Fast Failure Detection of OpenFlow Channels

Abstract

We propose a mechanism to detect OpenFlow channel fail- ures quickly in switches and controllers where multiple con- trollers are running. In OpenFlow networks, it is impor- tant to maintain OpenFlow channels between controllers and switches are up and to detect channel failures immediately, so that messages to notify events such as port down and mod- ification of flow tables are always delivered to the other side surely and quickly. Exchanging keep-alive messages fre- quently is undesirable for controllers because the controllers should handle many keep-alive messages from switches. This would be a significant overhead when the rate of other mes- sages than keep-alive ones is low, because the controllers are forced to handle many keep-alive messages although such keep-alive messages do not affect network control directly. Our proposed mechanism adaptively sends keep-alive mes- sages to detect quickly that a message is not reached to the other side, instead of checking whether a channel is up. A controller shares a message received from a switch with other controllers in a timely manner, and the controller regards a channel has been unavailable if the message notified by other controllers has not arrived via the channel. A switch sends keep-alive messages to all channels just after sending an im- portant asynchronous message such as a port status message, and regards all channels have been gone down if the switch does not receive any response. The evaluation shows that our proposed mechanism reduces failure detection delay to timeout until receiving a response, and that overhead on la- tency is negligible.