Fast and flexible: parallel packet processing with GPUs and click

摘要: We introduce Snap, a framework for packet processing that outperforms traditional software routers by exploiting the parallelism available on modern GPUs. While obtaining high performance, it remains extremely flexible, with tasks implemented as simple modular elements are composed to build fully functional and switches. Snap is based Click router, which extends adding new architectural features support batched processing, memory structures optimized offloading coprocessors, asynchronous scheduling in-order completion. show can run complex pipelines at speeds commodity PC hardware building an IP router incorporating both IDS-like full-packet string matcher SDN-like classifier. In this configuration, able forward 40 million packets per second, saturating four 10 Gbps NICs sizes small 128 byes. This represents increase in throughput of nearly 4x over baseline running comparable CPU.