[Ns-developers] NS3 Time objects

[Gustavo Carneiro]

On 11/2/06, Mathieu Lacage <Mathieu.Lacage at sophia.inria.fr> wrote:
>
> hi gustavo,
>
> On Thu, 2006-11-02 at 10:54 +0000, Gustavo Carneiro wrote:
> >   OK, let's finish this discussion, then; I do not wish to slow down
> NS-3
> > development.
>
> You are not slowing the development. This kind of discussion is
> necessary to fix and improve what we have and what we want to get. I
> think that both you and George have helped fix my original proposal into
> something much better. I really think that this sort of iterative
> discussion/implementation process is the way to go even though it takes
> time and a lot of efforts.


  Thanks; but  it's just that I really wanted to use NS-3 for some
simulations in 2-3 months for my phd, even if it's alpha by then :P  [ else
I'll have to use NS-2, nooo!!! :( ]

I would be fine with this. However, I have tried to actually implement
> the template thing I suggested yesterday. An not-too-simplistic
> implementation is attached to this email. I suspect you will like it a
> lot. As far as I am concerned, I do not see any downside to the approach
> so, I would be okay with pushing this new code to my personal repository
> with a bunch of extensive test cases. Help with the 128bit integer
> implementation of the HighPrecision class would be appreciated. It is
> not super hard but it is going to take time to implement and debug fully
> and my TODO list is getting longer every hour.


  I do not wish to comment on implementation details that have no impact on
the interface, though I'll be happy to spend some time to help with
implementation.  Here's some misc. comments:

-  About the design, what you call TimeUnit<N> I believe a better name would
be TimePower<N>.

-  Another item I would comment is that you implement TimeUnit<N> always as
a 128-bit number, in a separate HighPrecision class (in intent if not in
practice).  If one wanted to be precise we would consider that TimeUnit<N>
would require N times the storage space of Time<1>.  If we would take that
into consideration, we might as well get rid of HighPrecision and inline its
logic in the templated TimeUnit class itself, e.g.:

template <int N>
class TimeUnit
{
public:
  TimeUnit ();
  TimeUnit (TimeUnit const &o);
   /* ... */
private:
  int64_t m_high[N];
  int64_t m_low[N];
};

- I still have the opinion that a simple scalar value should be a plain
double, not some fancy class;

- I still believe that standard operators + - * / should be defined, not
methods like Add Sub, etc.

- The methods ApproximateToSeconds, ApproximateToMilliSeconds,
ApproximateToMicroSeconds, and ApproximateToNanoSeconds should only be
defined for TimeUnit<1>, though I don't know whether that can be achieved
with C++ templates...
  [ hm.. somehow I get the feeling looking at the code that this is already
implemented, but I don't have enough experience with templates to know for
sure ]

  I hope you find my comments useful.

 Regards,

-- 
Gustavo J. A. M. Carneiro
"The universe is always one step beyond logic."