3 ms·

Nice to see the product brief break down the write endurance rating by workload. They really want you to use at least 16kB block sizes with this drive. NAND pag

Nice to see the product brief break down the write endurance rating by workload. They really want you to use at least 16kB block sizes with this drive. NAND page sizes have been larger than 4kB for quite a while now, but most SSDs still put significant effort into handling 4kB IO operations well. This market segment of server drives optimized for capacity over performance or endurance is where vendors are most willing to be up front about how your workload really should be appropriate for the drive.

dylan604 2y ago
Or it could be a boon for video/audio/graphics types. I’m used to setting up my spinning rust arrays with 64k block sizes. Continuous reads from large files vs lots of tiny files is something that tripped up a lot of storage vendors trying to break into the media world.
- bufferoverflow 2y ago
  It's very expensive. Gen5 SSDs are around $100/TB. Hard drives are $16/TB. It makes sense to have a small fast SSD for editing, but HDDs for mass-storage. And with video you usually want redundancy, storage gets expensive fast.
  - dylan604 2y ago
    And with any data you usually want redundancy, FTFY But I know what you meant. I used to have systems built with multiple small arrays where the original sources were in RAID-5 (pre-raid6). The software would read from that array and write to a RAID0 array as these files were only ephemeral. The software would then read from that RAID0 and write to a second RAID0 to avoid reading/writing to the same volume. This was all before SSDs. For the lulz, I would love to see how that workflow would perform with NVMe storage. At the time of using these systems, spinning rust was not $16/TB and was probably closer to the $100/TB. I still remember $1k/GB in the early 90s
    - bufferoverflow 2y ago
      At the time you didn't have 8K or RAW videos or 200-megapixel photos. There's never enough storage.
      - dylan604 2y ago
        While not 8k, it was 4K-5k and definitely RAW. While never enough storage, we have achieved a point where disk I/O is not a bottle neck. It used to take a raid0 7 disk stripe to record SD video only while the audio was written to a separate drive. The fact we can record 8k RAW with audio on a single drive the size of a stick of RAM is crazy to compare. Smaller faster cheaper.
starspangled 2y ago
Physical NAND pages aren't really relevant. At least some high end SSDs I know of would gang multiple NAND chips together anyway and effectively the page size would be X*physical (where X might be 8 or 16). It could still do 512 byte writes well if it was configured that way. The limiting factor is the size of the forward map structures. To a lesser degree also things like garbage collection overheads and reverse map structures (particularly if you support deduplication or snapshots). But forward map is the main thing.
jauntywundrkind 2y ago
16kB pages is fine with me, so long as drives have Flexible Data Placement, where they can specify a placement group for their writes. This will let the drive write associated data together & avoid write-amplification problems. Alas I have yet to see many drives advertise FDP clearly. In fact I've seen deception, such as Micron claiming to demo FDP with their 7450 Pro series, but they aren't testing official NVMe FDP, they just namespaced their drive into four equal sizes. One major advantage of FDP is it doesn't need ahead of time subdividing! https://www.micron.com/about/blog/storage/ssd/benefits-of-flexible-data-placement-on-real-workloads-using-aerospike https://www.micron.com/about/blog/storage/ssd/benefits-of-fl... This was supposed to be a pretty low lift for driver makers, and the hyperscalers were beating down the door for FDP. I'm shocked it's a year latter and it's still not clear how we can get drives that will work with us to combat write amplification.
cm2187 2y ago
I believe they always do that with the ION line which has crap endurance for random writes (relative to the average datacenter SSD), so I assume it's a way to reassure customers that the non random is not that bad.

dylan604 2y ago

Or it could be a boon for video/audio/graphics types. I’m used to setting up my spinning rust arrays with 64k block sizes. Continuous reads from large files vs lots of tiny files is something that tripped up a lot of storage vendors trying to break into the media world.

bufferoverflow 2y ago
It's very expensive. Gen5 SSDs are around $100/TB. Hard drives are $16/TB. It makes sense to have a small fast SSD for editing, but HDDs for mass-storage. And with video you usually want redundancy, storage gets expensive fast.
- dylan604 2y ago
  And with any data you usually want redundancy, FTFY But I know what you meant. I used to have systems built with multiple small arrays where the original sources were in RAID-5 (pre-raid6). The software would read from that array and write to a RAID0 array as these files were only ephemeral. The software would then read from that RAID0 and write to a second RAID0 to avoid reading/writing to the same volume. This was all before SSDs. For the lulz, I would love to see how that workflow would perform with NVMe storage. At the time of using these systems, spinning rust was not $16/TB and was probably closer to the $100/TB. I still remember $1k/GB in the early 90s
  - bufferoverflow 2y ago
    At the time you didn't have 8K or RAW videos or 200-megapixel photos. There's never enough storage.
    - dylan604 2y ago
      While not 8k, it was 4K-5k and definitely RAW. While never enough storage, we have achieved a point where disk I/O is not a bottle neck. It used to take a raid0 7 disk stripe to record SD video only while the audio was written to a separate drive. The fact we can record 8k RAW with audio on a single drive the size of a stick of RAM is crazy to compare. Smaller faster cheaper.

starspangled 2y ago

Physical NAND pages aren't really relevant. At least some high end SSDs I know of would gang multiple NAND chips together anyway and effectively the page size would be X*physical (where X might be 8 or 16). It could still do 512 byte writes well if it was configured that way. The limiting factor is the size of the forward map structures. To a lesser degree also things like garbage collection overheads and reverse map structures (particularly if you support deduplication or snapshots). But forward map is the main thing.

jauntywundrkind 2y ago

16kB pages is fine with me, so long as drives have Flexible Data Placement, where they can specify a placement group for their writes. This will let the drive write associated data together & avoid write-amplification problems. Alas I have yet to see many drives advertise FDP clearly. In fact I've seen deception, such as Micron claiming to demo FDP with their 7450 Pro series, but they aren't testing official NVMe FDP, they just namespaced their drive into four equal sizes. One major advantage of FDP is it doesn't need ahead of time subdividing! https://www.micron.com/about/blog/storage/ssd/benefits-of-flexible-data-placement-on-real-workloads-using-aerospike https://www.micron.com/about/blog/storage/ssd/benefits-of-fl... This was supposed to be a pretty low lift for driver makers, and the hyperscalers were beating down the door for FDP. I'm shocked it's a year latter and it's still not clear how we can get drives that will work with us to combat write amplification.

cm2187 2y ago

I believe they always do that with the ION line which has crap endurance for random writes (relative to the average datacenter SSD), so I assume it's a way to reassure customers that the non random is not that bad.