Messages in this thread | | | Date | Thu, 23 Aug 2007 22:05:26 +0200 | From | Ingo Molnar <> | Subject | Re: [patch 0/2] s390 related scheduler patches and questions |
| |
* Christian Borntraeger <borntraeger@de.ibm.com> wrote:
> Another question: > > nanosecond resolution seems not ideal for 64bit values, at least if an > architecture has to do calculations. For example our cpu timer is > signed 64bit and bit 51 (63=LSB) steps by one each microsecond. To > create a nanosecond based timer we need: nsecs= clock*125/512 or nsecs > = clock/512*125. The first variant overflows in a time frame that is > still reasonable to be seen in practice (about 2 years if I made no > errors), the second variant introduces a stepping rate of 125ns. Of > course we could use nsec = (((((((clock/8)*5)/4)*5)/4)*5)/4), to have > a long overflow period and a 1.25ns stepping rate but this looks quite > ugly. Are you going to stick with nanosecond resolution? If yes, do > you think a stepping rate of 125ns is ok? Any chance to change the > scheduler resolution to microseconds? ;-)
there are noticeable accounting artifacts on fast boxes that do sub-microsecond scheduling, so getting the best sched_clock() resolution is certainly handy. (Also, nanoseconds gives us some rounding-error room.) But 0.125 usecs should still be fine.
the 2 years overflow is not an issue: you could solve that by only using the first 55 bits of the clock. This means the clock would wrap around every 1.14 years - the effects of that are that the "dont let time go backwards" code in the scheduler will ignore a very small interval (which happens at the wraparound) and will continue with the wrapped-around clock from that point on. The rq->clock itself is a true, monotonic 64-bit clock that wraps every 584.9 years.
[ and even after 584.9 years it should have no serious failure mode, as the timestamps are used in a relative manner. The only, minimal effect is on tasks that sleep for more than 584 years - which could get a few millisecs less sleeper fairness share. I am not overly worried about getting bugreports about that in my lifetime though =B-) (unless someone gets serious about bio-cryogenics R&D, real soon.) ]
Ingo - 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/
| |