Next: A. The grid types Up: 6 Compiling, running, debugging, Previous: 2 Time statistics files
An efficient use of the allocated computing resources in a coupled system requires the harmonisation of the component execution speed. This operation, called load balancing, is often neglected, either because of the apparent resource abundance or practical difficulties. To facilitate this work, a load balancing analysis functionality is included in OASIS3-MCT and can be activated by setting to 1 the third number under $NLOGPRT in the namcouple configuration file (see section 3.2). Some details on this functionality are provided here and more information can be found in the balancing_documentation.pdf file in oasis3-mct/util/load_balancing directory.
When activated, the load balancing analysis functionality outputs the full timeline of all OASIS3-MCT related events, for any of the allocated resources. This timeline is saved in one NetCDF file per coupled component, timeline_XXX_component.nc where XXX is the component name. It provides the comprehensive sequence of all operations related to the coupling (field send and receive through MPI, field output on disk, field interpolation and mapping, field reading on disk, restart writing, initialisation and termination phase of the OASIS3-MCT setup) so that any simulation slow down in link with the use of the OASIS3-MCT library can be identified.
The analysis of the coupling field exchanges, amongst all coupling events, allows to not only identify the waste of resources by components which are recurrently waiting for their coupling fields but it also reveals other bottlenecks such as disk access or model internal load imbalance. The full picture of these events makes possible an optimal load balancing, even for the most complex configurations.
In addition to the detailed timeline saved in the NetCDF file, more general computing information (simulation time, speed, waiting time, etc.) is also provided in a text file load_balancing_info.txt for the coupled model and for each component. In simple cases, this global information can help to allocate resources in a balanced way.