I have a very large hash (over 3 GB) that may be accessed 100,000 times or more during the life of a script (once for each record input to the script). Instead of accessing the large hash so many times, each time I run the script I could build a much smaller hash, with at most a few thousand keys, based on the initial response to the search of the large hash. I could then see if a key exists in the small hash before even referencing the large hash.

This is exactly what a Bloom filter is for. Once you have built a sufficiently large hash, performance suffers since you have to look up a key over the large hash. Bloom filters help offset some of the cost by introducing probabilistic lookups. IOW, the cost of checking for existence of a key is pretty small compared to performing the actual lookup. Have a look at this old perl.com article and there are a few CPAN modules available too - namely, Bloom::Filter and Bloom::Faster.

HTH, good luck!