2065840191340707905

2026-08-25 11:37:49 -04:00
1 parent 002b0e1fe0
commit 34075a79d5
@@ -0,0 +1,39 @@
+---
+title: "@NOTimothyLottes most futex impls i've seen check for that and early return before the syscall, but you likely already knew that"
+type: archive
+source: twitter
+source_url: "https://x.com/BonbliStar/status/2065944487075975648"
+author: "Bonbli ★"
+handle: BonbliStar
+post_id: "2065944487075975648"
+date: 2026-06-13
+archived: 2026-08-25
+draft: false
+tags:
+  - archive
+  - twitter
+  - NOTimothyLottes
+description: "@NOTimothyLottes most futex impls i've seen check for that and early return before the syscall, but you likely already knew that"
+in_reply_to: ""
+parent_post_id: "2065840191340707905"
+---
+
+## Source
+
+- URL: https://x.com/BonbliStar/status/2065944487075975648
+- Author: Bonbli ★ (@BonbliStar)
+- Posted: 2026-06-13 23:48:43
+
+## Branch
+
+**1/**
+
+@NOTimothyLottes most futex impls i've seen check for that and early return before the syscall, but you likely already knew that
+
+**2/**
+
+@BonbliStar Yes, the comment was more the point of designing for the worst case, the limit of possibly blocking calls that interact with the scheduling.
+
+## Related
+
+- Spine: [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops]]
@@ -7,14 +7,15 @@ author: "NOTimothyLottes"
 handle: NOTimothyLottes
 post_id: "2065872197147578751"
 date: 2026-06-13
-archived: 2026-08-23
+archived: 2026-08-25
 draft: false
 tags:
  - archive
  - twitter
  - NOTimothyLottes
 description: "@axelgneiting The glibc implementation would do that."
-in_reply_to: "2065851939917906296"
+in_reply_to: ""
+parent_post_id: "2065851939917906296"
 ---

 ## Source
@@ -23,10 +24,14 @@ in_reply_to: "2065851939917906296"
 - Author: NOTimothyLottes (@NOTimothyLottes)
 - Posted: 2026-06-13 19:01:28

-## Thread
+## Branch

 **1/**

@axelgneiting The glibc implementation would do that. I bypass libc and just do it manually with perf_event_open syscall and mmap.

-![](https://pbs.twimg.com/media/HKtzz5RXcAA6CIX?format=jpg&name=orig)
+![](https://pbs.twimg.com/media/HKtzz5RXcAA6CIX?format=jpg&name=orig)
+
+## Related
+
+- Spine: [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops]]
@@ -0,0 +1,41 @@
+---
+title: "@NOTimothyLottes nanosleep(0) will sleep for the timer slack"
+type: archive
+source: twitter
+source_url: "https://x.com/raggi/status/2065865513608307103"
+author: "James Tucker"
+handle: raggi
+post_id: "2065865513608307103"
+date: 2026-06-13
+archived: 2026-08-25
+draft: false
+tags:
+  - archive
+  - twitter
+  - NOTimothyLottes
+description: "@NOTimothyLottes nanosleep(0) will sleep for the timer slack"
+in_reply_to: ""
+parent_post_id: "2065840191340707905"
+---
+
+## Source
+
+- URL: https://x.com/raggi/status/2065865513608307103
+- Author: James Tucker (@raggi)
+- Posted: 2026-06-13 18:34:55
+
+## Branch
+
+**1/**
+
+@NOTimothyLottes nanosleep(0) will sleep for the timer slack
+
+_mm_pause() if you want to avoid a context switch and keep your scheduler quantum
+
+**2/**
+
+@raggi I do that via inline asm to save power. This was more a comment on the cost of actually using the futex system call, need to be careful about the call frequency
+
+## Related
+
+- Spine: [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops]]
@@ -0,0 +1,47 @@
+---
+title: "@NOTimothyLottes Seems like people trying to push high frame rates would be perfectly fine with just busy spinning at that point"
+type: archive
+source: twitter
+source_url: "https://x.com/rflaherty71/status/2065859701909459406"
+author: "Ryan Flaherty"
+handle: rflaherty71
+post_id: "2065859701909459406"
+date: 2026-06-13
+archived: 2026-08-25
+draft: false
+tags:
+  - archive
+  - twitter
+  - NOTimothyLottes
+description: "@NOTimothyLottes Seems like people trying to push high frame rates would be perfectly fine with just busy spinning at that point"
+in_reply_to: ""
+parent_post_id: "2065840191340707905"
+---
+
+## Source
+
+- URL: https://x.com/rflaherty71/status/2065859701909459406
+- Author: Ryan Flaherty (@rflaherty71)
+- Posted: 2026-06-13 18:11:49
+
+## Branch
+
+**1/**
+
+@NOTimothyLottes Seems like people trying to push high frame rates would be perfectly fine with just busy spinning at that point
+
+**2/**
+
+@rflaherty71 Not exactly true, the larger problem is getting context switched at a bad time, so getting non-blocking bound comes with another set of dangers (higher bad preemption risk)
+
+**3/**
+
+@NOTimothyLottes @rflaherty71 All the more reason to go GPU-side where that can’t occur.
+
+**4/**
+
+@AgileJebrim @rflaherty71 Consumer GPU side is littered with junk these days, starting with multi display compositors.
+
+## Related
+
+- Spine: [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops]]
@@ -0,0 +1,64 @@
+---
+title: "Why not do a gradual sleeping? Start with"
+type: archive
+source: twitter
+source_url: "https://x.com/winning_tactic/status/2065896862129733922"
+author: "OZ"
+handle: winning_tactic
+post_id: "2065896862129733922"
+date: 2026-06-13
+archived: 2026-08-25
+draft: false
+tags:
+  - archive
+  - twitter
+  - NOTimothyLottes
+description: "Why not do a gradual sleeping? Start with"
+in_reply_to: ""
+parent_post_id: "2065840191340707905"
+---
+
+## Source
+
+- URL: https://x.com/winning_tactic/status/2065896862129733922
+- Author: OZ (@winning_tactic)
+- Posted: 2026-06-13 20:39:29
+
+## Branch
+
+**1/**
+
+Why not do a gradual sleeping? Start with
+while (spin--) {
+    asm volatile("pause" ::: "memory");
+}
+then yield()
+then yield 'til timer fire
+gradually approaching longer spans?
+You can have power modes that sets a lower threshhold to always yield and not spin e.g. or to have the particular levels based on eco/perf, and also it depends on the rest of the system to ensure spinning isn't a waste of cycles.
+
+But I've never considered a write to a dummy file as a legitimate yield/wait trigger. I wonder how that would fit in to my solution...
+
+**2/**
+
+@winning_tactic I don’t design around waiting for a specific point in time, I do other things, but talk about them when I can easily show the results
+
+**3/**
+
+Correct. But that's the point of gradually backoff, active spinlocking, then green thread yielding (cooperative other tasks, maybe even the one contesting) then kernel_poke (I guess that's what I would call that writing to file its a return from kernel-land immediately) then timer fired, the idea isn't to wait on a time, it's wait on an increasing value that changes tactics from the previous waits so that you don't really have to keep track of when what happens.
+https://gitlab.com/qsrc.net/Lampyr/-/blob/dev/rt/time.c?ref_type=heads#L78
+But I've got to work on the power stuff and steal your idea of course.
+
+**4/**
+
+@winning_tactic Ahh the illusion of nonblocking execution on consumer devices. The moment the os decides to preempt your thread, there no longer exists a non blocking interface. A single thread never gets a guarantee of execution, so waiting for some exact point in time is failed design
+
+**5/**
+
+@NOTimothyLottes It's green threads multiplexed on pinned threads both cooperatively and preemptively. I beg your pardon!~
+https://gitlab.com/qsrc.net/Lampyr/-/blob/dev/rt/task.c?ref_type=heads
+But, that particular version is a WIP, multiple impl and messy.
+
+## Related
+
+- Spine: [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops]]
@@ -0,0 +1,36 @@
+---
+title: "Interesting how on this Linux box, scheduling NOPs like nanosleep of 0 and futex waiting on an already set memory location … both take around 60 microseconds, cannot afford to do many of those per millisecond."
+type: archive
+source: twitter
+source_url: "https://x.com/NOTimothyLottes/status/2065840191340707905"
+author: "NOTimothyLottes"
+handle: NOTimothyLottes
+post_id: "2065840191340707905"
+date: 2026-06-13
+archived: 2026-08-25
+draft: false
+tags:
+  - archive
+  - twitter
+  - NOTimothyLottes
+description: "Interesting how on this Linux box, scheduling NOPs like nanosleep of 0 and futex waiting on an already set memory location … both take around 60 microseconds, cannot afford to do many of those per millisecond."
+in_reply_to: ""
+---
+
+## Source
+
+- URL: https://x.com/NOTimothyLottes/status/2065840191340707905
+- Author: NOTimothyLottes (@NOTimothyLottes)
+- Posted: 2026-06-13 16:54:17
+
+## Thread
+
+**1/**
+
+Interesting how on this Linux box, scheduling NOPs like nanosleep of 0 and futex waiting on an already set memory location … both take around 60 microseconds, cannot afford to do many of those per millisecond. Food for thought for high fps targets
+
+Branches: [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops/2026-06-13-rflaherty71-seems-like-people-trying-to-push-high-frame-rates]], [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops/2026-06-13-raggi-nanosleep-0-will-sleep-for-the-timer-slack-mm]], [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops/2026-06-13-winning_tactic-why-not-do-a-gradual-sleeping-start-with-while]], [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops/2026-06-13-BonbliStar-most-futex-impls-ive-seen-check-for-that-and]]
+
+**2/**
+
+A write of zero bytes is closer to say 6 microseconds in contrast, so just the user/kernel transition is expensive but 10x faster
@@ -69,8 +69,8 @@ tags:
 - [[archive/threads/NOTimothyLottes/2026-06-03-back-running-640x480-on-my-vga-crts-to-prep-for]]
 - [[archive/threads/NOTimothyLottes/2026-06-10-so-im-on-linux-now-but-obviously-i-cheated]]
 - [[archive/threads/NOTimothyLottes/2026-06-13-cross-compilation-is-up]]
+- [[archive/threads/NOTimothyLottes/2026-06-13-interesting-how-on-this-linux-box-scheduling-nops]]
 - [[archive/threads/NOTimothyLottes/2026-06-13-nanorc-done-and-back-to-actual-programming-with]]
- [[archive/threads/NOTimothyLottes/2026-06-13-the-glibc-implementation-would-do-that]]
 - [[archive/threads/NOTimothyLottes/2026-06-14-on-linux-i-do-speculative-pinning-of-cpu-memory]]
 - [[archive/threads/NOTimothyLottes/2026-06-26-quick-summary-of-why-not-pselect-ppoll-epoll]]
 - [[archive/threads/NOTimothyLottes/2026-06-27-deterrent-to-alsa-is-just-parsing-through-snd-pcm]]