Flow Entry Timeouts Optimization over Software Defined Networks Supporting Elephant Flow Classification

Abstract

Elephant flow (elephant) classification is significant for network performance management and resource optimization. Classifying elephants over Software Defined Networks (SDNs) often relies on statistics counted by switches and transferred to controllers, leading to a huge control channel bandwidth occupation that degrades network latency and user experience. Classifying elephants using flow packets forwarded to controllers completely avoids moving statistics from switches to controllers, but couples flow entry timeouts to elephant classification accuracy, time efficiency, flow table size, control channel bandwidth usage, and network latency. This paper aims to find the best flow entry timeouts to optimize the objectives by modeling the objectives and formulating a multi-objective optimization problem. The number of objectives is further reduced to 3 and the problem is solved by Machine Learning (ML) approaches. Extensive evaluations are made over real packet traces. To the best of our knowledge, this work is the first effort that explores all the objectives related to flow entry timeouts when using flow packets forwarded to controllers to classify elephants over SDNs.