Version 1 (modified by mocavero, 2 years ago) (diff)

NEMO HPC subgroup: Fri 20 April 2018

Attendees: Claire Levy (IPSL), Sebastien Masson (IPSL), Mike Bell (Met Office), Mirek Andrejczuk (Met Office), Matthew Glover (Met Office), Andy Porter (STFC), Miguel Castrillo (BSC), Mario Acosta (BSC), Erwan Raffin (Bull), Silvia Mocavero (CMCC)

Apologies: Marie-Alice Foujols, Clement Bricaud

1. Work on OpenMP directives @MetO

Mirek presented the results of the OpenMP parallelisation (at loop level) developed and tested on the GO6 configuration. The results confirm that the hybrid version does not improve the pure MPI version. (see attached slides)

2. Hybrid parallelization, single-core performance and XIOS reading functionality tests @CMCC

Silvia presented the recent results achieved through the application of the hybrid coarse-grained approach to the NEMO tracers advection. An improvement of ~7% on the parallel efficiency compared with the pure MPI version has been achieved. The limited improvement of both the fine-grained and coarse-grained versions has been discussed. A detailed analysis of the current version of the NEMO code should be performed in order to understand which are the main bottlenecks (MPI overhead, memory access, …), how they are correlated and what we expect from the optimisation strategies. Silvia also presented some results about the integration of the XIOS reading functionality, tested on the GLOB16 CMCC configuration. Results show an improvement of both the I/O time variability and the execution time of about 20%. Finally, some tests on the optimisation of the single-core performance by applying the cache-blocking and vectorisation techniques have been presented. Cache-blocking results depend on the domain size, the block size, the compiler version and optimisation level. A methodology to theoretically estimate the code behaviour should be formulated. (see attached slides)

3. Next meeting call

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