OpenFlow Timeouts Demystified

This paper explores the impact of the timeout length on performance, measured through the miss rate, and table occupancy. It finds that as timeouts increase, the miss rate decays exponentially while the table size grows near-linearly. We observe there is an operating point where any further increases in the timeout lead to insignificant reductions in the miss rate while unnecessarily swelling expanding the table occupancy. In one dataset this timeout is at 5 seconds, while it is centered around 10 for three others. Additionally, this paper introduces hybrid flow table management that combines timeouts with explicit controller eviction messages. It establishes a lower bound of 57% fewer table entries in one dataset without impacting the miss rate, while a practical TCP-based implementation reduces the table size by 42%, or almost 32,000 entries. Finally, this work compares the performance of different flow table replacement policies. It identifies that while LRU outperforms the other policies, it cannot be implemented in current OpenFlow switches. A FIFO replacement policy performs [AZ: number!] worse than LRU but less than 0.1% better than Random replacements. Based on our observations, the future of OpenFlow timeouts is to delegate the responsibility of assigning timeouts to a controlloop within the controller so the timeout can both be dynamically tuned to individual flows, as well as individual network requirements.