it turns out, I can simply store the bi-gram array directly as part of the hash..

That will work, but be aware that accessing the data in a shared data structure is significantly slower than accessing non-shared memory. This is due to the need for internal locking amongst other things.

So if you run a second pool of threads, each accessing that shared hash containing shared arrays, the lock contention between the threads will likely slow your processing to a crawl. (A simple fact of life with perl's shared data structures).

Try the version I posted and see how you get on.