Version 1 (modified by epico, 2 weeks ago) (diff)

NEMO HPC WG: Tue 28 Jul 2020

0. Attendees

Oriol Tinto (BSC), Clement Bricaud (Mercator), Daley Calvert (Met Office?), David Guilbert (Atos/Bull?), Francesca Mele (CMCC), Kristian Mogensen (ECMWF), Maff Glover (Met Office?), Maicon Faria (BSC), Mario Acosta (BSC), Martin Shreiber (TUM), Michael Lange (ECMWF), Miguel Castrillo (BSC), Mirek Andrejczuk (Met Office?), Nils Wedi (ECMWF), Sam Hatfield (ECMWF), Silvia Mocavero (CMCC), Stella Paronuzzi (BSC), Xavier Yepes (BSC), Italo Epicoco (CMCC).

1. Presentations

  • Advances on mixed precision and its impact on developers (Oriol Tinto - BSC)
  • ESiWACE2: OIFS-XIOS benchmarking (Mario Acosta - BSC)
  • Scaling NEMO4 I/O using the new ORCA36 configuration (Miguel Castrillo - BSC)

(see attached slides)

2. Discussion

  • I.Epicoco introduces the meeting underlighting the role of the NEMO HPC-WG. + The HPC-WG should be more proactive in suggesting the HPC development actions for the next years moving and following the guidelines depicted in the NEMO Development Strategy Document + The HPC-WG should evaluate, assess and discuss about the matureness and effectiveness of the currently developing branches giving some feedback to the NEMO ST. + The HPC-WG should investigate about new HPC methods, solutions and optimizations that can be introduced in NEMO HPC development strategy.
  • Comments regarding mixed precision and its status + interesting are the first thoughts depicted by O.Tinto regarding the guidelines a NEMO developer should follow when facing up with the mixed precision code + I.Epicoco proposed to execute the performance assessment also on other HPC architectures; O.Tinto confirmed that an automatic procedure is already under development at BSC for automatic performance evaluation, it could be shared to make the tests easier.
  • Comments regarding the XIOS benchmarking + quite impressive the detailed test plan presented by M.Acosta to assess the XIOS performance with IFS + as partial conclusion after the first evaluation of XIOS with NEMO ORCA36 given by M.Castrillo is that XIOS 2.5 needs still further investigations in order to fully exploit all of its functionalities. At the moment XIOS suffers of a high memory footprint. I.Epicoco agrees on that. + K.Mogensen suggested to use XIOS by saving each variable to a different file, this should tradeoff between the advantage given by writing of different and independent files with the advantage of having the whole domain of a variable saved in a file.

Attachments (3)