Open sannant opened 3 years ago
It was surprising me that I need to allocate 60Gb of RAM in order to compute a free system with gridsize 8192^2 (so effectively quasi 16384^2).
Probably the Bugnicourt algorithm and the primal-objective are not optimized for memory efficiency.
However, I also made periodic calculations with 16384^2 with the old CCG nonadhesive solver and Bugnicourt with adhesion.
CCG without needs 40Gb Bugnicourt with adhesion needs 54Gb
This is not a huge difference, (corresponds to 7 more float fields à 2Gb), and some extra memory usage is expected from the adhesion.
So finally I am actually happy with it.
@sindhu2993 , @pastewka, if you share my opinion, I just close this again.
It was surprising me that I need to allocate 60Gb of RAM in order to compute a free system with gridsize 8192^2 (so effectively quasi 16384^2).
Probably the Bugnicourt algorithm and the primal-objective are not optimized for memory efficiency.
However, I also made periodic calculations with 16384^2 with the old CCG nonadhesive solver and Bugnicourt with adhesion.
CCG without needs 40Gb Bugnicourt with adhesion needs 54Gb
This is not a huge difference, (corresponds to 7 more float fields à 2Gb), and some extra memory usage is expected from the adhesion.
So finally I am actually happy with it.
@sindhu2993 , @pastewka, if you share my opinion, I just close this again.