P5

Abstract

The physical pipeline of flexible network switches is usually programmed using packet-level programs, such as P4 programs. However, those programs are low level and leave room for further optimization. We propose P5 (Policy-driven optimization of P4 Pipeline), a system that exploits knowledge of application deployments embedded in a high-level policy abstraction to: 1) detect features that are used by applications in a mutually-exclusive way and thereby remove inter-feature dependencies between the tables implementing these features in a network switch. This improves the pipeline concurrency of switches and hence its pipeline efficiency. 2) detect and remove the features that are not used by any application/traffic on the switch in a given topology. This reduces the number of tables and the resource consumed by switches, which also improves its pipeline efficiency. Our experiments on real P4 switch programs show that the resulting switch pipelines are up to 50% more efficient as compared to the cases that do not exploit this information.