4 ms·

The amount of space used by he ZIL on a SLOG is very modest. Typically less than a gigabyte. So a SSD used as a SLOG will be practically empty it's entire life.

The amount of space used by he ZIL on a SLOG is very modest. Typically less than a gigabyte. So a SSD used as a SLOG will be practically empty it's entire life. The reason one would use a large capacity SSD for this purpose is to take advantage of the large SSD's higher iops and greater wear leveling.

BrainInAJar 9y ago
it's unfortunate but not surprising that you can't buy very small, very fast SSD's for this purpose
- wmf 9y ago
  PMC-Sierra/Microsemi has some cool DRAM-based SSDs that do 1M IOPS (not a typo) with 16GB capacity. https://www.microsemi.com/product-directory/storage-boards/3690-flashtec-nvram-drives https://www.microsemi.com/product-directory/storage-boards/3... I don't think they're available in U.2 or M.2 format though. NVDIMM-N would probably also be great for slog. https://www.anandtech.com/show/12029/micron-announces-32gb-ddr4-nvdimmn-modules https://www.anandtech.com/show/12029/micron-announces-32gb-d...
  - carlob 9y ago
    Says over 10M IOPS in the link from microsemi, so… maybe a typo?
    - wmf 9y ago
      1M in NVMe mode, 10M in nonstandard mode. I don't know if there's any software that can use the nonstandard mode.
  - wtallis 9y ago
    Yeah, DRAM chips are the only ones that currently offer the right balance of performance and capacity per die. NAND and 3D XPoint dies are too big and too slow per die for this use case, so they're better used as nonvolatile storage to flush the log to in the event of power failure, not as the primary log storage media. I expect it will soon become common for high-end enterprise NVMe SSDs to include a few GB of DRAM accessible through the new Persistent Memory Region feature, so that they can be accessed through simple PCIe memory read and write operations, but will automatically be preserved by the SSD in the event of a power failure. It won't be as fast as a NVDIMM, but works with existing form factors and platforms.
    - StudentStuff 9y ago
      Battery backed DRAM is likely a better choice for performance & endurance in most applications, as the Optane product line is a lame duck with memory that dies around the 1TB of writes mark, leaving you with lost data and a dead SSD.
      - wtallis 9y ago
        > as the Optane product line is a lame duck with memory that dies around the 1TB of writes mark, leaving you with lost data and a dead SSD. You seem to be much more concerned with venting anger than conveying accurate or relevant information.
        Dylan16807 9y ago
        Considering that the Intel drive in https://techreport.com/review/26523/the-ssd-endurance-experiment-casualties-on-the-way-to-a-petabyte https://techreport.com/review/26523/the-ssd-endurance-experi... self-destructed when it ran out of guaranteed lifetime, I think it's a fair criticism.
        wtallis 9y ago
        There's nothing fair about making up a number that's wrong by multiple orders of magnitude. I'm not a fan of the end of life behavior of Intel SSDs either, but it's a consistent policy that does work well for enterprise usage scenarios, with no unexpected data loss. They probably shouldn't be applying the same policy to the consumer products. But the fact that there are valid criticisms to be made doesn't mean this particular one was fair or germane.
        Dylan16807 9y ago
        Oh, I wasn't even looking at the value of the number. Yeah, it's bad to exaggerate like that without being clear about it. But something performance-heavy can burn out an endurance of 10 drive writes per day pretty fast.
- benjaminl 9y ago
  Each NAND chip on an SSD is actually relatively slow. To get high read and write speeds, the access is multiplexed across all of the chips in the drive. Small capacity drives have a smaller number of NAND chips. This reduces the performance of the drives. In addition each NAND cell has a very limited number of lifetime writes it supports, typically 1k to 4k erase/write cycles. So drives do wear leveling, spreading writes over the whole drive. Therefore drives uses in extremely right heavy enterprise workloads may need to be very large just to support the required write endurance. From what I understand Optane is fast enough to not need this extra controller processing. Resulting in the incrediblly latency performance demonstrated by Optabe drives. I suspect if Intel wanted to sacrifice latency, they could dramatically increase top line read and write performance by performing the same read and write interleaving that NAND SSDs do.
  - wtallis 9y ago
    Intel's Optane SSDs do stripe across the controller's seven channels; that's the main reason why they're faster than the single-channel Optane Memory M.2 modules. At any capacity or channel configuration, 3D XPoint devices benefit from not having to perform large block erase operations. They can perform reads and writes at the same granularity, allowing for in-place modification of data instead of requiring a log structure with garbage collection under the hood.

BrainInAJar 9y ago

it's unfortunate but not surprising that you can't buy very small, very fast SSD's for this purpose

wmf 9y ago
PMC-Sierra/Microsemi has some cool DRAM-based SSDs that do 1M IOPS (not a typo) with 16GB capacity. https://www.microsemi.com/product-directory/storage-boards/3690-flashtec-nvram-drives https://www.microsemi.com/product-directory/storage-boards/3... I don't think they're available in U.2 or M.2 format though. NVDIMM-N would probably also be great for slog. https://www.anandtech.com/show/12029/micron-announces-32gb-ddr4-nvdimmn-modules https://www.anandtech.com/show/12029/micron-announces-32gb-d...
- carlob 9y ago
  Says over 10M IOPS in the link from microsemi, so… maybe a typo?
  - wmf 9y ago
    1M in NVMe mode, 10M in nonstandard mode. I don't know if there's any software that can use the nonstandard mode.
- wtallis 9y ago
  Yeah, DRAM chips are the only ones that currently offer the right balance of performance and capacity per die. NAND and 3D XPoint dies are too big and too slow per die for this use case, so they're better used as nonvolatile storage to flush the log to in the event of power failure, not as the primary log storage media. I expect it will soon become common for high-end enterprise NVMe SSDs to include a few GB of DRAM accessible through the new Persistent Memory Region feature, so that they can be accessed through simple PCIe memory read and write operations, but will automatically be preserved by the SSD in the event of a power failure. It won't be as fast as a NVDIMM, but works with existing form factors and platforms.
  - StudentStuff 9y ago
    Battery backed DRAM is likely a better choice for performance & endurance in most applications, as the Optane product line is a lame duck with memory that dies around the 1TB of writes mark, leaving you with lost data and a dead SSD.
    - wtallis 9y ago
      > as the Optane product line is a lame duck with memory that dies around the 1TB of writes mark, leaving you with lost data and a dead SSD. You seem to be much more concerned with venting anger than conveying accurate or relevant information.
      - Dylan16807 9y ago
        Considering that the Intel drive in https://techreport.com/review/26523/the-ssd-endurance-experiment-casualties-on-the-way-to-a-petabyte https://techreport.com/review/26523/the-ssd-endurance-experi... self-destructed when it ran out of guaranteed lifetime, I think it's a fair criticism.
        wtallis 9y ago
        There's nothing fair about making up a number that's wrong by multiple orders of magnitude. I'm not a fan of the end of life behavior of Intel SSDs either, but it's a consistent policy that does work well for enterprise usage scenarios, with no unexpected data loss. They probably shouldn't be applying the same policy to the consumer products. But the fact that there are valid criticisms to be made doesn't mean this particular one was fair or germane.
        Dylan16807 9y ago
        Oh, I wasn't even looking at the value of the number. Yeah, it's bad to exaggerate like that without being clear about it. But something performance-heavy can burn out an endurance of 10 drive writes per day pretty fast.
benjaminl 9y ago
Each NAND chip on an SSD is actually relatively slow. To get high read and write speeds, the access is multiplexed across all of the chips in the drive. Small capacity drives have a smaller number of NAND chips. This reduces the performance of the drives. In addition each NAND cell has a very limited number of lifetime writes it supports, typically 1k to 4k erase/write cycles. So drives do wear leveling, spreading writes over the whole drive. Therefore drives uses in extremely right heavy enterprise workloads may need to be very large just to support the required write endurance. From what I understand Optane is fast enough to not need this extra controller processing. Resulting in the incrediblly latency performance demonstrated by Optabe drives. I suspect if Intel wanted to sacrifice latency, they could dramatically increase top line read and write performance by performing the same read and write interleaving that NAND SSDs do.
- wtallis 9y ago
  Intel's Optane SSDs do stripe across the controller's seven channels; that's the main reason why they're faster than the single-channel Optane Memory M.2 modules. At any capacity or channel configuration, 3D XPoint devices benefit from not having to perform large block erase operations. They can perform reads and writes at the same granularity, allowing for in-place modification of data instead of requiring a log structure with garbage collection under the hood.