The position is part of the European project ENMASSE (Enhancing Nemo for Marine Applications and Services).
Climate models are extremely complex and require huge computational resources. Optimizing them is therefore both a necessity and a real challenge. The objective is to accelerate the NEMO code, which is one of the most popular components of these climate models. Since its creation, NEMO has been designed to run in double precision, i.e. using 64-bit floating-point numbers. Using lower precision can be beneficial in terms of performance by reducing memory usage, execution time (by two) and energy consumption. However, lower precision can also increase numerical instabilities. The main objective of this project is to automatically identify and control sources of instabilities when running NEMO in single precision. To support this, the plan is to use the CADNA library which relies on Discrete Stochastic Arithmetic to estimate rounding errors in numerical codes and detect numerical instabilities that may occur during their execution. The CADNA library will be used to control the numerical quality of the results computed by NEMO and point out the possible numerical instabilities. By leveraging CADNA, we will automatically identify instability sources and try to fix/bypass them by combining our expertise on NEMO algorithms with CADNA developers’ knowledge of numerical precision/instabilities. Different physical configurations of NEMO, either local or global, will be studied. We also plan to optimize the threshold values in NEMO, taking into account the various precisions used.
A complete description is given in:
https://www-pequan.lip6.fr/~jezequel/offers.html
