⚗️ FPMAS MetaModel

The purpose of the MetaModel is to provide a test bench to estimate load balancing and synchronization modes performances provided by FPMAS on different kinds of models, easily generated from a few configuration fields in a YAML file.

Several types of base graphs are provided by FPMAS:

• Grids
• Uniform random graphs
• Clustered graphs
• Small-World networks

Those graphs can easily be configured with a provided node count and an average outgoing neighbors count. Each graph can then be used to define a pure graph model, or as a spatial environment on which agents are randomly moving.

Agents can move uniformly, or according to an utility value assigned to each cell, in order to define spatial models with a non-uniform agent distribution, notably to test the behavior of each load balancing algorithm in this case.

Build

Ensures that the FPMAS platform is properly installed. An additional -DCMAKE_PREFIX_PATH=custom/installation/path/ might be specified to cmake if FPMAS is not installed in a standard directory (the installation path corresponds to the -DCMAKE_INSTALL_PREFIX eventually specified when building FPMAS).

git clone https://github.com/FPMAS/fpmas-metamodel
cd fpmas-metamodel
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make

Configuration

The MetaModel can easily be configured with many parameters in order to generate many kinds of models.

See the documented config.yml file, the documentation of the config.h file, the MetaModel doxygen documentation and the FPMAS wiki for more information about the MetaModel configuration fields.

Useless fields can be safely removed from the YML configuration file, so that the default value specified in the original config.yml configuration file is used. Some basic examples are provided in the examples directory.

Run

The model can be run with the following command:

mpiexec -n <N> ./fpmas-metamodel <config_file> [-s seed]

• N: processes count
• config_file: a .yml configuration file
• seed (optional): a random seed

The ./gen-seed N utility command can also be used to deterministically generate a set of N seeds that can be passed to the model.

Output

The MetaModel can generate several (and complex outputs).

CSV

The main and default MetaModel output consists in one CSV file by test case and by process, named <lb_algorithm>-<lb_period>.<process_rank>.csv, with the following fields. See the MetaModelCsvOutput class documentation for the description of each field.

The external analysis of the output data is a complex project on its own, that is not detailed here and not handled within this project.

JSON

The json_output and the json_output_period parameters can be used to generate some JSON outputs in order to perform external visualisation. The JSON output is currently only supported for GRID based environment.

When enabled, several .json files are produced:

• <lb_algorithm>-<lb_period>_cells.<time_step>.json: see CellsLocationOutput
• <lb_algorithm>-<lb_period>_agents.<time_step>.json: see AgentsOutput

DOT

Finally, a DOT can be produced at the end of the simulation in the file <lb_algorithm>-<lb_period>.<time_step>.dot, that can be rendered externally using the Graphviz tools. The size of the output can grow very fast as the size of the models grows, so such output is only relevant for relatively small models.

Examples of DOT rendering with the fdp tool for example configurations in examples/dot are presented below. See the graphviz manual pages to configure the rendering tools, or directly modify the generated .dot files.

small_world.yml	grid.yml

• Diamonds represent Cells
• Circles represent Agents
• The opacity of Cells is proportionnal to their utility
• Blue links represent the Cell network
• Red links represent location links
• Green links represent interactions

With the grid.yml configuration, that defines a grid_attractor at (2, 2), we can see that agent gather around those coordinates. It is the purpose of the MetaModel to easily produce those mechanics so the behavior of load balancing algorithms can be studied.

Graph stats

The fpmas-metamodel-graph-stats tool can be used to easily generate different environment to dump them to DOT files and to compute some graph statistics such as the average shortest path length, clustering coefficient, centrality or size of the biggest connected components.

mpiexec -n <N> ./fpmas-metamodel-graph-stats -i -s <seed> <graph_stats_config.yml>

• N: processes count
• graph_stats_config: a .yml configuration file. See graph_stats_config.yml for a simple example.
• seed (optional): a random seed that can be used to generate different graphs with the same FPMAS graph generators configuration.
• -i: if specified, only the initialisation and the DOT output of each graph is performed, without requiring FPMAS to output graph statistics.

Currently, FPMAS features about graph statistics are relatively experimental, so we recommend to use the -i option. The excellent Python graph-tool can be used instead to compute many statistics from the generated DOT files.

Contact

For more information about this model or its implementation, please contact:

• Paul Breugnot (developer): [email protected]
• Laurent Philippe (supervisor): [email protected]
• Bénédicte Herrmann (supervisor): [email protected]
• Christophe Lang (supervisor): [email protected]