[Tinyos Core WG] TEP102 and binary vs decimal units

[David Gay]

On 5/22/07, Philip Levis <pal at cs.stanford.edu> wrote:
> On May 22, 2007, at 10:03 AM, Cory Sharp wrote:
>
> > On 5/21/07, Philip Levis <pal at cs.stanford.edu> wrote:
> > On Mon, 2007-05-21 at 16:05, David Gay wrote:
> > > I'm finally getting to some final revisions for TEP 102, and the
> > only
> > > real open issue is the "binary vs decimal" milli and micro second
> > one.
> > > There's been a bunch of discussion on this, but I'm not clear we
> > > really reached consensus. Two options:
> > >
> > > - Leave the current TEP 102 more or less as is, i.e., our units are
> > > defined as 1/1024s, 1/32768s and 1/1048576s (TMilli, T32khz, TMicro)
> > >
> > > - Switch to the correspondiong decimal units (1/1000s, 1/32000s,
> > 1/1000000s)
> > >
> > > In practice, no implementations need change, as there's verbiage
> > about
> > > accuracy being loose.
> > >
> > > FWIW, on mica platforms the actual units (w/o drift or temperature
> > > effects) are 1/1024s, 1/28800s and 7.37/8000000s. On telos, I
> > believe
> > > it's 1/1024s, 1/32768s and 1/1000000s.
> > >
> > > If nobody answers this week, I'll switch TEP102 to the decimal
> > units.
> >
> > The consensus was to stick with the binary units, with the
> > understanding
> > that additional types could always be added for decimal units when/if
> > needed.
> >
> > Phil
> >
> > I thought we switched to decimal units with the understanding that
> > there could be a reasonable inaccuracy in the unit, since there
> > will be in practice, anyway.  Essentially: if our error bars are
> > huge, we may as well pick a standard unit instead of making one up.
>
> Is a 2% inaccuracy common? If not, then we should convert the calls
> to a best effort attempt at accuracy. I think it's disingenuous to
> say "they're decimal" but then give such a weak claim on accuracy
> that we make them binary. Doing this requires more bookkeeping of
> partial ticks, though, due to a need to keep timers in phase (e.g.,
> the simple (1000/1024) breaks if you ask for a periodic of 111ms and
> a periodic of 999ms).

I don't think TEP102-style timers should ever do that kind of
bookkeeping, for precisely the reasons you listed in that other e-mail
- the base timer level should be straightforward and low-overhead, at
the expense of accuracy.

David