5 ms·

Western Digital: We Are Sampling 32TB SMR Hard Drives

91 points by mfiguiere 2y ago

paulkrush 2y ago
"We are shipping samples of our 32TB UltraSMR/ePMR nearline hard drives to select customers"
amelius 2y ago
In case anybody got curious about the triple stage actuators, here is a nice article: https://www.mdpi.com/2076-0825/8/3/65 https://www.mdpi.com/2076-0825/8/3/65
- ChainOfFools 2y ago
  I wonder why drives still use mechanical actuators to fling a single read write head back and forth across the surface, rather than a bar fixed in place that spans entire radius of the writable surface, with an addressable strip of r/w heads that runs the length of it. I assume it's because the actual read-write head itself must among the most expensive parts to manufacture, but is there no photo mask approach to this that can manufacture such a head array (or staggered series of them if there is some interference problem that prevents putting them side by side in a single strip) the same way but processor dies are made?
  - dmitrygr 2y ago
    > with an addressable strip of r/w heads Because a typical drive has MILLIONS of tracks, and an RW head is quite wide, compared to its very narrow active area. It would not work. Let's try some math: A typical drive has 1.3 Tb/in^2 density. Assuming bits are approximately square (have a 1:1 aspect ratio), and the drive is locally linear, we can calculate that (ignoring ECC and metadata, which make the task even harder) that each square inch is 1.2e6 x 1.2e6 bits, meaning that you'd need a strip with a density of 1.2e6 read/write areas per inch. Assuming an outer platter diameter of 3 inches and inner diameter of 1 inch, you'd thus need this strip to be one inch long. So you need to engineer 1.2e6 RW heads instead of one and somehow fit 1.2e6 of them within one inch. I cannot even imagine how you'd do this, as each RW head needs to be a coil, and some way to get the data out (and possibly even a local amp)
    - ChainOfFools 2y ago
      What if the bar were able to travel linerarly via a voice coil, and had hundreds of heads, but the bar was able to shift in a linear fashion only about a micron back and forth in increments of thousandths of a micron. Wouldn't that get you to your millions of heads of effective track coverage, without all the complexity of having to move two or three different hinges on several different actuators sweeping the entire arc? Naively speaking it seems like you could mount such a bar even closer to the surface of the track since there is no chance of a head crash, as it could be supported at both ends of its mounting. Surely the entire armature that moves an articulated set of actuator assemblies plus the head itself would be heavier and more cumbersome and failure-proneto move than a linearly-fixed rail. Obviously someone has thought of this before me so there's a good reason for it, but track density alone can't be it.
      - wtallis 2y ago
        What problem does your proposal potentially solve? A strip of heads that move in unison won't be able to operate in parallel, for the same reason that on real hard drives heads serving different platters cannot operate in parallel without being mounted on independent actuators. So it seems like your suggestion only reduces the distance that any given head needs to move, at the cost of greatly increasing the mass of heads+wiring that needs to be positioned quickly and accurately. You've seriously inflated cost for no gain in bandwidth and likely no gain in latency (and even if everything worked out in favor, it couldn't do anything to improve rotational latency).
        ChainOfFools 2y ago
        I don't know I keep getting answers to different proposals than the one that I have made, all with a significant detail changed so that the answer makes sense and the proposal sounds absurd. In this case, I don't know how the impression could be gained that I was suggesting mounting 100 or so individual assemblies of the very same read write head currently used on mechanical drives, each with its own separate wiring harness. Obviously, yes that would substantially increase the cost and the mass and everything else, you're right. Thanks for setting me straight about the impracticality of very different solution than the one I described, which suggested the heads could be manufactured as a single integrated unit, the way every other matrix of active elements (OLED displays, memory cells, someone else mentioned mems, etc.) is made, which might require two additional wires for selection signalling, if that, depending on what could be multiplexed. In exchange for this weight and complexity you get rid of one, possibly two extra servo joints, one possibly two extra armatures, plus all the wiring and additional control logic required for positioning this very delicate articulated assembly. Now, if this array cannot be manufactured because no current solid state process exists that can reproduce the characteristics of a wound coil, duplicated a hundred times in a linear array of cells on silicon or some other substrate, that's an obstacle I can reason about and accept as a plausible dead end.
  - akira2501 2y ago
    The article itself gives you a good reason. Differential thermal properties which can be compensated for in the servo system but not in your fixed amerature.
    - ChainOfFools 2y ago
      The article does mention that this is a complication particular to SMR drives, but I was thinking about this not necessarily as a means of reaching higher density storage, but more reliable and perhaps higher performance solutions that would fall somewhere between the compromises of an SSD and a mechanical drive. It also seems possible that instead of a single super high density strip of heads which may be impossible to manufacture, that instead a series of much lower density array strips could be mounted in n ranks, each rank offset from the predecessor by n/(distance between heads in each strip). I assume this is not an insurmountable engineering, but that someone has already had this idea or something similar to it, and after running the numbers found it doesn't deliver enough of an advantage over the conventional design to be worth exploring further.
  - phire 2y ago
    I suspect the head flying just above the surface has something to do with it. A bar would need to be held at a fixed height above the surface. The flying head dynamically adjusts to both distortions across the surface, and temporal variations as the platter changes temperature. This height adjustment is entirely passive, dependant on aerodynamics rather than any electronics.
    - ChainOfFools 2y ago
      I remember someone explaining this to me years ago when IOmega released those removable Bernoulli drive cartridges. The (for the time) high densities were possible because even though the recording medium was a flexible bit of plastic, the head was designed to draw surface of the media up to itself using the namesake effect, rather than relying entirely om precision machine tolerances. And this effect was dynamically self limiting, such that there could not be a head crash as long is the unit wasnt disturbed during operation. What I did not know is that this is still and effect that is relied upon in modern, ultra high precision, fixed medium drives.
  - to11mtm 2y ago
    > I assume it's because the actual read-write head itself must among the most expensive parts to manufacture, Ehh a modern HDD is a marvel of precision manufacturing and engineering. > rather than a bar fixed in place that spans entire radius of the writable surface, with an addressable strip of r/w heads that runs the length of it. Your bigger challenge is getting a set of read/write heads that are across the platter able to maintain a consistent height/etc. The closest I'm aware of, are the Connor 'Chinook' drives [0] that had two actuators, one on each side of the drive. Broadly speaking, I'd ask if multiple heads too close throw off decades of head design experience... [0] - https://en.wikipedia.org/wiki/Conner_Peripherals#/media/File:Conner_Peripherals_%22Chinook%22_dual-actuator_drive.jpg https://en.wikipedia.org/wiki/Conner_Peripherals#/media/File...
  - throwaway81523 2y ago
    > bar fixed in place that spans entire radius of the writable surface, with an addressable strip of r/w heads that runs the length of it. That was called a fixed-head disk. They got rid of seek latency (still had rotational latency) and so there was a market niche for them as (e.g.) swap devices, but they had lower data density than moving-head disks did, so they went out of style. It looks like they are so obscure today that there's no wikipedia article, but here's an article from some other wiki: https://gunkies.org/wiki/Fixed-head_disk https://gunkies.org/wiki/Fixed-head_disk They are briefly mentioned in the general Wikipedia article on hard disk drives. There was also "drum memory" where there was a head per track on a cylindrical recording surface. Those seem to have dried up even earlier: https://en.wikipedia.org/wiki/Drum_memory https://en.wikipedia.org/wiki/Drum_memory Here's an info page about the DEC RS03 and RS04 from the 1970s. The RS04 weighed 120 pounds and had 1MB of capacity. https://www.storagenewsletter.com/2019/02/19/history-1974-dec-rs03-and-rs04-hdds/ https://www.storagenewsletter.com/2019/02/19/history-1974-de...
    - defrost 2y ago
      Drum memory is a particular case of Delay line Memory and what goes around comes around, .. it's back baby*: https://en.wikipedia.org/wiki/Delay-line_memory https://en.wikipedia.org/wiki/Delay-line_memory https://www.rp-photonics.com/optical_delay_lines.html https://www.rp-photonics.com/optical_delay_lines.html ( * For specific use cases ) There's probably an insane data throughput domain that requires autocorrelation filters that might benefit from an optic fibre bundle in which each fibre is slightly shorter than the previous one.
      - 082349872349872 2y ago
        compare http://tom7.org/papers/murphy2022harder.pdf http://tom7.org/papers/murphy2022harder.pdf
        defrost 2y ago
        Thank you, it's rare to find good papers anymore ... this, perhaps, was delayed 6 days in publishing out of pure spite no doubt
laluser 2y ago
My team is testing these out for our on-perm storage system. Not much to say vs previous generations unless you’re in the hardware engineering space.
Animats 2y ago
Seagate is already selling 30TB hard drives. But how do you buy a non-fake one?
- inhumantsar 2y ago
  you contact your disk array vendor account rep and ask for a couple dozen.
timschmidt 2y ago
The 8Tb WD SMR drives I have use a 40gb chunk of the surface for non-SMR reads and writes as a buffer. Altering anything in an SMR track requires copying the data from that track and tracks near it into the buffer, then laying them back down into the SMR region sequentially. Like erasing a flash block before writing. Unfortunately, this means that the drive's performance takes a dump from 200mb/s down to 300kb/s if you do enough transfers to fill the 40gb buffer, or attempt multiple transfers at once.
- 01HNNWZ0MV43FF 2y ago
  I sure hope it honors fsync!
  - AlexandrB 2y ago
    Not sure, but interestingly SMR drives support TRIM commands: https://superuser.com/questions/1407990/what-does-trim-on-an-hdd-mean https://superuser.com/questions/1407990/what-does-trim-on-an...
- matrix_overload 2y ago
  They still have valid use cases like backups or unedited video footage. It's just kinda lame that the manufacturers don't market them as "slow backup devices" clearly listing all limitations, and that you have to find it out first time you use it.
  - timschmidt 2y ago
    When I have used them for backups, the 40gb buffer quickly fills and then I'm stuck with speeds slower than my internet connection until the buffer empties, which it will only do if I kill the existing transfer. SSDs can dump 40gb of data very quickly. Annoyingly, since the r/w head has to do it's back-and-forth dance between the buffer area and the SMR areas, this condition even impacts read speeds. Consequently I wouldn't even use them for backups. Nor would I buy them again. I've set them up as Storj drives, which they seem to handle reasonably, and I am thankful to be done with them otherwise. If you're considering one, I would pay special attention to the buffer size, and ensure that all the transfers you want to do to or from the drive are significantly smaller than the buffer to ensure reasonable performance. That excludes most video too. Storj files are typically just a few megabytes, and typically arrive at a frequency of just one or two per second, which the drive can handle.
lofaszvanitt 2y ago
zzzz, where are the competing storage solutions?
watermelon0 2y ago
I might understand that businesses have use cases where price difference adds up, and they have workloads, where SMRs can satisfy the needs. However, what's the reason that they are selling SMRs to the consumers? Price difference in stores is not that big between SMR and CMRs, and there are a lot of downsides.
yencabulator 2y ago
To anyone complaining about SMR performance (and on Linux): try F2FS, it's designed to have more sequential write patterns.

91 points by mfiguiere 2y ago

"We are shipping samples of our 32TB UltraSMR/ePMR nearline hard drives to select customers"

In case anybody got curious about the triple stage actuators, here is a nice article: https://www.mdpi.com/2076-0825/8/3/65 https://www.mdpi.com/2076-0825/8/3/65

ChainOfFools 2y ago
I wonder why drives still use mechanical actuators to fling a single read write head back and forth across the surface, rather than a bar fixed in place that spans entire radius of the writable surface, with an addressable strip of r/w heads that runs the length of it. I assume it's because the actual read-write head itself must among the most expensive parts to manufacture, but is there no photo mask approach to this that can manufacture such a head array (or staggered series of them if there is some interference problem that prevents putting them side by side in a single strip) the same way but processor dies are made?
- dmitrygr 2y ago
  > with an addressable strip of r/w heads Because a typical drive has MILLIONS of tracks, and an RW head is quite wide, compared to its very narrow active area. It would not work. Let's try some math: A typical drive has 1.3 Tb/in^2 density. Assuming bits are approximately square (have a 1:1 aspect ratio), and the drive is locally linear, we can calculate that (ignoring ECC and metadata, which make the task even harder) that each square inch is 1.2e6 x 1.2e6 bits, meaning that you'd need a strip with a density of 1.2e6 read/write areas per inch. Assuming an outer platter diameter of 3 inches and inner diameter of 1 inch, you'd thus need this strip to be one inch long. So you need to engineer 1.2e6 RW heads instead of one and somehow fit 1.2e6 of them within one inch. I cannot even imagine how you'd do this, as each RW head needs to be a coil, and some way to get the data out (and possibly even a local amp)
  - ChainOfFools 2y ago
    What if the bar were able to travel linerarly via a voice coil, and had hundreds of heads, but the bar was able to shift in a linear fashion only about a micron back and forth in increments of thousandths of a micron. Wouldn't that get you to your millions of heads of effective track coverage, without all the complexity of having to move two or three different hinges on several different actuators sweeping the entire arc? Naively speaking it seems like you could mount such a bar even closer to the surface of the track since there is no chance of a head crash, as it could be supported at both ends of its mounting. Surely the entire armature that moves an articulated set of actuator assemblies plus the head itself would be heavier and more cumbersome and failure-proneto move than a linearly-fixed rail. Obviously someone has thought of this before me so there's a good reason for it, but track density alone can't be it.
    - wtallis 2y ago
      What problem does your proposal potentially solve? A strip of heads that move in unison won't be able to operate in parallel, for the same reason that on real hard drives heads serving different platters cannot operate in parallel without being mounted on independent actuators. So it seems like your suggestion only reduces the distance that any given head needs to move, at the cost of greatly increasing the mass of heads+wiring that needs to be positioned quickly and accurately. You've seriously inflated cost for no gain in bandwidth and likely no gain in latency (and even if everything worked out in favor, it couldn't do anything to improve rotational latency).
      - ChainOfFools 2y ago
        I don't know I keep getting answers to different proposals than the one that I have made, all with a significant detail changed so that the answer makes sense and the proposal sounds absurd. In this case, I don't know how the impression could be gained that I was suggesting mounting 100 or so individual assemblies of the very same read write head currently used on mechanical drives, each with its own separate wiring harness. Obviously, yes that would substantially increase the cost and the mass and everything else, you're right. Thanks for setting me straight about the impracticality of very different solution than the one I described, which suggested the heads could be manufactured as a single integrated unit, the way every other matrix of active elements (OLED displays, memory cells, someone else mentioned mems, etc.) is made, which might require two additional wires for selection signalling, if that, depending on what could be multiplexed. In exchange for this weight and complexity you get rid of one, possibly two extra servo joints, one possibly two extra armatures, plus all the wiring and additional control logic required for positioning this very delicate articulated assembly. Now, if this array cannot be manufactured because no current solid state process exists that can reproduce the characteristics of a wound coil, duplicated a hundred times in a linear array of cells on silicon or some other substrate, that's an obstacle I can reason about and accept as a plausible dead end.
- akira2501 2y ago
  The article itself gives you a good reason. Differential thermal properties which can be compensated for in the servo system but not in your fixed amerature.
  - ChainOfFools 2y ago
    The article does mention that this is a complication particular to SMR drives, but I was thinking about this not necessarily as a means of reaching higher density storage, but more reliable and perhaps higher performance solutions that would fall somewhere between the compromises of an SSD and a mechanical drive. It also seems possible that instead of a single super high density strip of heads which may be impossible to manufacture, that instead a series of much lower density array strips could be mounted in n ranks, each rank offset from the predecessor by n/(distance between heads in each strip). I assume this is not an insurmountable engineering, but that someone has already had this idea or something similar to it, and after running the numbers found it doesn't deliver enough of an advantage over the conventional design to be worth exploring further.
- phire 2y ago
  I suspect the head flying just above the surface has something to do with it. A bar would need to be held at a fixed height above the surface. The flying head dynamically adjusts to both distortions across the surface, and temporal variations as the platter changes temperature. This height adjustment is entirely passive, dependant on aerodynamics rather than any electronics.
  - ChainOfFools 2y ago
    I remember someone explaining this to me years ago when IOmega released those removable Bernoulli drive cartridges. The (for the time) high densities were possible because even though the recording medium was a flexible bit of plastic, the head was designed to draw surface of the media up to itself using the namesake effect, rather than relying entirely om precision machine tolerances. And this effect was dynamically self limiting, such that there could not be a head crash as long is the unit wasnt disturbed during operation. What I did not know is that this is still and effect that is relied upon in modern, ultra high precision, fixed medium drives.
- to11mtm 2y ago
  > I assume it's because the actual read-write head itself must among the most expensive parts to manufacture, Ehh a modern HDD is a marvel of precision manufacturing and engineering. > rather than a bar fixed in place that spans entire radius of the writable surface, with an addressable strip of r/w heads that runs the length of it. Your bigger challenge is getting a set of read/write heads that are across the platter able to maintain a consistent height/etc. The closest I'm aware of, are the Connor 'Chinook' drives [0] that had two actuators, one on each side of the drive. Broadly speaking, I'd ask if multiple heads too close throw off decades of head design experience... [0] - https://en.wikipedia.org/wiki/Conner_Peripherals#/media/File:Conner_Peripherals_%22Chinook%22_dual-actuator_drive.jpg https://en.wikipedia.org/wiki/Conner_Peripherals#/media/File...
- throwaway81523 2y ago
  > bar fixed in place that spans entire radius of the writable surface, with an addressable strip of r/w heads that runs the length of it. That was called a fixed-head disk. They got rid of seek latency (still had rotational latency) and so there was a market niche for them as (e.g.) swap devices, but they had lower data density than moving-head disks did, so they went out of style. It looks like they are so obscure today that there's no wikipedia article, but here's an article from some other wiki: https://gunkies.org/wiki/Fixed-head_disk https://gunkies.org/wiki/Fixed-head_disk They are briefly mentioned in the general Wikipedia article on hard disk drives. There was also "drum memory" where there was a head per track on a cylindrical recording surface. Those seem to have dried up even earlier: https://en.wikipedia.org/wiki/Drum_memory https://en.wikipedia.org/wiki/Drum_memory Here's an info page about the DEC RS03 and RS04 from the 1970s. The RS04 weighed 120 pounds and had 1MB of capacity. https://www.storagenewsletter.com/2019/02/19/history-1974-dec-rs03-and-rs04-hdds/ https://www.storagenewsletter.com/2019/02/19/history-1974-de...
  - defrost 2y ago
    Drum memory is a particular case of Delay line Memory and what goes around comes around, .. it's back baby*: https://en.wikipedia.org/wiki/Delay-line_memory https://en.wikipedia.org/wiki/Delay-line_memory https://www.rp-photonics.com/optical_delay_lines.html https://www.rp-photonics.com/optical_delay_lines.html ( * For specific use cases ) There's probably an insane data throughput domain that requires autocorrelation filters that might benefit from an optic fibre bundle in which each fibre is slightly shorter than the previous one.
    - 082349872349872 2y ago
      compare http://tom7.org/papers/murphy2022harder.pdf http://tom7.org/papers/murphy2022harder.pdf
      - defrost 2y ago
        Thank you, it's rare to find good papers anymore ... this, perhaps, was delayed 6 days in publishing out of pure spite no doubt

laluser 2y ago

My team is testing these out for our on-perm storage system. Not much to say vs previous generations unless you’re in the hardware engineering space.

Animats 2y ago

Seagate is already selling 30TB hard drives. But how do you buy a non-fake one?

inhumantsar 2y ago
you contact your disk array vendor account rep and ask for a couple dozen.

timschmidt 2y ago

The 8Tb WD SMR drives I have use a 40gb chunk of the surface for non-SMR reads and writes as a buffer. Altering anything in an SMR track requires copying the data from that track and tracks near it into the buffer, then laying them back down into the SMR region sequentially. Like erasing a flash block before writing. Unfortunately, this means that the drive's performance takes a dump from 200mb/s down to 300kb/s if you do enough transfers to fill the 40gb buffer, or attempt multiple transfers at once.

01HNNWZ0MV43FF 2y ago
I sure hope it honors fsync!
- AlexandrB 2y ago
  Not sure, but interestingly SMR drives support TRIM commands: https://superuser.com/questions/1407990/what-does-trim-on-an-hdd-mean https://superuser.com/questions/1407990/what-does-trim-on-an...
matrix_overload 2y ago
They still have valid use cases like backups or unedited video footage. It's just kinda lame that the manufacturers don't market them as "slow backup devices" clearly listing all limitations, and that you have to find it out first time you use it.
- timschmidt 2y ago
  When I have used them for backups, the 40gb buffer quickly fills and then I'm stuck with speeds slower than my internet connection until the buffer empties, which it will only do if I kill the existing transfer. SSDs can dump 40gb of data very quickly. Annoyingly, since the r/w head has to do it's back-and-forth dance between the buffer area and the SMR areas, this condition even impacts read speeds. Consequently I wouldn't even use them for backups. Nor would I buy them again. I've set them up as Storj drives, which they seem to handle reasonably, and I am thankful to be done with them otherwise. If you're considering one, I would pay special attention to the buffer size, and ensure that all the transfers you want to do to or from the drive are significantly smaller than the buffer to ensure reasonable performance. That excludes most video too. Storj files are typically just a few megabytes, and typically arrive at a frequency of just one or two per second, which the drive can handle.

lofaszvanitt 2y ago

zzzz, where are the competing storage solutions?

watermelon0 2y ago

I might understand that businesses have use cases where price difference adds up, and they have workloads, where SMRs can satisfy the needs. However, what's the reason that they are selling SMRs to the consumers? Price difference in stores is not that big between SMR and CMRs, and there are a lot of downsides.

yencabulator 2y ago

To anyone complaining about SMR performance (and on Linux): try F2FS, it's designed to have more sequential write patterns.