Fully Programming the Data Plane: A Hardware/Software Approach

ABSTRACT: Software-Defined Networking (SDN) has emerged in recent years as a new network paradigm to de-ossify communication networks. Indeed, by offering a clear separation of network concerns between the management, control, and data planes, SDN allows each of these planes to evolve independently, breaking the rigidity of traditional networks. However, while well spread in the control and management planes, this de-ossification has only recently reached the data plane with the advent of packet processing languages, e.g. P4, and novel programmable switch architectures, e.g. Protocol Independent Switch Architecture (PISA). In this work, we focus on leveraging the PISA architecture by mainly exploiting the FPGA capabilities for efficient packet processing. In this way, we address this issue at different abstraction levels: i) microarchitectural; ii) programming; and, iii) architectural. At the microarchitectural level, we have proposed an efficient FPGA-based packet parser architecture, which is a major PISA's component. The proposed packet parser follows a feedforward pipeline architecture in which the internal microarchitectural has been meticulously optimized for FPGA implementation. The architecture is automatically generated by a P4- to-C++ compiler after several rounds of graph optimizations. The proposed solution achieves 100 Gb/s line rate with latency comparable to hand-written packet parsers. The throughput scales from 10 Gb/s to 160 Gb/s with moderate increase in resource consumption. Both the compiler and the packet parser codebase have been open-sourced to permit reproducibility. At the programming level, we have proposed a novel High-Level Synthesis (HLS) design methodology aiming at improving software and hardware quality. We have employed this novel methodology when designing the packet parser. In our work, we have exploited features of modern C++ that improves at the same time code modularity and readability while keeping (or improving) the results of the generated hardware. Design examples using our methodology have been publicly released.

Abstract

Software-Defined Networking (SDN) has emerged in recent years as a new network paradigm to de-ossify communication networks. Indeed, by offering a clear separation of network concerns between the management, control, and data planes, SDN allows each of these planes to evolve independently, breaking the rigidity of traditional networks. However, while well spread in the control and management planes, this de-ossification has only recently reached the data plane with the advent of packet processing languages, e.g. P4, and novel programmable switch architectures, e.g. Protocol Independent Switch Architecture (PISA). In this work, we focus on leveraging the PISA architecture by mainly exploiting the FPGA capabilities for efficient packet processing. In this way, we address this issue at different abstraction levels: i) microarchitectural; ii) programming; and, iii) architectural. At the microarchitectural level, we have proposed an efficient FPGA-based packet parser architecture, which is a major PISA's component. The proposed packet parser follows a feedforward pipeline architecture in which the internal microarchitectural has been meticulously optimized for FPGA implementation. The architecture is automatically generated by a P4- to-C++ compiler after several rounds of graph optimizations. The proposed solution achieves 100 Gb/s line rate with latency comparable to hand-written packet parsers. The throughput scales from 10 Gb/s to 160 Gb/s with moderate increase in resource consumption. Both the compiler and the packet parser codebase have been open-sourced to permit reproducibility. At the programming level, we have proposed a novel High-Level Synthesis (HLS) design methodology aiming at improving software and hardware quality. We have employed this novel methodology when designing the packet parser. In our work, we have exploited features of modern C++ that improves at the same time code modularity and readability while keeping (or improving) the results of the generated hardware. Design examples using our methodology have been publicly released.