Efficiently executing the Dragonet network stack

Today’s network cards are getting more and more complex. One reason for this is that networks are still getting faster, while cores are not. Two common approaches for allowing network stacks to keep up, are distributing packets to multiple cores directly in the network card, and moving protocol processing fully or partially to the network card. However current network stacks were not designed to accommodate the varied sets of hardware features supported by different network cards, resulting in sub-optimal performance. This thesis discusses our approach of building a network stack based on modeling both the required network processing and the network card as dataflow graphs and then combining these graphs to arrive at a configuration for the network card and a description of what processing needs to be implemented in software. It extends on our earlier publications introducing the modelling based network stack approach, and discusses Dragonet, our implementation of a full network stack and its performance characteristics. Performance of the resulting implementation is evaluated and compared to Linux using multiple throughput and latency benchmarks. Our results showed that Dragonet is capable of providing applications with competitive and often superior throughput and latency, compared to the widely used Linux network stack.