If you intend to run mSWEEP on the machine used in compiling the source code, you might want to add the '-march=native -mtune=native' flags if compiling with GCC by running
> cmake -DCMAKE_CXX_FLAGS="-march=native -mtune=native" -DCMAKE_C_FLAGS="-march=native -mtune=native" ..
Using these options significantly reduces the runtime of mSWEEP in some environments (e.g. most HPC setups).
mSWEEP can be compiled with link time optimization. This can improve performance on large problems and can be enabled with
> cmake -DCMAKE_BUILD_WITH_FLTO=1 ..
The Intel C++ compiler can be enabled by compiling mSWEEP with
> cmake -DCMAKE_C_COMPILER=icc -DCMAKE_CXX_COMPILER=icpc ..
mSWEEP can be compiled with MPI support, distributing the mixture component estimation part of the program to several processes. To compile with MPI support, set your environment appropriately and build mSWEEP with the following commands:
> mkdir build
> cd build
> module load mpi/openmpi
> cmake -DCMAKE_ENABLE_MPI_SUPPORT=1 -DCMAKE_C_COMPILER=mpicc -DCMAKE_CXX_COMPILER=mpicxx ..
> make
The CMAKE_ENABLE_MPI_SUPPORT flag configures the project accordingly.
Using mSWEEP through MPI will result in increased memory usage.
Distribute computation to 4 processes by calling mSWEEP with:
mpirun -np 4 mSWEEP --themisto-1 fwd.txt --themisto-2 rev.txt -i cluster_indicators.txt
Enable hybrid parallellization with multiple threads per process through use of the -t flag:
mpirun -np 2 mSWEEP --themisto-1 forward_aln.gz --themisto-2 reverse_aln.gz -i cluster_indicators.txt -t 2
Hybrid parallelization might require binding the processes (refer to your HPC documentation on how to do this). The optimal configuration between ranks and threads will depend on the size and structure of your data.
When mSWEEP is called through MPI, the root process will handle all read and write operations and only the estimation part is distributed.