Re: fd_vector_max/1

[Daniel Diaz]

Hello Nicolas,

The current implementation of FD variables uses of a bit-vector to 
represent sparse values. This vector is used as soon as "holes" appear 
in the domain of a variable.
The current implementation uses a fixed-length vector whose length is 
controlled via the fd_vector_max/1 predicate.
This choice has been done for efficiency reasons. Other representations 
are possible but 1) imply a lot of code rewriting in the solver and 2) 
will penalize efficiency.

In your case (IP addr) it is obviously not a good idea to use a 
bit-vector for 2^32 values (and in fact you cannot define a so big 
length) !
Use instead 4 different variables ranging in 0..255 (or 2 variables in 
0..65535) and "link" them by constraints.

Daniel

Le 22/02/2011 08:12, Nicolas Pelletier a écrit :
> Hello,
>
> I would like to know how difficult it would be to lift the restriction
> that the maximum length for vector representation of FD variables is
> fixed, and what the expected impact on general performance would be. I
> am generating the configuration of a network equipment based on
> constraints on the functions that must be activated (and some other
> constraints on their respective configurations as well). Parts of the
> configuration constrain the domains of the variables used in other
> parts as some functions cannot be used together, or reserve some
> resources, such as IP address ranges. Some of my FD variables
> represent such ranges, and can sometimes span a big chunk of IP
> addresses. I am reluctant to split these into several FD variables,
> each spanning only 256 addresses or so, as this would make the program
> more complex and error-prone, but I am also concerned by the waste of
> memory I would incur if I were to set fd_vector_max to a huge value;
> hence my question.
>
> Thank you in advance for any answers.


--
Ce message a ete verifie par MailScanner
pour des virus ou des polluriels et rien de
suspect n'a ete trouve.