Enhanced algorithms for consistent network updates

Abstract

The basic algorithm that consistently updates the switches of a Software Defined Network while preserving the update property of per-packet consistency is the two-phase update. The two-phase update is underspecified on two matters: 1) how to detect when the last packet of the old rule set has left the network and therefore exactly when to delete the old rules 2) recoverability of the update. Recoverability ensures that those updates for which the two-phase algorithm is not completely executed do not change the semantics of those updates for which the algorithm is completely executed. This paper examines the failures that can occur during an update and how to handle a subset of those failures. It proposes an algorithm, enhancing the two-phase update to handle deletion of old rules and recoverability. It also specifies how to effectively use a software cache to supplement the TCAM, during an update. The paper extends the same algorithm for per-flow consistent updates, using a restricted number of exact-match rules for existing flows and specifying when to delete the exact-match rules. The two algorithms are also specified for switches that do not support a software cache. The paper also analyzes the algorithms quantitatively, identifying the parameters of interest and what they depend upon.