HLL is for calculating the number of unique objects when there are possible duplicates. Unless you're doing something really weird you shouldn't have any problem going through all 20 million objects without encountering duplicates.

You're probably better of just keeping a running total. You can either update this total online (keeping track of all state changes) or whenever you need to know (going through a mere 20 million items shouldn't take too long).

If, for whatever reason, you really want to approximate it quickly, just pick a subset at random and calculate the percentages from there. Some basic Bayesian statistics can tell you the accuracy of this approximation.