Note: This is a major update from GraphMat v1.0 (single node and distributed). Please see changelog for details.
- Intel compiler (icpc) + Intel MPI (mpiicpc + mpi libraries)
(or)
-
GCC + MPICH (Other MPI libraries not tested)
-
Boost serialization library (links to libboost_serialization)
make
make MPICXX=mpic++ CXX=g++
Set the following environment variables:
export OMP_NUM_THREADS=[ number of cores in system ]
export KMP_AFFINITY=scatter
Use numactl for NUMA (multi-socket) systems if you are running 1 MPI rank on all the sockets e.g.
mpirun -np <NRANKS> numactl -i all bin/PageRank < graph file >
mpirun -np <NRANKS> numactl -i all bin/BFS < graph file > < start vertex >
GraphMat uses Catch, a C++ based testing framework.
git submodule init
git submodule update
make test
To run all the tests with a single MPI rank,
./testbin/test
Tests are also runnable in distributed mode with multiple ranks,
mpirun -np <NRANKS> ./testbin/test
You can also do ./testbin/test -? to list all the options available
See wiki page - https://github.com/narayanan2004/GraphMat/wiki/Reading-graph-files
If you use GraphMat in your work, please cite the following papers:
-
Narayanan Sundaram, Nadathur Satish, Md Mostofa Ali Patwary, Subramanya R Dulloor, Michael J. Anderson, Satya Gautam Vadlamudi, Dipankar Das, Pradeep Dubey, "GraphMat: High performance graph analytics made productive", Proceedings of VLDB 2015, volume 8, pages 1214 - 1225.
-
Michael J. Anderson, Narayanan Sundaram, Nadathur Satish, Md Mostofa Ali Patwary, Theodore L. Willke and Pradeep Dubey, "GraphPad: Optimized Graph Primitives for Parallel and Distributed Platforms," 2016 IEEE International Parallel and Distributed Processing Symposium (IPDPS), Chicago, IL, USA, 2016, pp. 313-322.
Paper URL:
- www.vldb.org/pvldb/vol8/p1214-sundaram.pdf
- http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=7516027
More documentation coming soon. For questions, please email narayanan.sundaram@intel.com