You may find that a not-insignificant amount of the memory's being taken up by internal structure rather than extra preallocated space. SVs aren't small, and you've got a lot of pointers and pointers to pointers there. Not to mention hash key structs which are 8 bytes plus the key string, per key...

There's not much you're going to do to reduce in-memory costs, I think, without a lot of extra speed hit. Use a DB, as hasn been suggested--Berkeley DB or even a plain NDBM/GDBM file will probably serve you well here.