Demonstration of 10-Gbps Real-Time Reed–Solomon Decoding Using GPU Direct Transfer and Kernel Scheduling for Flexible Access Systems

In recent years, the applicability of network function virtualization and software-defined networks to optical access systems has been studied for reducing capital expenditure/operating expense and quickly responding to various service requests. We are aiming at expanding the footprint of software-defined functionality in optical line terminals by focusing on physical (PHY) layer functions as one key part of the flexible access system architecture. Utilization of common graphic processing units (GPUs) with a large amount of computation resources is promising for softwarizing PHY-layer functions that have high computation complexity. However, conventional studies do not consider sequential signal input and realize only Mbps-class processing. This paper proposes CPU control for GPU direct transfer in order to directly transfer an external signal input to GPU, and kernel startup scheduling so as to handle a sequential signal at 10 Gbps. For Reed-Solomon decoding, whose PHY-layer functions have high computation complexity, we demonstrate real-time performance from a general-purpose server with standard optical modules by implementing on it. The proposed configuration is evaluated using an optical system compliant with IEEE 802.3av. Evaluations show that the proposed GPU configuration realizes, for the first time, 10-Gbps real-time error correction.