Towards an in-network GPU-accelerated packet processing framework

Software-defined networking and data-plane programmability have opened up the possibilities for switches to be used for novel applications that are different than simple packet forwarding. Various tasks from low-level robot control to signal and data processing can be offloaded to network devices. In the past years, solutions exploiting programmable switching ASIC, FPGA or the combination of both have emerged. In this paper, we propose a GPU-accelerated switch design for supporting payload processing tasks in the network. The proposed design combines the processing capabilities of GPUs and the kernel-bypass library DPDK. We define different image processing use cases that can benefit from in-network computing, allowing execution without the need for an external server. The proposed method cannot only make the overall system performance better, but also reduce the power consumption since it requires less hardware elements. We evaluate and compare three models: Traditional external server with GPU in the local network, DPDK accelerated version of the previous model and the proposed GPU-accelerated in-network computing switch model. We investigate several benchmarks including both component-level and system-wide analysis. The examined use cases are related to video stream processing tasks like box blurring, Gaussian blurring and edge detection, demonstrating the performance improvement of our proposed design.