NPGPU: Network Processing on Graphics Processing Units

Abstract

The Internet is still expanding despite its already unprecedented complexity. To meet the ever-increasing bandwidth requirements under fast appearing new services and applications, today’s Internet routers and other key network devices are challenged by two conflicting requirements, high performance and good programmability. In this work, we propose a series of data-parallel algorithms that can be efficiently implemented on modern graphics processing units (GPUs). Experimental results proved that the GPU could serve as an excellent packet processing platform by significantly outperforming CPU on typical router applications. On such a basis, we proposed a hybrid microarchitecture by integrating both CPU and GPU. Besides dramatically enhancing packet throughput, the integrated microarchitecture could also optimize quality-of-service metrics, which is also of key importance for network applications. Our work suggests that an integrated CPU/GPU architecture provides a promising solution for implementing future network processing hardware.