How long does your nanosleep() really sleep? (2004)

How long does your nanosleep() really sleep? (2004)(dragonflybsd.org)

53 points by lkurusa 7 years ago | 16 comments

clarry 7 years ago |

This bug is still there in OpenBSD. With the 100Hz tick rate, sleep(10ms) turns effectively into 15 milliseconds, or -- more likely -- 20 milliseconds if the system is lightly loaded and the program doesn't spend much of its timeslice.

Of course, some programs don't expect this.

Fun story, me and my little sister played a game online years ago. The Mana World. A game with an OpenBSD port. She teamed up with some guy in game. She also told me that guy had a terrible internet connection & slow PC. So once I found them, I watched this guy lag and move about in slow motion. I asked if he's running OpenBSD. He said he's running OpenBSD. I sent him a little patch and it's like he got a new PC and a new internet connection, just like that.. :-)

ouid 7 years ago | |

do you have a link to this little patch so that I can refer to this post at some point in the future?

mikroskeem 7 years ago | |

can you put this patch up somewhere, e.g gist.github.com or something? or atleast point us where to look exactly? thank you :)

clarry 7 years ago | | |

Why, do you work on TMW?

Like I said, it was years ago. If I still have the patch, it's got to be on one of these spinning rust disks sitting in my closet. I'm on a trip so I really can't look for it now.

Assuming the client is still buggy (things like SDL2 have happened so who knows?), I'd probably look here first:

https://gitlab.com/manaplus/manaplus/blob/485e9bb12bf5dda318...

The logic won't run right if the timer is implemented with a sleep that ends up sleeping much longer than intended. If it works like it did back then (sleep is virtually always 20ms), simply incrementing the tick count by two every time should be enough to make it smooth on OpenBSD. Of course the correct fix is to keep track of actual time spent and adjust ticks accordingly. Or use a timer mechanism that has higher precision. IIRC setitimer[1] can get you a regular 100Hz alarm on OpenBSD.

http://man.openbsd.org/setitimer.2

If I'm not entirely mistaken, the old userspace implementation of pthreads worked with a 100Hz timer using this mechanism.

loeg 7 years ago |

On modern FreeBSD, when I run the article's demo program, by default I see even higher and more variable latency than the author observed in 2004:

However, when I set the kern.timecounter.alloweddeviation sysctl ("Allowed time interval deviation in percents") to 0 (from the default of 5%), I get very low and consistently low latency:

There are pros and cons to both defaults.

macdice 7 years ago | |

Can you elaborate on the cons?

loeg 7 years ago | | |

Sure thing.

Why make wakeup imprecise by default? The slop allows for some coalescing of nearby (in time) events into a single wakeup, saving power[0]. This article provides some figures that give a good intuition for why this might work: https://arstechnica.com/gadgets/2013/06/how-os-x-mavericks-w...

See also: tickless kernels in general[1]; FreeBSD switched to a tickless kernel with coalescing about 5.5 years ago[2].

[0]: https://en.wikipedia.org/wiki/Timer_coalescing

[1]: https://en.wikipedia.org/wiki/Tickless_kernel

[2]: https://svnweb.freebsd.org/base?view=revision&revision=24777...

waterhouse 7 years ago | | |

From what I've heard about other systems, if the system is given some wiggle room in when it runs tasks, then it can often run multiple tasks in a single wakeup, instead of waking up for one task, doing it, going to sleep for a very short time, waking up to do the second, etc. Wakeups, from what I've heard, carry overhead and increase power consumption, which may matter. (Also, if the CPUs are normally fully busy with user programs instead of fully asleep with zero tasks, replace "wakeup" with "context switch into the kernel".)

I don't know if this is what loeg had in mind, and I have not investigated the subject myself.

js2 7 years ago |

Solaris 8 briefly had a kernel bug such that usleep() didn't sleep at all. I don't recall the details, just that after patching the system and rebooting things were behaving very bizarrely. I eventually realised the sleep command was returning instantly and from that tracked it to the system call, and then found a patch for the issue. Google is failing me in digging up any details. It would've been in the 2000-2004 time frame.

cryptonector 7 years ago | |

EDIT: Oy, TFA is from 2004. Never mind my comment about Solaris 8 being ancient.