Optimization of Earth System Models in the path to the new generation of Exascale high performance computing Systems

 A Use Case by

Short description

In the last years, our understanding of climate prediction has increased successfully. This is possible thanks to the improvement of our global Earth System Models (ESMs), in order to represent an ever-increasing realism for our future climate, reducing latter uncertainties in these chaotic systems and reproducing feature impossible until now, using for this spatial resolutions of very high density. To make this possible, the new computational cost demanded is prohibitived and impossible to fulfill without a massive parallelization and the new era of exascale machines. However, the overhead produced by the new massive parallelization will be dramatic and new High Performance Computing techniques will be required. This will facilitate to exploit the exascale machines in a efficient way and set up ultra-high resolution configurations to run a ESMs as EC-Earth, which will be used to predict the climate change in the next decades and study extreme events such as hurricanes as examples. EC-Earth is a model system used in 11 different countries and up to 24 metereological or academic institutions to produce reliable climate predictions and climate projections. EC-Earth is composed by different components, being the most important ones OpenIFS as the atmospherical model and NEMO as the ocean model.

Results & Achievements

The new EC-Earth version on development is being tested for the main components (OpenIFS and NEMO) on Marenostrum IV, using a significant number of cores to test the new ultra-high resolutions of 10 kms in the horizontal side, using up to 2048 nodes (98,304 cores) for the NEMO component and up to 1024 nodes (49,152 cores) for the OpenIFS component.

Different optimizations (developed in the framework of ESiWACE and ESiWACE2 projects) included in these components have been tested to evaluate the computational efficiency achieved. For example, the OpenIFS version including the new IO parallel integrated make possible to produce the output of hundreds of Gigabytes increasing the execution time only a 2% comparing to the execution without IO, much better than the previous version witch produced an overhead close to 50%. In addition, this approach will allow to use the same I/O sever for both components, enabling more complex computations online and using a common file format (netCDF). On the other hand, preliminary results using the new mixed precision version integrated in NEMO proved an improvement of almost 40% in the execution time, without losing accuracy in the simulation results.

Objectives

EC-Earth is one of the ESMs which suffers from lack of scalability using higher resolutions, with an urgent improvement of its capability and capacity in the path to exascale. Our main goal is to achieve the good scalability of EC-Earth using resolutions up to 10km of horizontal spatial resolution, when the parallelization will be extreme. In order to achieve this, different objectives are in progress:

(1) The computational profiling analysis of EC-Earth. Analysing the main bottlenecks of the main the main components when an extreme parallelization is used.

(2) Trying to exploit high-end architectures efficiently, reducing the energy consumption of this model to achieve a minimum efficiency in order to be ready for the new hardware. For this purpose, different High Performance Computing techniques are being applied, such as the integration of a full parallel IO for all the components of EC-Earth known as XIOS, or the reduction of the precision from double to single of some variables used by the model, maintaining the same accuracy in the results but improving the final execution time of the model.

(3) Evaluating if massive parallel execution and the new methods implemented could impact the quality of the simulations or lose reproducibility.