BabelFlow is an Embedded Domain Specific Language to describe algorithms using a task graph abstraction which allows them to be executed on top of one of several available runtime systems.
The code is released under BSD 3 licence (attached) and currently includes example use cases using MPI and Charm++ runtime systems.
The project can be build using CMake as following
mkdir build
cd build
cmake ..
make
The default CMake configuration will build the BabelFlow using the MPI runtime.
To change runtime you can modify the field RUNTIME_TYPE in the CMake
configuration to MPI, CHARM or LEGION.
To build and run your application using the Charm++ runtime you need to provide the Charm++ compiler path setting the field CHARM_COMPILER as follow:
ccmake .. -DRUNTIME_TYPE=CHARM -DCHARM_COMPILER=/path/to/charmc
To build and run your application using the Legion runtime you need to provide the Legion cmake configuration folder as follow:
ccmake .. -DRUNTIME_TYPE=LEGION -DLegion_DIR=/path/to/legion/install/share/Legion/cmake
If your Legion build is using GASNet you should also provide the path to it adding the following flag to your ccmake:
-DGASNet_INCLUDE_DIR=/path/to/gasnet/release/include
Reduction
The reduction graph example executes a k-way reduction dataflow. The tasks defined in this examples perform two simple operations: sum and print (the final result).
The executable requires 2 inputs: <nr-of-leafs> <fan_in>
To execute the application using the MPI runtime you can run (using for example 8 cores):
mpirun -np 8 examples/reduction/mpi/reduction 16 2
To execute the version built using the Charm++ runtime you can run (8-way SMP):
examples/reduction/charm/reduction +p8 16 2
Note that the size of the graph (e.g., number of leafs) is not bounded by the number of cores you are using. For example we are using 8 ranks to execute a reduction graph with 16 leafs.
This means that you can easily debug bigger scale runs on a reduced number of computing resources.
Broadcast
The broadcast graph example executes a k-way broadcast dataflow. This example defines only one task which computes the sum of the numbers from 0 to 99 and compares the result with the input data which has been propagated through the broadcast dataflow.
The executable requires 2 inputs: <nr-of-leafs> <fan_out>
To execute the application using the MPI runtime you can run (using for example 8 cores):
mpirun -np 8 examples/broadcast/mpi/broadcast 16 2
To execute the application using the Charm++ runtime you can run (8-way SMP):
examples/broadcast/charm/broadcast +p8 16 2