In-Network Edge Split Inference via Named Data Networking

Split inference (SI) has been devised as a valuable solution to enable the execution of computation-heavy deep neural network (DNN) inference models on resource-constrained edge devices. The different layers of a DNN model are distributed to one or several nodes (mainly edge/cloud servers) cooperating with the end-device requesting the inference. Boosted by the sixth generation (6 G) trends, programmable network nodes equipped with computing, caching and intelligence capabilities can be involved in such a cooperative task. In this work, we propose Named Data Networking (NDN) as a key enabler of in-network SI. Its connectionless communication model coupled with routing-by-name and native in-network caching capabilities can facilitate dynamic splitting operations on nodes throughout the cloud-to-things continuum. We show how NDN design principles and communication primitives can be leveraged to support in-network SI. Then, preliminary results are reported to showcase the benefits of the conceived proposal.