High-Throughput GPU Acceleration of MS-OSM for Multiple DM Trials with Optimized Kernel and Asynchronous I/O

Pulsar and fast radio burst observations widely adopt coherent dedispersion to compensate for dispersion effects introduced by the interstellar medium. The multisegment overlap-save method (MS-OSM) was proposed to alleviate the extremely large FFT requirement of conventional overlap-save (OSM) coherent dedispersion. In this paper, we present a high-throughput GPU implementation of MS-OSM for multiple DM trials (GMSDT). GMSDT integrates multiple DM trials into a unified processing pipeline while leveraging GPU streams and concurrent CPU threads to improve computational efficiency and GPU resource utilization. It is demonstrated that GMSDT significantly reduces GPU memory usage compared with conventional OSM processing, enabling the simultaneous processing of more DM trials on a single GPU. With a carefully designed GPU implementation, parallel processing of multiple DM trials can outperform the DM-by-DM sequential process. Furthermore, a CPU-GPU cooperative data persistence pipeline is proposed to support high-throughput dedispersion processing and storage for pulsar and FRB observations. Experimental results demonstrate high numerical consistency, significant performance improvements over conventional methods, and successful coherent dedispersion of real pulsar observation data.