Changes in Intent: Behavioral Predictions of Distributed SDN Controller Reconfiguration

Abstract

Intent-based programming enables software-defined networks (SDN) to be able to dynamically reconfigure themselves through automatic intent recomputation in response to network events, such as host mobility. This allows SDN to be used as a platform for new technologies such as swarms of drones in data-driven agriculture. At the same time, this dynamicity results in SDN networks having a very large state space - whose size is further exacerbated when SDN controllers are distributed for reliability and scalability. This renders infeasible comprehensive testing or verification of network performance prior to deployment, necessitating the use of monitoring at run-time, together with associated abortive or healing actions to ensure reliability. However, as intent recomputation time can vary significantly based on the underlying network topologies, it is very difficult to experimentally determine the boundary between normal expected performance and anomalous performance at scale, and hence to specify when these actions should take place. In this paper, we demonstrate the use of machine learning to automatically learn intent recomputation performance; the resulting predictions can be used as input into the specification of run-time monitors and the determination of associated reliability mitigations. More specifically, we describe our proof-of-concept case study on using linear regression to predict the expected time for intent recomputation due to host mobility on the distributed ONOS open-source SDN controller.