5 ms·

Could someone say something about the new "BFQ I/O scheduler"? I was not aware of the work being done, and the goal to have a more reponsive desktop

zackelan 9y ago
http://algo.ing.unimo.it/people/paolo/disk_sched/ http://algo.ing.unimo.it/people/paolo/disk_sched/ https://lwn.net/Articles/601799/ https://lwn.net/Articles/601799/ tl;dr: > CFQ separates each process's I/O requests into a separate queue, then rotates through the queues trying to divide the available bandwidth as fairly as it can. > The BFQ I/O scheduler also maintains per-process queues of I/O requests, but it does away with the round-robin approach used by CFQ. Instead, it assigns an "I/O budget" to each process.
- ausjke 9y ago
  CFQ is best for mechanical hard drives not sure if it is updated for SSDs, deadline and noop seem better for SSD
  - the8472 9y ago
    There's another new IO scheduler for SSDs: Kyber, https://patchwork.kernel.org/patch/9672023/ https://patchwork.kernel.org/patch/9672023/
    - gigatexal 9y ago
      anyone do benches say using fio with this vs. the default scheduler? nvm: http://www.phoronix.com/scan.php?page=article&item=linux-412-io&num=1 http://www.phoronix.com/scan.php?page=article&item=linux-412...
      - the8472 9y ago
        Aiui kyber is designed for providing consistently low latency, e.g. for desktop users. And maybe specific user-facing server workloads? Those benchmarks mostly seem to test throughput, so are not good at showing the intended effect. You would want to see something that shows 99th percentile latency of individual reads/writes.
        gigatexal 9y ago
        yeah i came to that same conclusion. I may need to do my own testing using Fio.
  - kaixi 9y ago
    Why is this being downvoted? AFAIK, this is mostly true.
    - saghm 9y ago
      Random OT question: how can you tell when a comment that isn't yours is being downvoted? I've seen comments like yours a few times before, but I can't see the karma on individual posts. Is there a karma threshold for being able to see this (like the 500 karma threshold for being able to downvote)?
      - jsjohnst 9y ago
        You can tell because the font color starts to fade out till you can barely read the text.
        saghm 9y ago
        Interesting; I'm unable to notice this on the comment being discussed with my current setup (Firefox on Linux)
        icebraining 9y ago
        That's because it was upvoted back to black since then :)
        saghm 9y ago
        Haha fair enough
    - _arvin 9y ago
      wasn't the downvoter, but general statements like that don't include context of workload (i.e. is it a desktop or server, gaming pc, etc.) so different i/o schedulers would be optimal for different scenarios

zackelan 9y ago

http://algo.ing.unimo.it/people/paolo/disk_sched/ http://algo.ing.unimo.it/people/paolo/disk_sched/ https://lwn.net/Articles/601799/ https://lwn.net/Articles/601799/ tl;dr: > CFQ separates each process's I/O requests into a separate queue, then rotates through the queues trying to divide the available bandwidth as fairly as it can. > The BFQ I/O scheduler also maintains per-process queues of I/O requests, but it does away with the round-robin approach used by CFQ. Instead, it assigns an "I/O budget" to each process.

ausjke 9y ago
CFQ is best for mechanical hard drives not sure if it is updated for SSDs, deadline and noop seem better for SSD
- the8472 9y ago
  There's another new IO scheduler for SSDs: Kyber, https://patchwork.kernel.org/patch/9672023/ https://patchwork.kernel.org/patch/9672023/
  - gigatexal 9y ago
    anyone do benches say using fio with this vs. the default scheduler? nvm: http://www.phoronix.com/scan.php?page=article&item=linux-412-io&num=1 http://www.phoronix.com/scan.php?page=article&item=linux-412...
    - the8472 9y ago
      Aiui kyber is designed for providing consistently low latency, e.g. for desktop users. And maybe specific user-facing server workloads? Those benchmarks mostly seem to test throughput, so are not good at showing the intended effect. You would want to see something that shows 99th percentile latency of individual reads/writes.
      - gigatexal 9y ago
        yeah i came to that same conclusion. I may need to do my own testing using Fio.
- kaixi 9y ago
  Why is this being downvoted? AFAIK, this is mostly true.
  - saghm 9y ago
    Random OT question: how can you tell when a comment that isn't yours is being downvoted? I've seen comments like yours a few times before, but I can't see the karma on individual posts. Is there a karma threshold for being able to see this (like the 500 karma threshold for being able to downvote)?
    - jsjohnst 9y ago
      You can tell because the font color starts to fade out till you can barely read the text.
      - saghm 9y ago
        Interesting; I'm unable to notice this on the comment being discussed with my current setup (Firefox on Linux)
        icebraining 9y ago
        That's because it was upvoted back to black since then :)
        saghm 9y ago
        Haha fair enough
  - _arvin 9y ago
    wasn't the downvoter, but general statements like that don't include context of workload (i.e. is it a desktop or server, gaming pc, etc.) so different i/o schedulers would be optimal for different scenarios