It is estimated that the amount of data coming out of an optical fiber is doubling every nine months and, thus, the growth rate in network bandwidth by far exceeds that of transistor density stated by Moore's law. This causes excessive strain on network infrastructure nodes such as routers which need to operate at line rate in order to keep up with the external bandwidth requirements. Consequently, manufacturers of network processors have developed a wide range of technologies including highly parallel and specialised architectures to cope with ever increasing processing demands. Software tool support, however, lags behind and most research in compiling for network processors has focused on improved sequential and parallel code generation. In this paper we show that not code, but data organisation is the key obstacle to overcome in order to achieve high performance on network infrastructure applications. We evaluate three specialised data transformations (structure splitting, array regrouping, and software caching) against the industrial EEMBC networking benchmarks and real-world data sets. We demonstrate that speedups of up to 2.62 can be achieved, but at the same time no single solution performs equally well across all network traffic scenarios. This clearly indicates that adaptive data transformation schemes are necessary to ensure optimal performance under varying network loads.