Mistill: Distilling Distributed Network Protocols from Examples

Patrick Krämer, Johannes Zerwas, Andreas Blenk

Keywords: communication networks, distributed protocols

Abstract: New applications and use-cases in data center networks require the design of Traffic Engineering (TE) algorithms that account for application-specific traffic patterns. TE makes forwarding decisions from the global state of the network. Thus, new TE algorithms require the design and implementation of effective information exchange and efficient algorithms to compute forwarding decisions. This is a challenging and labor and time-intensive process. To automate and simplify this process, we propose MISTILL. MISTILL distills the forwarding behavior of TE policies from exemplary forwarding decisions into a Neural Network. MISTILL learns which network devices must exchange state with each other, how to process local state to send it over the network, and how to map the exchanged state into forwarding decisions. We show the ability of MISTILL to learn distributed protocols with three examples and verify their performance in simulations. We show that the learned protocols closely implement the desired policies.