Achieving high performance is a complicated task. Different levels have to be exploited - from single core performance up to cluster-wide performance. Furthermore, different techniques and skills have to be used.
Held virtually on October 26-29, 2020 the PRACE course "Tools & Techniques to quickly improve performances of HPC applications in Solid Earth" was developed in the CHEESE framework by CINECA staff to mostly focus on the main HPC issues for Earth Science.
The main objective of the course was to help Solid Earth researchers improve performance and face some important features like urgent computing, hazard assessment and early warning forecast for Exascale supercomputing.
In three full days, the course instructors provided in-depth analysis of all main issues such as:
- Single core performance
- vectorization
- data access
- Intranode node parallelization
- Shared Memory optimization (OpenMP)
- OpenACC (for GPU)
- inter-node parallelization
- Message passing optimization (MPI)
- Putting all together (MPI+OpenMP, MPI+Openacc)
- I/O issues
The main aspect of the course was to show participants typical problems that they could face when working with HPC systems and in future Exascale supercomputers. They were provided with some possible solutions such as parallel paradigm and the technological solutions.
In a survey sent after the end of the course, participants said that they would recommend this course to others and that they were satisfied with the training materials and the course instructor, Giorgio Amati.
