Version 1 (modified by mcastril, 18 months ago) (diff)

Discussion of Development of a mixed-precision option within NEMO. Grenoble, France.

Last edition on 02/01/19 09:52:59 by mcastril

Participants

  • BSC: Miguel Castrillo, Oriol Tintó
  • CNRS: Gurvan Madec, Claire Levy, Sebastien Masson, Rachid Benshila
  • UKMO: Mike Bell
  • ECMWF: Nils Wedi
  • INRIA: Laurent Debreu
  • Ocean Next: Laurent Brodeau

Agenda

  • Presentation of the work by the BSC
  • Discussion

Minutes

Oriol gave a brief explanation of the motivation and the methodology, describing the initial test used to assess the impact (using differences from simulations with different timesteps).

Gurvan noted that the Asselin filter parameter should be modified when the timestep length is changed and that certain parts of the code would be expected to be sensitive to precision: for example the tri-diagonal solver in the vertical diffusion, the 2D barotropic mode, accumulation of truncation error in the surface pressure gradient and the vertical scale factors. Other aspects like biharmonic dissipation should not need high precision calculation. It was agreed that it would be interesting to see whether Oriol's results are consistent with these insights.

Nils asked whether there was a concensus that mixed precision is worth investment of effort. It was agreed that it is because it can give nearly a factor of 2 reduction in cost. Recent simplification work in the communication interfaces should also assist the creation of single-precision versions of the communication routines.

Nils asked what approach would be taken to implementation. It was agreed that code maintenance is a priority and mechanisms to do this using cpp/fpp substitutions were discussed.

It is difficult to define "one size fits all" tests for mixed precision calculations. A useful approach is to explore the sensitivity of results for specific processes in idealised test cases. There was a consensus that tests to assess the accuracy of the results would not only be useful for the work on mixed precision, but also for the NEMO community in general.

It was agreed to explore mixed precision further. To take things forward, Oriol will visit IPSL for a few days in March when Gurvan is back from Australia.