I've used both a hierarchial directory scheme based on a numerical ID as well as relational blobs in postgres to store large amounts of pr0n data for my last employer. With the high performance storage hardware and frontend (web farm) caching framework I put together, the performance was excellent.