Hardware is constantly evolving; improving in efficiency and adding new features. Here, we look at the current status of relevant hardware solutions and talk to vendors about new developments on the horizon that will be available in the near future. We will then determine specific hardware solutions that may be of interest for MAELSTROM and machine learning applications in weather and climate science.
In MAELSTROM, many different scientists come together to work on different aspects of weather and climate simulations. The applications they are using have different requirements to the underlying hardware that is executing their tasks. In this deliverable, we take stock of the flavors of the applications and what explicit and implicit requirements they have. The results will be documented and published.
Using first versions of the MAELSTROM data-sets in the respective machine learning solutions, we study the performance on several computing resources that are available to us. We look how the solutions behave on the hardware platforms and use performance engineering tools to make quantitative and qualitative assessments. The results will be documented and published.
In order to run the MAELSTROM ML solutions at large scales, an HPC hardware infrastructure is needed. On this infrastructure, we establish a workflow to quickly determine performance of pre-defined workloads of the ML solutions.