What is it about?
We present the first performance measurements on HPC-grade hardware of DPEcho, our MPI-parallel SYCL porting of the General-Relativistic Magneto-Hydrodynamics (GR-MHD) code ECHO for numerical simulations of Astrophysical Black Holes. We take advantage of the recent additions in SYCL2020 to benchmark GPUs and CPUs of different vendors with a representative workload from numerical astrophysics with DPEcho.
Featured Image
Photo by BoliviaInteligente on Unsplash
Why is it important?
Having chosen SYCL, we can take advantage of the unprecedented power offered by heterogeneous computing systems, including the latest HPC-grade GPUs. Unlike proprietary programming models (e.g. CUDA), SYCL is a cross-platform layer, not llimited to single vendors' hardware: this alleviates the need for redundant development, maintenance and coordination of multiple GPU portings. At the same time, SYCL allows for more fine-granular control than other, more general programming methods such as OpenMP offload.
Perspectives
Developing SYCL code for numerical astrophysics and designing algorithms around native parallel SYCL constructs has been a fun and rewarding experience. Topped off by being able to run it independently on hardware offered by any computing cente, and even home or laptop GPUs.
Salvatore Cielo
Leibniz-Rechenzentrum
Read the Original
This page is a summary of: Portability and Scaling of the DPEcho GR-MHD SYCL code, April 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3585341.3585382.
You can read the full text:
Resources
Contributors
The following have contributed to this page
