Network Function Offloading through Classification of Elephant Flows

With the move from traditional hardware appliance based network functions to Network Function Virtualization, software development is decoupled from the hardware. However, as a network function is no longer optimized for hardware, beneficial features of networking hardware may not be used any more. Solutions such as SDN or NIC offloading aim to overcome this antipodes by integrating networking hardware into the packet processing pipelines. On the one hand, offloading traffic of network functions to hardware can increase throughput and reduce resource consumption. On the other hand, the number of parallel flows in a network can be very high, exhausting the capacity of the tables of the networking hardware and force the system to fall back to software processing. Fortunately, it is known that a large portion of the flows in the Internet are mice flows, whereas the majority of the traffic is constituted by elephant flows. If the elephant flows can be detected efficiently, the hardware tables can be used more efficiently as a larger share of the traffic can be offloaded. We introduce a machine learning based approach that takes its decision with the first packet of a flow. A fundamentally different approach is using packet sampling for the offloading decision. We are evaluating both approaches in terms of complexity, offloaded share of the traffic and table occupation. The results show that a machine learning based offloading decision is possible with the first packet. The sampling approach only reaches a comparable performance at very high sampling rates.