This was a Master's project by the Operating Systems and Middleware Group at the Hasso-Plattner-Institute (HPI) in Potsdam, Germany. Our goal was to define and implement a parallel benchmark suite for heterogeneous many-core systems, based on the 13 Berkeley Dwarfs.

Three representative parallel algorithms were chosen and implemented in a GPU and a CPU variant:

• Matrix multiplication
• Quadratic Sieve
• Othello game simulation

Furthermore, a problem-specific scheduler was implemented for each of these algorithms, which uses MPI to distribute an input problem among several physical machines, and to distribute the workload among all CPUs and GPUs in each machine. So ultimately, every processor and graphics processor in a cluster are combined to solve a single problem. The run time of the whole calculation is measured and printed out at the end, possibly using the average of several runs.

Before the actual benchmark stage, each participating processor (GPU or CPU) is rated according to it's speed regarding the chosen algorithm. These ratings are used to distribute the workload accordingly - faster processors get bigger slices of the input data than slower processors, thus maximizing processor utilization in a heterogeneous environment.

This is a CMake based project, but it has only been compiled and tested in Linux environments so far. The CMake scripts assume GCC, but in principle, any compiler with sufficient C++ 11 support should work. In order to build and run the GPU implementations, a Fermi-class NVIDIA graphics card is required. Setting the BUILD_WITH_CUDA CMake option to OFF will disable building the CUDA implementations.

• GCC 4.7 or higher
• CUDA 5.0 or higher (optional)
• OpenMPI 1.6.4 or higher, with multithreading support
• Boost 1.49 or higher
• GMP 5.0.5 or higher, GMPXX C++ bindings

NOTE: The OpenMPI package provided in Debian's package management system is built without multithreading support and therefore is not compatible with this project. To run all scenarios successfully, you have to build MPI from source:

wget http://www.open-mpi.org/software/ompi/v1.6/downloads/openmpi-1.6.4.tar.bz2
tar -xvf openmpi-1.6.4.tar.bz2
cd openmpi-1.6.4
./configure --prefix=/usr/lib/openmpi --enable-mpi-thread-multiple
make -j
sudo make install
sudo cp --symbolic-link /usr/lib/openmpi/bin/* /usr/bin/
sudo cp --symbolic-link /usr/lib/openmpi/lib/* /usr/lib/

Either run build_all.sh, which will generate a Debug build by default, or follow these steps:

mkdir build
cd build
cmake ..
make

After building, you should run all the tests using test.sh or <BUILD_DIR>/src/test/test (if your build directory is something else than build).

A recorded shell session demonstrating some basic usage of the program can be found in doc/demo.ttyrecord. The ttyplay tool is required to play the demo.

An example MPI appfile (needed to benchmark a cluster) can be found in doc/example_appfile.

The source code is licensed under the MIT license.