Skip to content

Replication package for empirical evaluation on d-DNNFs for feature-model counting

License

Notifications You must be signed in to change notification settings

SoftVarE-Group/exploiting-ddnnfs-eval

Repository files navigation

Empirical Evaluation on Exploitation of d-DNNFs

Replication package for empirical evaluation that examines the efficiency of reusing d-DNNFs for repetitive counting queries on feature models.

General information

Important note: Due to githubs 100Mb per file limit, not all d-DNNFs are part of the repository, but they are included in the docker image we provide. If you want to run the benchmark without using the image, you have to first run the compiler. Additionally, you have to run the preprocessing run configuration if you want to run dsharp related configurations.

The python benchmark script can be used as it comes. Nevertheless, we highly recommend using the provided docker image because it is easier to use and for our results we also used docker. Furthermore, in solvers/ the tools are provided as built binaries and the different run configurations are in run_configurations/. We divide those in configurations that include confidential models, those that exclude confidential models, and tests configuration. The tests holds the same configurations as the other two, but only considers one model (i.e. Smarch/toybox_0_7_5), runs that model only once, and the output of the tools gets printed onto stdout.

How to run

For both ways you either have to specify a run configuration (.json or .yaml) file, respectively a folder that contains one or multiple run configurations or use the provided configurations. We provide exclusively .yaml files.

Run benchmark via Docker

First, if not done already you have to install docker. You can find information on how to build Docker here.

Second, we have to pull the image using docker. Note that it might be necessary (depending on your setup) to use sudo for Docker.

docker pull ghcr.io/softvare-group/reuse-ddnnf-eval-february24:latest

Now we are ready to run a container with one or multiple run configurations. For instance, the following example starts a docker container from the image rp:nc-latest, with the name test, allocates a tty for the container process (-it flags), and runs all run configurations that are included in the test directory and its subdirectories.

docker run --name test -it ghcr.io/softvare-group/reuse-ddnnf-eval-february24:latest run_configurations/test

Another example could be the following command which differs from the previous by running the configuration for the cardinality of features for reasoners based on the d-DNNFs of c2d.

docker run --name test -it ghcr.io/softvare-group/reuse-ddnnf-eval-february24:latest run_configurations/non_confidential/reasoner/features_c2d.yaml

You can get the result data by copying it from the container onto your system. Therefore, you can use the following command which accesses the test container and copies the results directory into the DET_PATH_ON_HOST. This can simply be the current working directory (i.e. ./).

docker cp test:benchmark/results DEST_PATH_ON_HOST

Alternative way to run the benchmark (not recommended)

In general, an experiment specified by a .json or a .yaml file can be executed with Python directly. Nonetheless, we recommend using docker because when using python you have to make sure to install all dependencies on your own and meet all requirements as mentioned in the dockerfile.

For instance, the following command executes all run configurations in the test directory and its subdirectories.

python3 run.py run_configurations/test