NEMO HPC subgroup: Thu 21 June 2018

Attendees: Sebastien Masson (IPSL), Eric Maisonnave (CERFACS), Mike Bell (Met Office), Mirek Andrejczuk (Met Office), Matthew Glover (Met Office), Martin Price (Met Office), Andy Porter (STFC), Miguel Castrillo (BSC), Oriol Tinto (BSC), Mario Acosta (BSC), Erwan Raffin (Bull), David Guibert (Atos), Clement Bricaud (MERCATOR), Michael Lange (ECMWF), Kristhian Mogensen (ECMWF), Nils Wedi (ECMWF), Silvia Mocavero (CMCC)

1. Mixed-precision (BSC)

Oriol presented the methodology implemented to test the impact of numerical precision in NEMO. An automatic tool to implement the Reduced Precision Emulator to a computational model has been presented, as well as the tests on groups of variables in order to identify which variables can use less precision without affecting the numerical accuracy. Tests have been performed on NEMO 3.6 GYRE1 and showed that only ~10% of the variables need double precision (see attached slides)

Open questions and future plan: need to test the approach on high-resolution (eddy resolving scale) and to evaluate the error. Which is the tolerance to safely use reduced precision?

2. NEMO-DSL (STFC)

Andy presented the light-DSL approach designed to apply DSL in NEMO. An interface to the PSyclone tool has been implemented in order to apply PSyclone transformations without impacting on the NEMO coding structure rules. The approach would allow to introduce the OpenACC parallelisation through a workflow based on two steps: processing the NEMO code to create an internal representation compliant to the PSyclone tool, manipulating the intermediate code to perform PSyclone transformations (see attached slides)

Open questions and future plan: considering the integration of optimisations for CPUs

3. Benchmark setup (CERFACS-CNRS)

Eric presented the NEMO benchmark implemented to identify the main bottlenecks to the scalability. Some results on the BENCH-1 on Intel Broadwell system at Meteo-France have been shown comparing the reference simulation (without MPI collectives) with the pseudo double size halo and no-communication experiments. The trend of the time spent waiting for communications and for computation load imbalance has been analysed (see attached slides)

Open questions and future plan: extension of the tests to high-resolution configurations

4. Next meeting call

Next meeting will be in July. Silvia will send the doodle poll