NetBSD configures a timer device to deliver a periodic timer interrupt, usually every 10 ms, in order to count time, wake threads that are sleeping, etc. This made sense when timer devices were expensive to program and CPUs ran at MHz. But today, CPUs run at GHz; timers on modern x86, arm, mips, etc. hardware are cheap to reprogram; programs expect greater than 10 ms resolution for sleep times; and mandatory periodic activity on idle machines wastes power.

There are four main milestones to this project:

  1. Choose a data structure for high-resolution timers, and a way to request high-resolution vs low-resolution sleeps, and adapt the various timeout functions (cv_timedwait, etc.) to use it. The current call wheel data structure for callouts provides good performance, but only for low-resolution sleeps. We need another data structure that provides good performance for high-resolution sleeps without hurting the performance of the existing call wheel for existing applications.

  2. Design a machine-independent high-resolution timer device API, implement it on a couple machines, and develop tests to confirm that it works. This might be done by adapting the struct timecounter interface to arm it for an interrupt, or might be done another way.

  3. Convert all the functions of the periodic 10 ms timer, hardclock, to schedule activity only when needed.

  4. Convert the various software subsystems that rely on periodic timer interrupts every tick, or every second, via callout(9), either to avoid periodic work altogether, or to batch it up only when the machine is about to go idle, in order to reduce the number of wakeups and thereby reduce power consumption.