Reducing Data Copies between GPUs and NICs

Abstract

Cyber-physical systems (CPS) must perform complex algorithms at very high speed to monitor and control complex real-world phenomena. GPU, with a large number of cores and extremely high parallel processing, promises better computation if the data parallelism often found in real-world scenarios of CPS could be exploited. Nevertheless, its performance is limited by the latency incurred when data are transferred between GPU memory and I/O devices. This paper describes a method, based on zero-copy processing, for data transmission between GPUs and NICs. The arrangement enables NICs to directly transfer data to and from GPU. Experimental results show effective data throughput without packet loss.