WD 22TB CMR and 26TB SMR HDDs Deliver Small Incremental Capacity Jump

[Cliff Robinson]

WD has 22TB Ultrastar, Red Pro, Purple, and Gold drives for CMR markets and a 26TB SMR Ultrastar for lower cost and performance storage

The post WD 22TB CMR and 26TB SMR HDDs Deliver Small Incremental Capacity Jump appeared first on ServeTheHome.

Continue reading...

 

[oneplane]

Regardless of the rated workload oddities this makes me wonder if we're going to have SMR-awareness in filesystems and logical volume managers because at this point the internal implementation of the recording method is having too much of an impact to be "the disk controller's problem" and we can't really depend on just the interface (be it ATA or SCSI commands) being enough to know how to handle the disk in the most efficient way.

 

[Terry Kennedy]

Regardless of the rated workload oddities this makes me wonder if we're going to have SMR-awareness in filesystems and logical volume managers because at this point the internal implementation of the recording method is having too much of an impact to be "the disk controller's problem" and we can't really depend on just the interface (be it ATA or SCSI commands) being enough to know how to handle the disk in the most efficient way.

We had the same thing when 512e drives came out - certain workloads exhibited abysmal performance due to non-aligned writes. We got somewhat better partition managers to ensure that partitions started on a native 4K boundary. And that was enough to make a majority of users happy (possibly because they don't know any better). Some of us only buy 4Kn drives to avoid that issue entirely.

And we had a combination of Band-Aids on many filesystems and somewhat-clever firmware in SSD drives so that they didn't exhibit incredibly bad performance under normal workloads. Just thinking of Windows XP's defragmentation tool running on a SSD makes me cringe.

I don't see detailed specs including how much RAM and flash is available on those drives for acceleration - maybe WD is hoping to similarly compensate for SMR limitations with huge amounts of caching. Given that we're seeing DRAMless 1TB SSDs with only a single flash chip and controller chip, that may be what they're thinking. Are there actually any production-grade filesystems that are HA-SMR aware?

Regarding the original article's comments about the small number of remaining drive manufacturers, I don't really think that is the limiting factor here - if it was practical to go beyond 10 or 11 platters or increase areal density, I'm pretty sure either WD or Seagate would have released those. And I've seen some bizarre designs in my time (beyond 15K RPM, linear head no-actuator designs, etc.). I do think the duopoly is leading to products differentiated only by firmware (and marketing things like price and warranty).

 

[oneplane]

I suspect we're also running into physical limits of the current technology. We're making the materials thinner and running them faster with closer tolerances, but once we're in the microns it becomes very difficult to "do the same thing, but better". SMR as well as helium-filled drives just smell like life-support efforts to get the most out of what we have without having to create a full replacement (whatever that may be).

While I don't think we'll get hardware-aware filesystems any time soon (at least not beyond 512/520/528 and 4k awareness), we're going to have to eventually, even just for hierarchical storage on device level where the OS and the drive have to work together to find out what blocks go where and at what guarantee in performance and reliability/availability/durability. Simply 'dumping data to disk' and having it try to find out when to flush and what to keep warm isn't going to be enough when SMR and the likes are all we're getting out of drive manufacturers as 'improvements'.

Then again, in the solid state space it isn't that much better. We're either getting weird RAID-ish cards doing opaque things with SSDs or Apple-style Flash-on-a-board disks that don't have a controller at all. The latter might even be the better way forward considering the lack of transparency and awareness in the SSD controllers. So far all I have seen (which isn't that much variation TBH) is controllers just guessing what data types and structures are flowing through it with no information sharing with the OS at all.