| 60 | | '''NEMO benchmark''' |
| 61 | | |
| 62 | | ''Description'': development of a benchmark (idealised geometry, equal MPI sub-domain size, realistic parameters, parametrisable communications, internal timings ...) at different resolutions to identify the bottlenecks to NEMO scalability. |
| 63 | | |
| 64 | | ''Involved people'': Sebastien Masson (CNRS/ST), Eric Maissonave (CERFACS), Louis Douriez (ATOS), David Guibert (ATOS), Erwan Raffin (ATOS) |
| 65 | | |
| 66 | | ''WP action'': HPC09_ESIWACE? |
| 67 | | |
| 68 | | '''Intra-node performance''' |
| 69 | | |
| 70 | | ''Description'': analysis and improvement of NEMO intra-node performance to move real towards peak performance and to reduce intra-node communications. Analysis at kernel level to understand which kernels are limiting performance; improving utilisation of memory hierarchy; loop vectorisation. Investigation and integration of a second level of parallelism based on the shared memory paradigm. Use of OpenMP (fine and coarse-grained parallelisation). |
| 71 | | |
| 72 | | ''Involved people'': Silvia Mocavero (CMCC/ST), Francesca Mele (CMCC/ST), Mike Bell (MetO), Miroslaw Andrejczuk (MetO/ST), Matt Glover (MetO) |
| 73 | | |
| 74 | | ''WP actions'': WP2018-01_Silvia Mocavero_singlecoreperf, WP2018-02_Francesca_Mele_hybrid |
| 75 | | |
| 76 | | |
| 77 | | '''Inter-node communications''' |
| 78 | | |
| 79 | | ''Description'': limiting the communication overhead through the reduction of the MPI communications number and frequency, the investigation of advanced MPI communications and techniques to overlap communication and computation. |
| 80 | | |
| 81 | | ''Involved people'': Silvia Mocavero (CMCC/ST) |
| 82 | | |
| 83 | | ''WP actions'': WP2018-03_Silvia Mocavero_globcomm, WP2018-04_Silvia Mocavero_mpi3, WP2018-05_AndrewC-extendedhaloes |
| 84 | | |
| 85 | | |
| 86 | | '''Separation of concerns''' |
| 87 | | |
| 88 | | ''Description'': investigation of a light-DSL approach to apply DSL in NEMO without impacting on the NEMO coding structure rules. The approach consists in processing the NEMO code to create an internal representation compliant to the PSyclone tool and manipulating the intermediate code to perform PSyclone transformations. |
| 89 | | |
| 90 | | ''Involved people'': Andrew Porter (STFC) |
| 91 | | |
| 92 | | ''WP action'': - |
| 93 | | |
| 94 | | |
| 95 | | '''I/O''' |
| 96 | | |
| 97 | | ''Description'': Investigation of techniques to reduce the I/O overhead which increases with the number of processors. Extension of the use of XIOS for reading/writing restart files. |
| 98 | | |
| 99 | | ''Involved people'': Miroslaw Andrejczuk (Met Office/ST) |
| 100 | | |
| 101 | | ''WP actions'': WP2018-06_andmirek-XIOSread, WP2018-07_andmirek_XIOSwrite |
| 102 | | |
| 103 | | |
| 104 | | '''Mixed-precision''' |
| 105 | | |
| 106 | | ''Description'': optimization of the NEMO model by using a mixed precision approach. Study on the precision needed by the different processes in NEMO. Implementation of the mixed precision approach in NEMO. |
| 107 | | |
| 108 | | ''Involved people'': Miguel Castrillo (BSC), Oriol Tinto (BSC) |
| 109 | | |
| 110 | | ''WP action'': WP2018-08_Mixed_precision |
| 111 | | |
| 112 | | |
| 113 | | |
