eccerr0r]

Well apparently there's a lot of hubbub about shingled magnetic recording hard drives and how it kills ZFS.

However I wonder, how about MDRAID or DMRAID?

I know it might be sacrilege to buy a SMR drive as it's known that it kills ZFS RAIDZ resilver, but I'm seeing inconsistent results for MD and DM RAID. My current concern is that one of my disks in my MD RAID is failing and should I really be worried about getting a SMR disk? I suspect a device managed SMR would be better than ending up with a host managed SMR as MD doesn't know anything about SMR, but will I see the failing resync behavior in MD or DM RAID?

I'm also not familiar with DM RAID and how it differs in write patterns from MD RAID but seen some anecdotal reports that SMR disks will perform acceptably for MD RAID. As in not a 9-day resilver but a 24 hour resync deal (if a comparable CMR disk would take 18 hours to resilver). While it would be nice if I keep my zfs options open, if it works for MD or DM, it's fine for now...

(BTW, in terms of theoretical speeds, a 4TB disk would take a bit over 11 hours to write sequentially at 100MB/sec. At 18 hours to resilver even a CMR, it seems that zfs RAIDZ is doing a lot of metadata writes which may be what's killing things. I suspect a MD resync would also need to do metadata writes as it syncs which would likewise cause issues, though I wonder if it's any different as the vast majority of MD/DM resyncs are sequential writes, which theoretically an SMR disk should be able to optimize to one contiguous shingle minus the metadata.)
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?

[molletts]

You may be able to dodge the effects of shingling if you can find out how big the SMR zones are on the disk(s) you're planning to use by setting the RAID stride (chunk size) to be the same or a multiple and ensuring that the array is aligned appropriately to the zones.

As long as the RAID implementation always writes an entire chunk at a time (offhand, I don't know whether either DMRAID or MDRAID guarantees or can be configured to do this), the SMR drives should never have to do read-modify-write cycles. Don't forget to tell the filesystem about the underlying RAID geometry too so it can try to play nicely - for ext4, this would be by specifying '-E stride=x,stripe_width=y' on the mkfs.ext4 command line with x being the number of filesystem blocks per RAID chunk and y the number per stripe (chunks * data disks).

I've never used SMR disks so I don't know how effective this is in the real world.

[NeddySeagoon]

eccerr0r,

dmraid and mdraid use the same kernel code.
SMR is reported to kill raid altogether when a drive is slow to respond because its busy doing reshingling.
It gets kicked out of the raid set.
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

[Anon-E-moose]

I wouldn't mix cmr and smr drives.

What makes smr drives "bad" for any raid, is the amount of data that's streaming to the disk and not giving the disk time enough to push the data to the physical sectors before it goes back for more data.

ZFS doesn't handle smr drives too well, but not 100% sure why. (google would probably give results)

Edit to add: To me, the price difference isn't great enough to mess with smr drives.
_________________
PRIME x570-pro, 3700x, 6.1 zen kernel
gcc 13, profile 17.0 (custom bare multilib), openrc, wayland

[eccerr0r]