Dynamic Flow Entry Timeouts Based Packet Nonuniform Sampling for Elephant Flow Detection

This paper proposes a controller-based elephant detection approach based on packet statistics sampled by dynamic timeouts of flow entries. It avoids transporting statistics generated at switches to controllers by directly counting statistics using the packets forwarded to controllers for the installation and reactivation of entries. By initializing entries with hard timeout of 0.008s, and keeping doubling it whenever the entries timed out until flows are detected as elephants or out of life spans (the timeouts of flow entries of flows that have been detected as elephants are switched to idle timeout of 5s), the approach achieves a detection accuracy around 90% while reducing the bandwidth usage to less than 30% of the original over packet traces caught from the same network. By giving various initial timeout adjustment strategy to TCP and UDP flows, the approach can maintain the same high detection accuracy and low bandwidth usage over various networks.