I want to like SMR disks, but WD pushing SMR disks into product lines where it doesn't belong, and the disks not being host-aware makes actually using them a losing proposition. Not to mention that SMR disks often aren't actually cheaper than equivalent CMR disks, at least in the consumer channels (the opposite is almost certainly the case in B2B, but I don't care about that).
If the disks were host-aware, Linux (and preferably ZFS) could handle host-aware SMR disks non-catastrophically, and SMR disks were part of a separate product line with prices and capacities that make some sort of sense when you compare with the CMR line, I'd be all over them as cold storage, or just treating them as (re)write-once-read-many drives.
But it's clear WD and Seagate don't care about the consumers. They're just a rounding error in comparison to B2B sales, especially given the current times.
As a relatively unserious data hoarder with a ZFS NAS, SMR drives are unusable, and are famous for resilver times ranging from weeks to never (drive times out). I'm sure there are workloads where they make sense (maybe archiving huge binary blobs). A while back Dropbox was using mostly SMR drives, so maybe the firmware and software ecosystem available for enterprise use cases is more approachable than I thought. And with the AI-driven price explosion for storage, perhaps the total cost of ownership even with the software investment starts to make sense.
SMR drives are the worst, even had issue resilvering a 2TB mirror.
What's frustrating is even just a little info from the manufactures would be beneficial, like the size of a shingle so the computer could batch writes, or even if the drive is shingled!
That's why libzbc exists. Host-aware and host-managed SMR give you the tools to optimize I/O around the batching needed. It doesn't change the fundamental outlook - you still have to organize writes in these huge blocks and do some math to figure out the best timing to commit cache to re-writes - but it's at least a tractable problem compared to chucking everything over the wall with device-managed SMR.
In any case, I don't think hobbyists get to play with the non-disk-managed drives, and anyone with the money and the special disks is probably not using ZFS because it's a bit of an uphill battle to fit SMR to a copy-on-write filesystem.
How come copy-on-write is not easier with SMR? Each SMR zone a.k.a. shingle or flash erase block is effectively a tiny tape drive. It is basically append only.
Shouldn't this be especially useful with ZFS doing copy-on-write with linked lists as a reversed heap?
+1 SMR drives are really great if you pack your data into large enough groups so that you occupy exacly N SMR zones, and allocate your files to specific SMR zones that you then only rewrite atomically. The sibling comment is on point - it's pretty similar to how to work effectively with SSDs.
The drive doesn't have enough information to do this itself, we need SMR aware filesystems.
There are very good deals currently on eBay for 14TB HM-SR drives! I just got quite a few - should be interesting to see how the current zone aware filesystems perform. XFS with an optane metadata device? Could be pretty okay!
When SSDs appeared, many filesystems performed badly on them, or degraded them much more quickly than needed, because they were not SSD-aware. It took a while to fix the issues, add TRIM, etc.
Sounds like ZFS has the same problem today, it's not aware of SMR, and it treats those drives as if they are regular hard-drives.
Which also makes it super clear that the intended use case is customers willing to use SMR-aware software to get a generation or two of density-per-bay improvement, not folks trying to jam it into current workflows.
THE issue with SMR is that they use designated area on the disk on the same head as disk as staging area.
They do like "ls /sda0/dirty-tracks/ | xargs -P 1 dd if=/sda0/dirty-tracks/{} of=/dev/sda{} size=512M && rm /sda0/{}", but onboard the disk, not on the host, and that's both wildly slow and imcompatible with anything but super light or niche use cases.
If only they at least used a NAND flash for that staging or at least coordinated with the host system, but no, HDD industry chose not to cooperate with anyone and to hide the fact that some disks are DM-SMRs.
More capacity density, if you completely ignore IO density. But many HDD workloads were already IO bound, and SMR is much, much worse. (HDD vendors won't sell CMR to hyperscalers anymore -- it's like they are trying to get us to buy QLC flash instead of their products.)
Think of SMR as "fast tape with random reads" rather than anything like a CMR product.
I think the future of hard drives involves absolutely huge error correction block sizes - 1 gigabyte or more.
The way it would work is:
* The underlying disk would do sequential writes only, in 1Gb blocks, rather like a log filesystem.
* The large block size enables much better information density due to being able to tolerate far more errors.
* Random reads within blocks can usually be done - but occasionally most or all of the whole block will need to be read to decode the data.
* A small capacitor+disk momentum ensures any writes that happened before a power failure can always be completed - either to the real disk or to flash.
* The OS still only sees 4k pages - versioning, packing those into 1Gb blocks, mappings and garbage collection are all handled inside the drive.
> any writes that happened before a power failure can always be completed
It can be also the data milliseconds after loosing power. The minimal amount of data in question is the data that the drive has acknowledged to the OS after a special command.
Anyway, this problem is already solved at least few years ago.
Has Western Digital found yet another way to further degrade the performance of their drives beyond their initial deployment of OptiNAND and SMR? At this rate the tail latency of a tape robot is going to outperform WD spinning rust.
I want to like SMR disks, but WD pushing SMR disks into product lines where it doesn't belong, and the disks not being host-aware makes actually using them a losing proposition. Not to mention that SMR disks often aren't actually cheaper than equivalent CMR disks, at least in the consumer channels (the opposite is almost certainly the case in B2B, but I don't care about that).
If the disks were host-aware, Linux (and preferably ZFS) could handle host-aware SMR disks non-catastrophically, and SMR disks were part of a separate product line with prices and capacities that make some sort of sense when you compare with the CMR line, I'd be all over them as cold storage, or just treating them as (re)write-once-read-many drives.
But it's clear WD and Seagate don't care about the consumers. They're just a rounding error in comparison to B2B sales, especially given the current times.
As a relatively unserious data hoarder with a ZFS NAS, SMR drives are unusable, and are famous for resilver times ranging from weeks to never (drive times out). I'm sure there are workloads where they make sense (maybe archiving huge binary blobs). A while back Dropbox was using mostly SMR drives, so maybe the firmware and software ecosystem available for enterprise use cases is more approachable than I thought. And with the AI-driven price explosion for storage, perhaps the total cost of ownership even with the software investment starts to make sense.
SMR drives are the worst, even had issue resilvering a 2TB mirror.
What's frustrating is even just a little info from the manufactures would be beneficial, like the size of a shingle so the computer could batch writes, or even if the drive is shingled!
That's why libzbc exists. Host-aware and host-managed SMR give you the tools to optimize I/O around the batching needed. It doesn't change the fundamental outlook - you still have to organize writes in these huge blocks and do some math to figure out the best timing to commit cache to re-writes - but it's at least a tractable problem compared to chucking everything over the wall with device-managed SMR.
In any case, I don't think hobbyists get to play with the non-disk-managed drives, and anyone with the money and the special disks is probably not using ZFS because it's a bit of an uphill battle to fit SMR to a copy-on-write filesystem.
How come copy-on-write is not easier with SMR? Each SMR zone a.k.a. shingle or flash erase block is effectively a tiny tape drive. It is basically append only.
Shouldn't this be especially useful with ZFS doing copy-on-write with linked lists as a reversed heap?
+1 SMR drives are really great if you pack your data into large enough groups so that you occupy exacly N SMR zones, and allocate your files to specific SMR zones that you then only rewrite atomically. The sibling comment is on point - it's pretty similar to how to work effectively with SSDs.
The drive doesn't have enough information to do this itself, we need SMR aware filesystems.
What we need is host-based SMR. SMR-aware filesystems exist (e.g. Zoned XFS, btrfs' zone mode, f2fs' zone mode).
There are very good deals currently on eBay for 14TB HM-SR drives! I just got quite a few - should be interesting to see how the current zone aware filesystems perform. XFS with an optane metadata device? Could be pretty okay!
Handling SMRs well probably would require a custom replacement for SPA (block layer) component in ZFS, talking directly with DMU (object storage).
Could also expose the drives as SCSI object storage devices
And all of that for a pretty paltry gain in capacity.
Uh, they do advertise zone size. At least, host-aware SMR. Yes, yes, host-unaware SMR is even more terrible.
I read a nice write-up of Dropbox's use of SMR drives here:
https://dropbox.tech/infrastructure/four-years-of-smr-storag...
This is a bug/design flaw in ZFS, they work fine in md IME.
When SSDs appeared, many filesystems performed badly on them, or degraded them much more quickly than needed, because they were not SSD-aware. It took a while to fix the issues, add TRIM, etc.
Sounds like ZFS has the same problem today, it's not aware of SMR, and it treats those drives as if they are regular hard-drives.
I could only find an explanation of what it actually is here: https://documents.westerndigital.com/content/dam/doc-library...
Which also makes it super clear that the intended use case is customers willing to use SMR-aware software to get a generation or two of density-per-bay improvement, not folks trying to jam it into current workflows.
Thanks. Looks like SMR drives will get a bit more dense, but with the same limitations as before — no update in place without bunch of rewriting.
Means they are still bad in RAID setups
Host-Managed SMR is a good middle-ground, but the file system support is still poor.
Exactly. I already use a CoW filesystem so SMR ought to be a non-issue but last I checked the software side of things wasn't ready yet.
Hardware-RAID doesn't make sense for more than 10 years now.
THE issue with SMR is that they use designated area on the disk on the same head as disk as staging area.
They do like "ls /sda0/dirty-tracks/ | xargs -P 1 dd if=/sda0/dirty-tracks/{} of=/dev/sda{} size=512M && rm /sda0/{}", but onboard the disk, not on the host, and that's both wildly slow and imcompatible with anything but super light or niche use cases.
If only they at least used a NAND flash for that staging or at least coordinated with the host system, but no, HDD industry chose not to cooperate with anyone and to hide the fact that some disks are DM-SMRs.
These "UltraSMR" don't work that way.
More capacity density, if you completely ignore IO density. But many HDD workloads were already IO bound, and SMR is much, much worse. (HDD vendors won't sell CMR to hyperscalers anymore -- it's like they are trying to get us to buy QLC flash instead of their products.)
Think of SMR as "fast tape with random reads" rather than anything like a CMR product.
I think the future of hard drives involves absolutely huge error correction block sizes - 1 gigabyte or more.
The way it would work is:
* The underlying disk would do sequential writes only, in 1Gb blocks, rather like a log filesystem.
* The large block size enables much better information density due to being able to tolerate far more errors.
* Random reads within blocks can usually be done - but occasionally most or all of the whole block will need to be read to decode the data.
* A small capacitor+disk momentum ensures any writes that happened before a power failure can always be completed - either to the real disk or to flash.
* The OS still only sees 4k pages - versioning, packing those into 1Gb blocks, mappings and garbage collection are all handled inside the drive.
> any writes that happened before a power failure can always be completed
It can be also the data milliseconds after loosing power. The minimal amount of data in question is the data that the drive has acknowledged to the OS after a special command.
Anyway, this problem is already solved at least few years ago.
Most enterprise storage arrays have writes going to some kind of very fast NVRAM and batched so that data can be written to storage in larger blocks.
SMR is honestly just an "avoid at all costs" technology, abysmal performance in any application.
Has Western Digital found yet another way to further degrade the performance of their drives beyond their initial deployment of OptiNAND and SMR? At this rate the tail latency of a tape robot is going to outperform WD spinning rust.