Towards a Compilation Infrastructure for Network Processors

Modern network processors (NPs) typically resemble a highly-multithreaded multiprocessor-ona-chip, supporting a wide variety of mechanisms for on-chip storage and inter-task communication. NP applications are themselves composed of many threads that share memory and other resources, and synchronize and communicate frequently. In contrast, studies of new NP architectures and features are often performed by benchmarking a simulation model of the new NP using independent kernel programs that neither communicate nor share memory. In this paper we present a NP simulation infrastructure that (i) uses realistic NP applications that are multithreaded, share memory, synchronize, and communicate; and (ii) automatically maps these applications to a variety of NP architectures and features. We use our infrastructure to evaluate threading and scaling, on-chip storage and communication, and to suggest future techniques for automated compilation for NPs.