= Minutes NEMO HPC Working Group - 10 Feb 2016 meeting ==

1.       Check attendance (5 minutes)
Present: A. Porter, Matthew Glover, Tim Grahm, Jason Holt, Martin Schreiber, A. Coward, Dmitry Kuts, J. Appleyard, Kim Serradel, Lanton, M. Castrillo, Oril Tinto, Silvia Mocavero, Stan Posey, T. Bradley F. Krueger, C. Lévy

2.       Actions from previous meetings(20 mins)

- Action 1:  Miguel, Silvia, Tim and Claire to coordinate the set up and review of these two configurations 
Tim has created a branch in NEMO repository in branches/2016/dev_HPC_Gyre_benchmark_test/NEMOGCM/BENCHMARK to set up the benchmark for GYRE (first configuration chosen). He has started to add counters (more work to be done).
Miguel has worked on the ORCA025_LIM3 configuration to be added in this benchmark branch

- Action 2: Mike to put information from group meeting on the NEMO HPC WG wiki page. Not done yet.

- Action 3: Mondher will email Silvia, Miguel, Tim and Mike information on computers accessible to Mercator.
No answer yet from Mondher.

- Action 4: Claire, Tim, Silvia, Miguel to agree what version of NEMO to use for these benchmarks 
Decided to set up the benchmark in head of the trunk (future 4.0 release)

- Action 5:Mike to circulate the present draft of the HPC task description and the WP structure. Members to respect the confidentiality of this information.

3.       H2020 Copernicus call:  (Silvia) (10 mins)
Silvia has shared the HPC workpackage with the working group. The workpackage mainly includes:
- improving // efficiency
- improving single node performances
- real test cases

4.         Presentation: NEMO hybridization project at Intel, OpenMP parallelization (D. Kuts, title to be confirmed)
See slides attached.
Development work started from nemo v3_4_STABLE and the hybrid branch (MPI+OMP) developped by Silvia Mocavero
For future: needs longer inegration (more time steps) and a few different platforms, probably.
Discussion on wether it is useful to update from 3_4_STABLE to 3_6_STABLE or even to trunk?
Dmitri is willing to share his developments and results

5.       Discussion on single-core performance measurements and optimisation (M. Schreiber)
- Need to agree on how to measure performances
- how to cover all architectures?
- measure computation phase only, or include start up phase
- instumentationthrough timing
- single core performance
- cache
- portability
- possible performance tools: papi into timing function 
- detect bottlenecks
Document sent to the whole group for approval

6.       AoB (10 mins)
- Possible topics for next meetings?