2nd P4 Workshop by Stanford/ONRC

GPGPUs, as an SIMD computing architecture, have shown scalable performance on network applications, therefore have gained increasing interest from the P4 community to explore as a potential target. In our work, we study the mapping of a P4 program into GPGPU by dissecting and analyzing the “simple router” example given by P4 community, and implementing an equivalent simple router on a low-end NVIDIA GPU, following the forwarding model as defined in P4 language. In our experiments, we use a GPGPU with 384 CUDA cores, each running at 500 MHz. Our preliminary evaluation of the prototype demonstrates a promising throughput of the simple router on GPU. Experimental results show that the P4 equivalent simple router on GPU can achieve over 582 Gbps of bandwidth with near-ideal packet transmission I/O, and up to 21 Gbps of bandwidth with the socket-based I/O.