Guest Talk: Enabling Fundamental Cacheability for Distributed Deep Learning Training