Remove the nexus between trial_buffs and trial_conns, and don't test shared buffers lower than or equal to a value we already know is good, but use that value instead. This will make it easier to adjust the trial values and/or formulae in future if necessary.