Do you know the way the cells are reshuffled among the GP boards?
It is not based on peak traffic, but on other inputs. Maybe it is based on MIN PDCH, or MAX PDCH, or the number of TRX per cell.
I would first analyse the reshuffling algo, and try to force it to spread the cells in a more harmonious way. However, it may be easier to either add more GP boards, or globally reduce the PDCH consumption of each cell.
Could you remind me of your ALU release, and BSC generation ?