Re: [RFC][PATCH] Adjust SHIFT_PLL to improve NTP convergence.

["George Spelvin" <linux@horizon.com>]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1125 bytes --]

> IMHO, the "value" of calibration during boot is dubious (unless you just
> reboot a "hot" machine): Usually the machine is cold on reboot, and the
> oscillators will drift a lot initially. Depending on your environment,
> the temperatures inside your box may rise about 10°C or more within a
> few minutes.

> Maybe an experimental feature "recalibrate anytime (and as long as you'd
> like)" would be interesting...

Well, a couple of points come to mind:
- Rebooting hot is not all that unusual.  Less so that with Windows, but
  some of us install test kernels from time to time. :-)
- If you're calibrating the CPU against the PIT, they're often derived
  via PLL from the same clock source, so the *ratio* is fixed, and thus
  the calibration won't drift at all.
- If there were some mechanism to make the divisor consistent across
  boots, that would be helpful.  Imagine a file in /sys that I could
  write with an absolute frequency at boot time as long as it was within
  1000 ppm (0.1%) of the kernel's measurement.

Thinking about it, the third option seems the most useful.  It's
basically a cleaned-up tickadj.