Re: GEQO: ERX

Adriano Lange escreveu:
> I implemented the 2PO algorithm last month but I didn't have much time
> to do an extensive test and to comment all code. The code was posted in
> this list in a previous thread. In that occasion, I was interested in a
> kind of cache structure to avoid the constructing a complete RelOptInfo
> from scratch every time when the cheapest total_cost must be calculated
> (this occur in GEQO).
>
> I’m sending a patch for the 8.3 release.

I forgot it.

>
> I also changed some GUC variables to facilitate some tests:
> (remove) geqo
> (remove) geqo_threshold
> (add) ljqo_enable (bool) – activate Large Join Query Optimizer
> (add) ljqo_algorithm {geqo|twopo}
> (add) ljqo_threshold (int) – like geqo_threshold
> (add) twopo_heuristic_states (bool) – initial heuristic states
> (add) twopo_ii_stop (int) – II phase loop
> (add) twopo_sa_phase (bool) – enable SA phase
> (add) twopo_sa_initial_temperature (float)
> (add) twopo_sa_temperature_reduction (float)
> (add) twopo_sa_equilibrium (int)
>
> In my little tests, this algorithm seems equal or worse than geqo,
> except when using heuristic in order to bias the initial state. Maybe
> some tunings are needed but I prefer spend yet some time reading more
> about the compressed annealing, cited in TODO list. Anyway, I think that
> to build another annealing-like algorithm might be easier if some
> structures and functions in 2PO source code are correct.
>
> Sincerely,
>
> Adriano Lange
>