Current Version: 1.0.1-BETA
Arabesque is a distributed graph mining system that enables quick and easy development of graph mining algorithms, while providing a scalable and efficient execution engine running on top of Hadoop.
Benefits of Arabesque:
- Simple and intuitive API, specially tailored for Graph Mining algorithms.
- Transparently handling of all complexities associated with these algorithms.
- Scalable to hundreds of workers.
- Efficient implementation: negligible overhead compared to equivalent centralized solutions.
Arabesque is open-source with the Apache 2.0 license.
- Linux/Mac with 64-bit JVM
- A functioning installation of Hadoop2 with MapReduce (local or in a cluster)
Arabesque currently takes as input graphs with the following formats:
- Graphs label on vertex(default)
# <num vertices> <num edges>
<vertex id> <vertex label> [<neighbour id1> <neighbour id2> ... <neighbour id n>]
<vertex id> <vertex label> [<neighbour id1> <neighbour id2> ... <neighbour id n>]
...
- Graphs label on edges To enable processing label on edges, in the yaml file, add the following lines
arabesque.graph.edge_labelled: true
arabesque.graph.multigraph: true # Set this to true if multiple edges
# exist between two vertices.
Input format
# <num vertices> <num edges>
<vertex id> <vertex label> [<neighbour id1> <edge label> <neighbour id2> <edge label>... ]
<vertex id> <vertex label> [<neighbour id1> <edge label> <neighbour id2> <edge label>... ]
...
Vertex ids are expected to be sequential integers between 0 and (total number of vertices - 1).
You can find an execution-helper script and several configuration files for the different algorithms under the scripts folder in the repository:
run_arabesque.sh
- Launcher for arabesque executions. Takes as parameters one or more yaml files describing the configuration of the execution to be run. Configurations are applied in sequence with configurations in subsequent yaml files overriding entries of previous ones.cluster.yaml
- File with configurations related to the cluster and, so, common to all algorithms: number of workers, number of threads per worker, number of partitions, etc.<algorithm>.yaml
- Files with configurations related to particular algorithm executions using as input the provided citeseer graph:fsm.yaml
- Run frequent subgraph mining over the citeseer graph.cliques.yaml
- Run clique finding over the citeseer graph.motifs.yaml
- Run motif counting over the citeseer graph.triangles.yaml
- Run triangle counting over the citeseer graph.
Steps:
- Compile Arabesque using
mvn package
You will find the jar file under target/
-
Copy the newly generated jar file, the
run_arabesque.sh
script and the desired yaml files onto a folder on a computer with access to an Hadoop cluster. -
Upload the input graph to HDFS. Sample graphs are under the
data
directory. Make sure you have initialized HDFS first.
hdfs dfs -put <input graph file> <destination graph file in HDFS>
-
Configure the
cluster.yaml
file with the desired number of containers, threads per container and other cluster-wide configurations. -
Configure the algorithm-specific yamls to reflect the HDFS location of your input graph as well as the parameters you want to use (max size for motifs and cliques or support for FSM).
-
Run your desired algorithm by executing:
./run_arabesque.sh cluster.yaml <algorithm>.yaml
-
Follow execution progress by checking the logs of the Hadoop containers.
-
Check any output (generated with calls to the
output
function) in the HDFS path indicated by theoutput_path
configuration entry.
The easiest way to get to code your own implementations on top of Arabesque is by forking our Arabesque Skeleton Project. You can do this via Github or manually by executing the following:
git clone https://github.com/Qatar-Computing-Research-Institute/Arabesque-Skeleton.git $PROJECT_PATH
cd $PROJECT_PATH
git remote rename origin upstream
git remote add origin $YOUR_REPO_URL