Tweetovi

Blokirali ste korisnika/cu @ZebaHrvoje

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @ZebaHrvoje

  1. proslijedio/la je Tweet
    17. pro 2019.

    A while back I wrote an internal note on why you can sometimes see discrepancies between application IO sizes and what is observed on the device side. Figured the greater public might want to read it as well, so here it is: "When 2MB turns into 512KB"

    Prikaži ovu nit
    Poništi
  2. proslijedio/la je Tweet
    4. pro 2019.
    Poništi
  3. proslijedio/la je Tweet
    5. stu 2019.

    Posted the initial version of support for IORING_OP_LINK_TIMEOUT commands. They allow tying a timeout to a given request, which will then cancel if not completed before the timeout. More details in the posting below: and tests:

    Prikaži ovu nit
    Poništi
  4. proslijedio/la je Tweet
    28. lis 2019.

    By popular demand, here's support for cancellation of any pending request that has been offloaded to async context. I do have a test case for liburing, I'll be posting that tomorrow. For now, the commit message below has all the details.

    Prikaži ovu nit
    Poništi
  5. proslijedio/la je Tweet
    23. lis 2019.

    Spend a few days starting the workqueue replacement for io_uring. Reasons for this work is stated in the commit message. Next stable, thorough testing, and then syslets/threadlets await (who doesn't love a good pun).

    Prikaži ovu nit
    Poništi
  6. 9. lis 2019.

    I've done some comparative tests against . ScyllaDB did a great job bounding latencies due to their advanced I/O scheduler (I guess). shows 5x improvement. For more information, take a look at

    Poništi
  7. 6. lis 2019.

    Last thing left to do was to run another 100M keys insert at 10k keys/s on top of 10.1B already inserted keys while running 16 readers to see how things scale. Checkout the results. Also, checkout the reader response tails after the writer finishes ;)

    Prikaži ovu nit
    Poništi
  8. 6. lis 2019.

    After the last test, I tried inserting 100M keys on top of existing 10B at the rate of 10k keys/s. This time, both the writer and the readers (two of them) were extreamely stable with readers averaging under 0.8ms responses.

    Prikaži ovu nit
    Poništi
  9. 6. lis 2019.

    Just to keep things in perspective, inserting 10B keys in took ~14h. Other databases would take dozens of days just to complete this test ;) At an insert rate of 200k keys/s, things look a bit erratic. Normal productions should see lower per cpu rates.

    Prikaži ovu nit
    Poništi
  10. proslijedio/la je Tweet
    28. ruj 2019.

    Killing time on a long flight, doubled the performance of linked SQEs for some conditions. 4GB copy with examples/link-cp goes from: real0m5.272s user0m0.160s sys 0m5.039s to real0m2.374s user0m0.124s sys 0m1.985s

    Poništi
  11. proslijedio/la je Tweet
    26. ruj 2019.
    Odgovor korisniku/ci

    That’s exactly right. But you forgot efficient. And scalable.

    Poništi
  12. 21. ruj 2019.

    I did some digging around mongodb and found a way to use LSM trees as an underlying data structure. Got better results. With it, mongodb ran for ~38h. Also did some work on to level out update response spikes.

    Poništi
  13. proslijedio/la je Tweet
    20. ruj 2019.

    Aristotle's Wheel Paradox When you roll a rigid wheel over a distance equal to its circumference you seem to find that the circumferences of the inner (blue) and outer (red) circles are equal. 🤔 Can you guess the fallacy?

    Poništi
  14. proslijedio/la je Tweet
    19. ruj 2019.

    Sneak peek from my slides - basic peak perf testing with aio, io_uring, and polled io_uring. Single core. Get wrecked, aio.

    Poništi
  15. 18. ruj 2019.

    Another view of the data from the test we ran last week

    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·