Chris Vest

@chvest

Abstraction | Simplicity | Concurrency. Dragon keeper at Zoo.

Not enough memory for annotating '__entry_SYSCALL_64_trampoline' symbol
Vrijeme pridruživanja: svibanj 2009.

Tweetovi

Blokirali ste korisnika/cu @chvest

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

  1. proslijedio/la je Tweet
    prije 17 sati

    I wrote some thoughts on software performance:

    Poništi
  2. proslijedio/la je Tweet
    prije 9 sati

    The deadly Virus you don't hear so much about: Measles killed over 6000 in Congo.

    Poništi
  3. prije 5 sati

    Continuing the string searching story, I had the brain wave to try ripgrep. Searching for fixed strings in binary files, it tops out at 900 MiB/s. That's the target to meet. I should inspect the code to see how it's done.

    Prikaži ovu nit
    Poništi
  4. prije 5 sati

    I wanted to know how effective the clean-up efforts of the nuclear materials on the Thule base had been, but the report from the official study is written in Greenlandic... that's, uh, a fairly niche language.

    Poništi
  5. prije 7 sati

    Finally a way to put the PCI-e SSDs in the drive bay where they belong.

    Poništi
  6. proslijedio/la je Tweet
    prije 12 sati

    Enjoyed my first a lot! Memorable moment during the presendation by in the rust track: "Programming languages are just Domain-Specific Languages for system calls"

    Poništi
  7. 2. velj

    And Apostolico claims to place an O(2n) upper bound on BM. That's the same worst case as KMP.

    Prikaži ovu nit
    Poništi
  8. 2. velj

    Oh, I had missed that the Galil rule gives Boyer-Moore linear worst case performance. This might mean that BM is always a better choice than KMP.

    Prikaži ovu nit
    Poništi
  9. 2. velj

    Just 10% of the time goes into IO. 70% is in the tight KMP loop looking for matching bytes. Might just be doing too many instructions per byte to go any faster in a single thread.

    Prikaži ovu nit
    Poništi
  10. 2. velj

    Release build goes at ~300 MiB/s. A lot better but still less than 1/3rd of what I was hoping for.

    Prikaži ovu nit
    Poništi
  11. 2. velj

    The IO part is actually a tiny fraction of the work. The trick is making the bytes easily accessible, and the iterators couldn't do that. Even when accessing the buffer directly, most time goes into silly stuff like dereferencing, casting, error checking.

    Prikaži ovu nit
    Poništi
  12. 2. velj

    Turns out (all in debug mode) grabbing a file byte iterator can only let through 1.5 MiB/s. Using the iterator through a BufReader brings it up to 4 MiB/s. Doing manual buffer management let's me get roughly 16 MiB/s.

    Prikaži ovu nit
    Poništi
  13. proslijedio/la je Tweet
    1. velj

    It's not every day that a read-after-write makes a server spend 4% of its CPU cycles in one instruction. Wrote that one up here:

    Poništi
  14. proslijedio/la je Tweet

    President Trump was impeached for abuse of power and obstruction of Congress. By refusing to call witnesses and compel documents, Senate Republicans have chosen to become accomplices in his cover-up.

    Poništi
  15. proslijedio/la je Tweet
    29. sij

    Posted without comment

    Prikaži ovu nit
    Poništi
  16. proslijedio/la je Tweet
    30. sij

    The nature of Brexit was never positive or hopeful. Brexit was simply the repository for well of bitterness, resentment, entitlement & prejudice. Ugly feelings looking for a home. They found it in Brexit. Now they have it… the ugliness will go in search of another target.

    Prikaži ovu nit
    Poništi
  17. 30. sij

    If I can make it go at 1 GiB/s on my NVMe drive, then that’ll be a satisfactory solution.

    Prikaži ovu nit
    Poništi
  18. 30. sij

    Since binary files very often have looong stretches of the same bytes, or otherwise low-entropy structures, the worst-case performance is important. So I think I’ll go with KMP.

    Prikaži ovu nit
    Poništi
  19. 30. sij

    Working on adding search to XV. Looks like there are basically two options for how to do it: Boyer-Moore, and Knuth-Morris-Pratt. BM has better best and average case perf., but worse worst case compared to KMP, if I understand correctly.

    Prikaži ovu nit
    Poništi
  20. 30. sij

    One smol IDIV for page offset, one giant regression for databass kind.

    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:

    ·