Coded Caching: Content-centric wireless networks

Communication networks are traditionally connection-centric, i.e., establish a reliable data connection between end-nodes. Recently, there has been a significant shift in network usage, where the users’ intent is to access some specific (broadband) content rather than connect to a specific node. Recognizing this, there has been a significant effort in “information-centric networks”, where the focus is on how to name data and access information rather than connecting to particular servers. However, these solutions are insufficient for wireless networks where the bottlenecks are in communication as well as the memory needed to store/cache data.

We have developed a new approach based on “coded caching”, where one jointly optimizes what is stored along with delivery. The challenge is that one needs to store before knowing the demand, but make the limited storage (e.g., in wireless access points) enable most efficient delivery of demanded data. We harness (instead of avoid) wireless broadcast and superposition properties along with strategic storage to satisfy user demands. We have characterized the fundamental trade-off between the memory used for storage with the rate (or number of bits) needed to satisfy any user demand, for many different scenarios and network topologies, with both explicit schemes as well as information-theoretic impossibility results.

In hierarchical (tree) networks, where there are layers of caches and broadcast transmissions, we have shown how to strategically place data as well as what should be the relaying/forwarding strategies to obtain this optimal trade-off. The optimality is shown with respect to information-theoretic impossibility results.

For the case where the content popularity has a non-uniform distribution (as is usually the case), we have shown how one should optimally share the memory between the popularity classes. The information-theoretic impossibility results here needed new tools using subset entropy inequalities.

We solved the open question of coded caching in interference networks, enabled by a fundamental separation result which showed that one can layer the coded caching with the underlying noisy interference network. This separation principle is a fundamental architectural insight that we think holds for more general architectures. We have summarized this perspective in a book chapter on this topic.