Branch Prediction For Network Processors

Originally designed to favour flexibility over packet processing performance, the future of the programmable network processor is challenged by the need to meet both increasing line rate as well as providing additional processing capabilities. To meet these requirements, trends within networking research has tended to focus on techniques such as offloading computation intensive tasks to dedicated hardware logic or through increased parallelism. While parallelism retains flexibility, challenges such as load-balancing limit its scope. On the other hand, hardware offloading allows complex algorithms to be implemented at high speed but sacrifice flexibility. To this end, the work in this thesis is focused on a more fundamental aspect of a network processor, the data-plane processing engine. Performing both system modelling and analysis of packet processing functions; the goal of this thesis is to identify and extract salient information regarding the performance of multi-processor workloads. Following on from a traditional software based analysis of programme workloads, we develop a method of modelling and analysing hardware accelerators when applied to network processors. Using this quantitative information, this thesis proposes an architecture which allows deeply pipelined micro-architectures to be implemented on the data-plane while reducing the branch penalty associated with these architectures