Skip to content

Latest commit

 

History

History
129 lines (100 loc) · 3.96 KB

README.md

File metadata and controls

129 lines (100 loc) · 3.96 KB
Raw

This Software is a test project to study the acyclic partitioning of DAG (of tasks), see the related paper at https://peerj.com/articles/cs-247/

The official repository is https://gitlab.inria.fr/bramas/dagpar/

Compilation

Simply go in the build dir and use cmake as usual:

mkdir build
cd build
cmake ..
make

If everything is fine, the tests should be correct:

make test

Adding a test file

Any .cpp file added in tests will be automatically compiled after using cmake .. and make in the build dir.

Where to make modifications

The partition method of the Graph class should do the partitioning. Anyone interested to improve the method should modify it, or duplicate an existing method and look at the obtained results.

Running the example

One can execute the main binary (from the tests/main.cpp file, compiled with make or make main). This file generates a graph and tries to partition it with different strategies. For each partition method a .dot file is generated and can be converted in to a pdf/png using the dot tool. Also, each execution emulation will give an execution trace in .svg.

## Help

$ ./main --help
[HELP] Asked for help.
[HELP] You can only pass the graph generation or a filename.
[HELP] $ ./main [generation method]
[HELP] Where generation method is among: tree, deptree, 2dgrid, doubletree, filename

## Example

Each tested method will create .dot/.svg files and inform about an emulated execution time.

$ ./main tree
[INFO] use tree
nbThreads : 2 / partMinSize : 4 / partMaxSize : 4
Number of nodes : 36
Degree of parallelism one the original graph : 8  4.88889
Generate pdf of the original graph with: dot -Tpdf /tmp/agraph-original.dot -o /tmp/agraph-original.pdf
Generate pdf of the graph with: dot -Tpdf /tmp/agraph-rand.dot -o /tmp/agraph-rand.pdf
Degree of parallelism after random partitioning : 2  1.11111
Number of partitions : 9 -- avg part size : 4
[EXECUTION] Ready tasks at starting time => 1
[EXECUTION] Total duration => 32
...

Will create

$ ls /tmp/*.svg /tmp/*.dot
/tmp/agraph-backtrack.dot   /tmp/agraph-rand.dot                 /tmp/depgraph-diamond.dot
/tmp/agraph-diamond.dot     /tmp/dep-graph-2trace-backtrack.svg  /tmp/depgraph.dot
/tmp/agraph.dot             /tmp/dep-graph-2trace-greedy.svg     /tmp/depgraph-greedy.dot
/tmp/agraph-greedy.dot      /tmp/dep-graph-2trace-rand.svg       /tmp/depgraph-horizontal.dot
/tmp/agraph-horizontal.dot  /tmp/dep-graph-2trace.svg
/tmp/agraph-original.dot    /tmp/depgraph-backtrack.dot

Converting the .dot can be done one by one using the commands given in the output of the execution, or for all files at once with (be aware that if the graphes are huge this might take forever)

for fl in /tmp/*.dot ; do dot -Tpdf $fl -o $fl.pdf ; done

** Note that the files called depgraph-... represent the new clustered graphes. **

Currently, the size of the clustering and the number of threads in the emulated executions must be changed in the code:

    const int nbThreads = 2;
    const int partMaxSize = 4;

## Chukrut example

The main binary can execute .txt and .dot files (it will simply looks at the dependency in the .dot). Therefore a Chukrut data could be run with:

./main ../data/graph-140721493545840.dot

Install Metis

wget http://glaros.dtc.umn.edu/gkhome/fetch/sw/metis/metis-5.1.0.tar.gz
tar -xvf metis-5.1.0.tar.gz
rm metis-5.1.0.tar.gz
cd metis-5.1.0
make config prefix=$(pwd)/install
make
make install
export METIS_DIR=$(pwd)/install

# Then
cmake .. -DMETIS_DIR=$METIS_DIR

Install Scotch

wget https://gforge.inria.fr/frs/download.php/file/37622/scotch_6.0.6.tar.gz
tar -xvf scotch_6.0.6.tar.gz
rm scotch_6.0.6.tar.gz
cd scotch_6.0.6/src
cp Make.inc/Makefile.inc.x86-64_pc_linux2 ./Makefile.inc
make
make prefix=$(pwd)/install install
export SCOTCH_DIR=$(pwd)/install

# Then
cmake .. -DMETIS_DIR=$METIS_DIR -DSCOTCH_DIR=$SCOTCH_DIR -DUSE_SCOTCH_AS_METIS=ON