high performance computing on graphics processing units: hgpu.org

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 20μs, representing less than 5% of time variation.