calling functions at the same time

[Grant Edwards]

In article <r4idnSA-J8HtSw7dRVn-hQ at powergate.ca>, Peter Hansen wrote:

>>>You can't get greater then 1ms accuracy when measuring a time
>>>on todays machines, and that will be less when measuring time
>>>over a network...
>> 
>> Not sure what you mean by "todays machines", but on a
>> Pentium-class machine running Linux, you get approx 1us
>> resolution with calls to time.time():
> 
> Keep in mind the difference between "accuracy" and
> "resolution" (or "precision").

I do.

Not only do you get 1us resolution, you get packet timestamp
_accuracy_ to well under 1ms according to tests I've run 
comparing packet timestamps against an oscilloscope trace.

Delta-T accuracy of calls to time() are accurate to 1us as
well.  Absolute accuracy depends on how you set your system
clock.  Running NTP with in-kernel FLL control of the system
tick will generally allow you to maintain absolute accuracies
of under 100us.

> I think Knio is probably right that, unless you are running on
> a realtime OS, you won't be able to guarantee anything better
> than 1ms accuracy, and quite probably not even that.

Under Linux, I can guarantee you the answer you get back from
time.time() is accurate to 1us.  However, it is a multitasking
system, and there are other things running.  While it's easy to
_determine_ exactly when you called time.time(), it's difficult
to _control_ the point in time when you call time.time(). When
your task does get a chance to run, and you do get to call
time.time(), you'll generally get a result accurate to 1us.

If I do a busy-wait loop with nothing else running, I've been
able to accurately measure pulsewidths of a few microseconds
under Linux using the gettimeofday() call (which I believe is
what time.time() calls.

As soon as there are other ready tasks, the accuracy of the
measurement quickly deteriorates to tens of millisconds due to
scheduling latencies.

-- 
Grant Edwards                   grante             Yow!  It's today's SPECIAL!
                                  at               
                               visi.com