6 ms·

Rebuilding is not random i/o though, it goes sequentially and that's what SMRs are good at. Still, I did recently reshape from 4x6TB raid5 to 5x6TB raid6. The

Rebuilding is not random i/o though, it goes sequentially and that's what SMRs are good at. Still, I did recently reshape from 4x6TB raid5 to 5x6TB raid6. The new drive was SMR WD Red, it took 12 days and I was quite pissed when I found out the SMR bit. It wasn't in the datasheet, if I knew upfront, I would never bought it.

syshum 7y ago
>> I knew upfront, I would never bought it. SMR has a terrible rep for a reason, and no one with any knowlege of what SMR is would buy which is why they have to be unethical by removing this spec from the spec sheets SMR is cheaper for them to manufacture, and is TERRIBLE technology that really has very little actual use case, except Archive storage. The manufacturers want to push SMR for more than that because it is more profitable for them
- zozbot234 7y ago
  > SMR is cheaper for them to manufacture, and is TERRIBLE technology It's not that terrible if the customer knows what they're getting. It's just that SMR competes with tape more so than with conventional (CMR) hard drives. Many workloads are good enough for SMR.
  - syshum 7y ago
    >>Many workloads are good enough for SMR. One workload is, which is what you said.. Archiving, what one would normally do to Tape. Everything else is is substandard and should not be used
    - chmod775 7y ago
      But what if you need random reads? SMR sounds like it's great (value) at providing random read access to large amounts of data if you only have sequential writes or infrequent writes.
      - samstave 7y ago
        Serious question: so basically wikipedia is the only use case?
        tfigment 7y ago
        Time series historian or some siem logging or similar journaling seem applicable to me. Maybe continuous loop video recording.
        samstave 7y ago
        Ah video makes sense. Thanks.
        eeZah7Ux 7y ago
        wikipedia is not a good use case.
        klodolph 7y ago
        You don’t need to write the entire drive sequentially with SMR, it’s just that you need to write sequentially within a given “chunk” of data. So there are some workloads that can, at least in theory, work on SMR. Think of object storage or log-structured data.
        toast0 7y ago
        With proper system design and tiered storage, it would fit a lot of archival uses if it's visible. Things like data warehousing, where you would want to store on something else for the current time period, but could store a big blob as the data gets finalized (once all nodes have reported in, or maybe after N days if you drop some columns at that time). Or household archival where data basically doesn't get changed once it's put on the storage --- wait for a block sized write to accumulate and then do it (need a different tier for the intermediate writes... Some SMR drives have a PMR section for this too). But, device managed SMR is not a very good solution, even if the use case would be good for SMR, because the filesystem is most likely not aware of the need to cluster writes.
        7y ago
    - wbl 7y ago
      Host managed SMR is similar to SSD in terms of what needs to be done.
      - monocasa 7y ago
        With a huge latency difference which makes them totally different in practice.
        wtallis 7y ago
        The latency difference means the two are suited for very different real-world workloads. But they're similar enough in the access patterns they prefer that the software work to make your IO fit those patterns can be shared between SMR and flash.
        monocasa 7y ago
        The latency difference means that the access patterns are extremely different in practice. Even random reads get really painful on SMR since they get structured as logs that need to be replayed. SMR is way more like tape than an SSD.
    - Dylan16807 7y ago
      With a filesystem that understands the layout, they're almost as good as normal hard drives. With a filesystem that understands the layout and $3 of flash, you can make something that solidly outperforms a normal hard drive, at a much lower cost. Also tape drives are too expensive for normal consumers.
  - bsanr2 7y ago
    >It's not that terrible if the customer knows what they're getting. Between cigarettes, credit default swaps, and asbestos, I think this is a phrase we should become very wary of saying with a straight face.
- 8fingerlouie 7y ago
  > SMR has a terrible rep for a reason, and no one with any knowlege of what SMR is would buy which is why they have to be unethical by removing this spec from the spec sheets I have a couple of 8 TB Seagate SMR drives i use for backup, and for that purpose they've been great. They're cheaper than PMR drives, and i honestly don't care if the backup takes an hour extra to complete. Unless you're in a business scenario, how much of your data on your x*8TB drives at home do you actually write to on a regular basis ? Half of mine is family photos, backups of clients, movie/music media, etc. For this purpose SMR also works well. I'm far more worried about power consumption for NAS drives than write performance.
  - hu3 7y ago
    To be fair your parent did say SMR is fine for your usecase of backups (archive).
- bluedino 7y ago
  Dropbox has half their data on SMR drives. They did customize their storage software to work well with SMR, however.
  - mikepurvis 7y ago
    They keep around old versions of your files for a time anyway, so the use case is almost a perfect fit.
- TheKarateKid 7y ago
  I’m still confused how SMR is suitable even for archive storage. The drives slow down tremendously when doing long sequential writes. I don’t think tape even does this.
  - justinclift 7y ago
    Tape is pretty constant. It has a (let say) "optimum" throughput, and as long as you get the data to the drive controller fast enough that's what it'll do. If you don't then the tape drive write speed falls off pretty quick as it does stuff to cope.
- ashtonkem 7y ago
  Archival storage shouldn’t be use anything other than tape, in my opinion.
tjoff 7y ago
With regular raid yes, not with something like ZFS. (which is often argued as something positive but I would argue is a bad aspect of ZFS (regardless of whether SMRs are used or not))
- magicalhippo 7y ago
  There's been work done to make ZFS rebuilding (or resilvering as ZFS calls it) less random. See for example "New prefetcher for sequential scrub" and "Improving resilver: results & operational impacts" from http://www.open-zfs.org/wiki/OpenZFS_Developer_Summit_2017 http://www.open-zfs.org/wiki/OpenZFS_Developer_Summit_2017
- seabrookmx 7y ago
  Yeah I just upgraded a ZFS mirror from 2 x 3TB to 2 x 8TB, using Seagate SMR drives (picked them up for $185 CAD which is super cheap). This is for a media server so figured SMR would be fine, but the resilver took 28hours! No checksum errors though, so now that it's done things seem to be humming along fine.
  - jandrese 7y ago
    That has to hurt the aggregate write performance on the array. You hand it a big chunk of data and it goes ok, calculate the FEC and write to disks 1, 2, 3, 4. And disk 1 is done, disk 2 is done, disk 4 is done, ... , ... , ... , disk 3 is done, now to tell the OS the write is complete and to send me the next chunk.
marcan_42 7y ago
DM-SMRs aren't good at sequential I/O. They're bad, just not pathologically bad. SMR drives still need to stage sequential I/O to the cache area, but they at least get to write it out without a ton of write amplification. With extra inefficiencies, yeah, SMR drives will only be a few times slower than regular drives for sustained sequential write workloads, as opposed to thousands of times slower for random writes. Real host-managed SMR with SMR aware filesystems and storage engines might be OK, but that's not a thing in the consumer space.

syshum 7y ago

>> I knew upfront, I would never bought it. SMR has a terrible rep for a reason, and no one with any knowlege of what SMR is would buy which is why they have to be unethical by removing this spec from the spec sheets SMR is cheaper for them to manufacture, and is TERRIBLE technology that really has very little actual use case, except Archive storage. The manufacturers want to push SMR for more than that because it is more profitable for them

zozbot234 7y ago
> SMR is cheaper for them to manufacture, and is TERRIBLE technology It's not that terrible if the customer knows what they're getting. It's just that SMR competes with tape more so than with conventional (CMR) hard drives. Many workloads are good enough for SMR.
- syshum 7y ago
  >>Many workloads are good enough for SMR. One workload is, which is what you said.. Archiving, what one would normally do to Tape. Everything else is is substandard and should not be used
  - chmod775 7y ago
    But what if you need random reads? SMR sounds like it's great (value) at providing random read access to large amounts of data if you only have sequential writes or infrequent writes.
    - samstave 7y ago
      Serious question: so basically wikipedia is the only use case?
      - tfigment 7y ago
        Time series historian or some siem logging or similar journaling seem applicable to me. Maybe continuous loop video recording.
        samstave 7y ago
        Ah video makes sense. Thanks.
      - eeZah7Ux 7y ago
        wikipedia is not a good use case.
      - klodolph 7y ago
        You don’t need to write the entire drive sequentially with SMR, it’s just that you need to write sequentially within a given “chunk” of data. So there are some workloads that can, at least in theory, work on SMR. Think of object storage or log-structured data.
      - toast0 7y ago
        With proper system design and tiered storage, it would fit a lot of archival uses if it's visible. Things like data warehousing, where you would want to store on something else for the current time period, but could store a big blob as the data gets finalized (once all nodes have reported in, or maybe after N days if you drop some columns at that time). Or household archival where data basically doesn't get changed once it's put on the storage --- wait for a block sized write to accumulate and then do it (need a different tier for the intermediate writes... Some SMR drives have a PMR section for this too). But, device managed SMR is not a very good solution, even if the use case would be good for SMR, because the filesystem is most likely not aware of the need to cluster writes.
      - 7y ago
  - wbl 7y ago
    Host managed SMR is similar to SSD in terms of what needs to be done.
    - monocasa 7y ago
      With a huge latency difference which makes them totally different in practice.
      - wtallis 7y ago
        The latency difference means the two are suited for very different real-world workloads. But they're similar enough in the access patterns they prefer that the software work to make your IO fit those patterns can be shared between SMR and flash.
        monocasa 7y ago
        The latency difference means that the access patterns are extremely different in practice. Even random reads get really painful on SMR since they get structured as logs that need to be replayed. SMR is way more like tape than an SSD.
  - Dylan16807 7y ago
    With a filesystem that understands the layout, they're almost as good as normal hard drives. With a filesystem that understands the layout and $3 of flash, you can make something that solidly outperforms a normal hard drive, at a much lower cost. Also tape drives are too expensive for normal consumers.
- bsanr2 7y ago
  >It's not that terrible if the customer knows what they're getting. Between cigarettes, credit default swaps, and asbestos, I think this is a phrase we should become very wary of saying with a straight face.
8fingerlouie 7y ago
> SMR has a terrible rep for a reason, and no one with any knowlege of what SMR is would buy which is why they have to be unethical by removing this spec from the spec sheets I have a couple of 8 TB Seagate SMR drives i use for backup, and for that purpose they've been great. They're cheaper than PMR drives, and i honestly don't care if the backup takes an hour extra to complete. Unless you're in a business scenario, how much of your data on your x*8TB drives at home do you actually write to on a regular basis ? Half of mine is family photos, backups of clients, movie/music media, etc. For this purpose SMR also works well. I'm far more worried about power consumption for NAS drives than write performance.
- hu3 7y ago
  To be fair your parent did say SMR is fine for your usecase of backups (archive).
bluedino 7y ago
Dropbox has half their data on SMR drives. They did customize their storage software to work well with SMR, however.
- mikepurvis 7y ago
  They keep around old versions of your files for a time anyway, so the use case is almost a perfect fit.
TheKarateKid 7y ago
I’m still confused how SMR is suitable even for archive storage. The drives slow down tremendously when doing long sequential writes. I don’t think tape even does this.
- justinclift 7y ago
  Tape is pretty constant. It has a (let say) "optimum" throughput, and as long as you get the data to the drive controller fast enough that's what it'll do. If you don't then the tape drive write speed falls off pretty quick as it does stuff to cope.
ashtonkem 7y ago
Archival storage shouldn’t be use anything other than tape, in my opinion.

tjoff 7y ago

With regular raid yes, not with something like ZFS. (which is often argued as something positive but I would argue is a bad aspect of ZFS (regardless of whether SMRs are used or not))

magicalhippo 7y ago
There's been work done to make ZFS rebuilding (or resilvering as ZFS calls it) less random. See for example "New prefetcher for sequential scrub" and "Improving resilver: results & operational impacts" from http://www.open-zfs.org/wiki/OpenZFS_Developer_Summit_2017 http://www.open-zfs.org/wiki/OpenZFS_Developer_Summit_2017
seabrookmx 7y ago
Yeah I just upgraded a ZFS mirror from 2 x 3TB to 2 x 8TB, using Seagate SMR drives (picked them up for $185 CAD which is super cheap). This is for a media server so figured SMR would be fine, but the resilver took 28hours! No checksum errors though, so now that it's done things seem to be humming along fine.
- jandrese 7y ago
  That has to hurt the aggregate write performance on the array. You hand it a big chunk of data and it goes ok, calculate the FEC and write to disks 1, 2, 3, 4. And disk 1 is done, disk 2 is done, disk 4 is done, ... , ... , ... , disk 3 is done, now to tell the OS the write is complete and to send me the next chunk.

marcan_42 7y ago

DM-SMRs aren't good at sequential I/O. They're bad, just not pathologically bad. SMR drives still need to stage sequential I/O to the cache area, but they at least get to write it out without a ton of write amplification. With extra inefficiencies, yeah, SMR drives will only be a few times slower than regular drives for sustained sequential write workloads, as opposed to thousands of times slower for random writes. Real host-managed SMR with SMR aware filesystems and storage engines might be OK, but that's not a thing in the consumer space.