This repository contains benchs for evaluating performance of some message passing patterns.

To get sources files for the benchs:

git clone https://github.com/arcaneframework/arcane-benchs

You need to compile and install Arcane in Release configuration following the instructions in the github repository. Version 3.5.7+ of Arcane is required. We suppose Arcane will be installed in the directory ${ARCANE_INSTALL_PATH}. Instead of using a separate distribution of Arcane You can use the sources of Arcane provided in this repository if you use git submodules:

git submodule update --init --recursive
cmake -S framework -B /tmp/arcane_build_path -DCMAKE_INSTALL_PREFIX=${ARCANE_INSTALL_PATH}
cmake --build /tmp/arcane_build_path --target install

A recent (3.19+) version of CMake is needed to compile. In-source compilation is forbidden. You can configure and compile the benchs using the following command:

# To configure if you are in the source directory
cmake -S . -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH=${ARCANE_INSTALL_PATH} -B build_bench
# To compile
cmake --build build_bench

Two benchs are availables:

• MicroHydro for load balance
• MaHyCo for synchronisation

The benchs are launched by shell scripts generated in the build directory. You can specifiy the environment variable MPI_ARGS if you want to add some arguments to the MPI launch command.

It is important to note that by default only the processus with rank 0 will display information in the listing. You may set the environment variable ARCANE_PARALLEL_OUTPUT to 1 to allow all other ranks to print their corresponding listing in a file output%R where %R is the rank.

This bench is used to evaluate cell migration during load-balancing. During cell migration, a subdomain send cells and its associated variables to other subdomains. All the informations for a subdomain are sent in one message. In this benchmark, the size of one message is in the range 10Mo to 100Mo.

Two implementation are available:

• The default implementation use MPI_Irecv/MPI_Isend/MPI_Wait.
• A collective implementation using 'MPI_Alltoallv'. This implementation is available if you set the ARCANE_MESH_EXCHANGE_USE_COLLECTIVE environment variable to TRUE.

The important information is the time spent to send cell information. The following line in the listing gives the information:

*I-Mesh       ParallelExchanger Cell: ProcessExchange (end)
total_time=0.0058741569519043

You can launch the test and get the results with the following commands:

sh launch_micro_hydro_lb8.1.sh > result1.txt
grep 'total_time=' result1.txt | grep Cell

The results will be similar to this listing:

*I-Mesh       ParallelExchanger Cell: ProcessExchange (end) total_time=0.0058741569519043
*I-Mesh       ParallelExchanger Cell: ProcessExchange (end) total_time=0.00423097610473633
*I-Mesh       ParallelExchanger Cell: ProcessExchange (end) total_time=0.00581598281860352
*I-Mesh       ParallelExchanger Cell: ProcessExchange (end) total_time=0.00305461883544922
*I-Mesh       ParallelExchanger Cell: ProcessExchange (end) total_time=0.00515961647033691
*I-Mesh       ParallelExchanger Cell: ProcessExchange (end) total_time=0.00465488433837891
*I-Mesh       ParallelExchanger Cell: ProcessExchange (end) total_time=0.00334906578063965

There is two tests using 512 sub-domains:

launch_micro_hydro_lb512.1.sh
launch_micro_hydro_lb512.2.sh

If you want to test if anything is OK you can launch smaller tests:

launch_micro_hydro_lb32.1.sh
launch_micro_hydro_lb8.1.sh

This bench is used to evaluate weak scaling of synchronisations. In this configuration, a subdomain is the owner of approximately 18000 cells with 3 layers of ghost cells. The scripts are provided to launch the test from 128 subdomains to 2048. The important information is the total time of the execution loop. If the scalability of synchronisations is perfect, that time should be constant. You can launch a test and get the results with the following commands:

sh launch_mahyco_512.sh > result2.txt
grep '  loop:' result2.txt

The output is:

*I-Internal   TotalReel = 15.5567035675049 secondes (init: 0.841777086257935  loop: 14.7149264812469 )