Messages in this thread | | | Date | Fri, 16 Jun 2006 17:33:39 +0200 (CEST) | From | Roman Zippel <> | Subject | Re: clocksource |
| |
Hi,
On Thu, 15 Jun 2006, john stultz wrote:
> I've also been working on improving the adjustment algorithm. Paul > Mckenney enlightened me to the established concepts in control theory, I > started reading up on PID control (see: > http://en.wikipedia.org/wiki/PID_controller ). While I have understood > the basic concept, it was useful to read up on it. I've tried to rework > the adjustment code accordingly. > > The method I came up with is really just P-D (proportional-derivative) > control, but that should be ok since the adjustments are all linear so I > don't think the integral control is necessary (control theorists can > pipe in here).
This makes it more complex than necessary. AFAICT this controller calculates the adjustment solely based on the current error, but we have more information than this, which make the current error rather uninteresting. We know the clock frequency and the NTP frequency so we can easily precalculate, how the error will look like at the next few ticks. Based on this we can calculate how we have to adjust the clock frequency to reduce the error. Overshooting is also not a real problem as long as the absolute error gets smaller.
An important point about the last patch is not just robustness but also speed, it tries to keep the fast path small, which is basically:
interval = clock->cycle_interval; if (error > interval / 2) { adj = 1; if (unlikely(error > interval * 2)) { ... } } else if (error < -interval / 2) { adj = -1 interval = -interval; offset = -offset; if (unlikely(error < interval * 2)) { ... } } else return;
clock->mult += adj; clock->xtime_interval += interval; clock->xtime_nsec -= offset; clock->error -= interval - offset;
You'll need a very good reason to do anything more than this for small errors and I would suggest you start from something like this, as this is the very core of the error adjustment.
bye, Roman - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |