Given two very large sparse matrices (won't fit into memory), how do you store them to serve a function to do matrix multiplication?

(This is more of a design question)