A Preliminary Performance Model for Optimizing Software Packet Processing Pipelines

• DOI: 10.1145/3124680.3124747
• Publication Date: 2017-09-05T12:23:44Z
• Authors: Ankit Bhardwaj; Atul Shree; V. Bhargav Reddy; Sorav Bansal
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology

Software packet processing is increasingly commonplace, especially for software-defined networking constructs. Previous work has investigated methods to efficiently map packet processing pipelines to general-purpose processor architectures. Concurrently, novel high-level domain-specific languages (DSLs) for specifying modern packet processing pipeline functionality, are emerging (e.g., P4 [5]). An attractive goal is develop a compiler that can automatically map a high-level pipeline specification (specified in a high-level DSL) to an underlying machine architecture. Ideally, the compiler should automatically exploit the available parallelism, make intelligent scheduling decisions, and adapt to the workload needs in an online fashion, to provide maximum performance. An important pre-requisite for the development of such a compiler is a performance model of the underlying machine architecture, for the applications of interest. We report our experiences with adding an optimizer to the P4C compiler [14], which compiles a high-level P4 program to a lower-level C-based implementation that runs with the DPDK infrastructure [1], and gets eventually executed on a multi-socket x86 machine. We make two contributions: (a) we show that significant performance improvements (up to 55%) can be gained by adding scheduling and prefetching optimizations to the P4C compiler; and (b) we develop a preliminary performance model for reasoning about the expected throughput and latency of a packet-processing workload, on a modern machine architecture. Our model can be used by a compiler, to reason about the expected performance of a packet-processing workload for different code configurations, and can thus be used to optimize the generated code accordingly.