Kugelblitz: Streamlining Reconfigurable Packet Processing Pipeline Design and Evaluation (2305.08435v1)

Abstract: Reconfigurable packet processing pipelines have emerged as a common building block for offloading fast and efficient packet processing to programmable switches and SmartNICs. However, designing or extending such architectures is complex, laborious, and resource intensive, due to the large design space coupled with the need for building architecture-specific compilers and RTL implementations before any evaluation is feasible. In this paper, we present Kugelblitz, a framework for rapidly designing, implementing and evaluating reconfigurable packet processing pipeline architectures. Kugelblitz comprises domain specific abstractions for separately representing pipeline architectures and protocol processing programs, a compiler for automatically mapping programs to specific architectures, and a hardware generator outputting complete RTL implementations for specific pipeline architectures. Kugelblitz enables meaningful full-system evaluation including chip area and timing from VLSI synthesis of the generated RTL, and end-to-end system benchmarks running real application workloads in full system simulation. In our evaluation, we demonstrate that Kugelblitz generates realistic and efficient RTL implementations, scales to large pipelines, enables fast design exploration, and provides meaningful application offload performance results.