[ntp:questions] Re: clock resolutions for different OS

David L. Mills mills at udel.edu
Thu Aug 28 19:05:50 UTC 2003


Ulrich,

Glad you found the jitter.c program useful. Try the following:

Run multiple copies of that program in an otherwise underutilized CPU
and make a histogram in log-log scales of the frequency values. The
resulting graph gives an inicredible amount of interesting data. You can
find little bumps in the frequency characteristic due to cache misses,
TLB misses, context switches, timeslice-end, I/O latencies and, as I
found with the Alpha, unexplained rare latencies over several hundred
milliseconds.

The reason I built the program in the first place was to search for
system clock glitches like non-monotonic behavior with multiple CPUs and
the like. That proved very useful with multi-CPU machines and the
nanokernel distribution, since the timer oscillator is not syntonic with
the process cycle counters in the various CPUs.

Dave

Ulrich Windl wrote:
> 
> What do you want? Is using the "jitter.c program from the NTP sources
> not sufficient to get your machine's clock resolution? Resolution
> depends on system load, believe me. Even dirty-ing the CPU's cache may
> affect timing (ia32).
> 
> Regards,
> Ulrich



More information about the questions mailing list