gcvb

(G)enerate (C)ompute (V)alidate (B)enchmark

gcvb is a Python 3 module aiming at facilitating non-regression, validation and benchmarking of simulation codes.

gcvb is not a complete tool of continuous integration (CI). It is rather a component of the testing part of a CI workflow. It can compare the different metrics of your computation with references that can be a file, depends of the "configuration" or are absolute.

Install gcvb

Clone this repository and in the directory run :

pip3 install .

or without git:

pip install --upgrade "gcvb[dashboard] @ https://github.com/jm-cc/gcvb/archive/master.zip"

Getting started

As of today, the documentation is mainly available under the form of examples in the gcvb-example repository.

General philosophy

Input :

• The "data" folder contains the input files and references.
• A config.yaml file contains relevant information for the current computing environment used.
• Multiple yaml files can then be use to indicate which tests to launch, store the references to detect a regression.

Output :

• a gcvb.db sqlite3 database that contains the metrics stored during runs, but also files you want to keep for each run.
• a "results" folder where the computation are actually computed.

Process :

• Given the input, the user can generate a base, that base can then be used for multiple runs. Each run compares with the reference and allows to check that everything is in order (or not). A run can launch a subset of the test of a base through filter options.

gcvb is mainly used with subcommands. It must be launch in the folder containing the input.

To access the help, just use gcvb -h. Help is also available for each subcommands (e.g. gcvb generate -h).

Jobrunner

When launching computation with compute, by default a script is submitted. It is also possible to launch a jobrunner that computes job in parallel. (gcvb compute --with-jobrunner <num_cores>)

Multiple jobrunner can be launched through gcvb jobrunner <num_cores>.

Note that there is no difference between a jobrunner launched through the compute options and by the jobrunner command. It is possible to create the right entries in the database without launching a jobrunner or submit the default launch script with the option --dry-run

Copyright and license

Copyright 2019 Airbus S.A.S

Code released under the MIT License.