|P is for Practical|
If I am understanding this correctly, basically you are masking the 20 LSB so that every 32 bit integer with the same 12 MSB will get packed into the same hash value string, meaning that there could be up to 2^12 = 4096 entries per hash key.
However, since I am packing inodes and since inodes are more-or-less sequentially assigned, it would seem that until 2^20 inodes have been assigned, there is no packing. So, wouldn't it be better to mask the MSB, rather than the LSB since the LSB would be relatively randomly uniformly assigned in most disk usage cases. i.e., wouldn't it be better to do something like:
or even maybe:
Indeed, your masking may partially explain why for 10e6 indexes, it takes 260MB, while doubling to 20e6 only increases to 426MB.
Also, since there are at most 2^12 duplicates, couldn't you save 2 bytes by packing just the parts that are not masked by fffff, so that you could pack it in a 'v', rather than a 'V'. i.e,. in my masking scheme:
I am no perl monk, so I may be missing something of course...
Finally, it might be interesting to play with masking different amount of bits to see the space-saving vs. lookup time tradeoffs for different degrees of sparseness.