On Sun, May 14, 2017 at 20:29 +0000, Carl Mascott wrote:
> It looks to me like you are rounding on each iteration of the for-loop:
>
> + for (i = 0; rate > 9999 && i <= 3; i++) {
> + rtmp = rate / 1000;
> + if (rtmp < 9999)
> + rtmp += (rate % 1000) / 500;
> + rate = rtmp;
> + }
>
> Am I missing something?
>
rtmp is less than 9999 only on the last iteration.
Perhaps you're referring to "rate / 1000" as rounding
but in fact it just throws those digits away.
> I'll post my patch on Tuesday.
>
> Yes, I understand that systat can display only 4 digits for BW.
> That's 5 digits with my guard digit, which is shifted out (and not displayed)
> at the end.
> So, with the guard digit, 6 digits is too many.
>
>
> --------------------------------------------
> On Sun, 5/14/17, Mike Belopuhov <m...@belopuhov.com> wrote:
>
> Subject: Re: pf queue definition: bandwidth resolution problem
> To: "Carl Mascott" <cmasc...@yahoo.com>
> Cc: tech@openbsd.org, t...@openbsd.org
> Date: Sunday, May 14, 2017, 4:05 PM
>
> On Sun, May 14, 2017 at 19:48 +0000, Carl
> Mascott wrote:
> > I have a suggestion RE
> your pftop.c patch. You are rounding
> >
> multiple times, after each scale operation. This is known
> as
> > rounding the intermediate results of
> a calculation and degrades
> > accuracy.
> If you're not familiar with the issue do a Google
> search
> > on rounding intermediate.
> >
>
> I round
> only once, that's what I've explained in my
> mail.
>
> > I suggest assigning the
> pf rate to an unsigned long and multiplying by 10.
> > This makes the LSD a guard digit.
> > After all scaling, round once (if guard
> digit >= 5 then add 10).
> > Yes, this
> may require one more scaling operation if it rounds up to 6
> digits (including guard digit).
> > At the
> very end divide by 10.
> > Note: This is
> essentially fixed point arithmetic with one decimal
> digit.
> >
> > I modified
> pftop.c v1.37 to do this earlier today.
> >
> It was kind of tricky.
> > I need to see if
> it still looks OK on Tuesday (I'm busy Monday).
> > Let me know if you want a patch then.
> > I won't have actually tested it,
> though, so it's highly likely to have bugs.
> > You might be better off writing your own
> (and then perhaps comparing to mine).
> >
>
> > RE my email client: It's Yahoo
> webmail -- nothing I can do about it.
> >
>
> pftop in
> systat can only display 4 digits and 1 unit symbol
> so you can't display all digits of your
> bandwidth spec.
> But please by all means,
> send your diff and I'll take a look at it.
>
>
