2026-04-14T04:46:17Zhttps://recercat.cat/oai/request

oai:recercat.cat:2117/4177622026-01-30T08:35:32Zcom_2072_1033col_2072_452950

Event-based OpenMP tasks for time-sensitive GPU-accelerated systems Cetre, Cyril Yu, Chenle Royuela Alcázar, Sara Barrère, Rémi Quiñones Moreno, Eduardo Gratadour, Damien Barcelona Supercomputing Center Àrees temàtiques de la UPC::Informàtica::Arquitectura de computadors Time-sensitive systems GPU High-performance computing OpenMP CUDA ROCm The throughput-centric design of GPUs poses challenges when integrating them into time-sensitive applications. Nevertheless, modern GPU architectures and software have recently evolved, making it possible to minimize overheads and interference along the critical path through advanced mechanisms, such as GPU graphs, while sustaining high throughput. However, GPU vendors provide programming ecosystems specific to their products, raising concerns about code portability. Hence, there is a need for a hardware-agnostic API capable of managing time-sensitive GPU-accelerated pipelines. In this context, we propose integrating event-based synchronizations into the high-level OpenMP programming model to, in combination with GPU graphs, notably reduce interference and overheads over the critical path. This work showcases how this combination offers significant performance improvements and time consistency. We also enable portability across several vendor ecosystems and demonstrate our work on a set of representative applications for cyber-physical systems. According to our experiments, we measured a maximum jitter below 20 µs, representing less than 5% of time variation. This work is supported by the RisingStars project, under the Marie Skłodowska-Curie grant agreement No 873120 from the European Union’s Horizon 2020 research and innovation programme. This work is co-financed by the ASCENDER project of the UNICO I+D Cloud program that has MINECO and the EU-Next Generation EU as financing entities, within the framework of the PRTR and the MRR. This work is also partly supported by the HiPERT project, with reference PID2023148117NA-I00, financed by MCIU/AEI/10.13039/501100011033/ FEDER, UE. Peer Reviewed Postprint (published version) 2024 Part of book or chapter of book Cetre, C. [et al.]. Event-based OpenMP tasks for time-sensitive GPU-accelerated systems. A: "Advancing OpenMP for Future Accelerators: 20th International Workshop on OpenMP, IWOMP 2024: Perth, WA, Australia, September 23-25, 2024: proceedings". Springer, 2024, p. 31-45. 978-3-031-72567-8 https://hdl.handle.net/2117/417762 10.1007/978-3-031-72567-8_3 eng https://link.springer.com/book/10.1007/978-3-031-72567-8 info:eu-repo/grantAgreement/EC/H2020/873120/EU/RISE International Network for Solutions Technologies and Applications of Real-time Systems/Rising STARS info:eu-repo/grantAgreement/AEI//PID2023148117NA-I00 http://creativecommons.org/licenses/by/4.0/ Open Access Attribution 4.0 International 15 p. application/pdf Springer