application-user/maswa

Accelerated algorithms for Multichannel Analysis of Surface Waves (MASW) using MPI and CUDA

MASWA

Accelerated algorithms for Multichannel Analysis of Surface Waves (MASW) inversion using MPI and CUDA.

Separate Versions:

Currently the MPI and Cuda implementations are two separate algorithms, so the user chooses which form of parallelization best suites their resources and uses that version of MASWA (MPI if they have many CPU cores, Cuda if they have a GPU). The header and c files for the MPI version are in include and src, respectively, while the files for the Cuda version are in include_cuda and src_cuda.

Both versions are compiled using the compilation.py script, written in Python. Both versions mostly use C code, with some C++ functionality to handle complex numbers. The Cuda version obviously uses Cuda for GPU kernels.

Requirements:

Although both versions use the same compilation script, additional requirements for one version are not necessary to run the other. For example, a GPU is not needed at all to compile and run the MPI algorithm, while no version of Open MPI or Open RTE is needed to run the Cuda version.

Requirements for both versions:

• a CPU with one or more cores
• python 3 to run the compilation.py script (it was tested with Python 3.7.3)
• g++ compiler (tested and proven to work on: GCC version 5.2.0, Ubuntu version 7.4.0, and Apple LLVM version 10.0.1)
• some common C libraries: stdio.h, stdlib.h, string.h, math.h, and time.h
• the c++ complex library for complex numbers

Additional requirements for MPI version:

• Open MPI (versions that have been tested are 1.8.5, 2.1.1, and 3.1.3)
• alternatively, Open RTE also works for running the MPI code (tested versions are 2.1.1 and 3.1.3)
• ideally, as many CPU cores as processes you wish to use for running the code

Additional requirements for Cuda version:

• an Nvidia GPU (the code was tested on an Nvidia Quadro K620, and is designed to work with older GPUs as well as newer ones)
• the nvcc Cuda compiler driver (tested on V9.1.85)
• the cuComplex library in Cuda

Running MASWA

MPI version:

To compile MPI MASWA, first make sure all of the required software is downloaded. Then enter this command:

python compilation.py mpi

to compile the mpi code. From there you can run it with this command is using Open MPI:

mpirun -np <number of processes> ./maswa_mpi

If using Open RTE instead, then replace mpirun with mpiexec. To delete the .o files generated by compilation, run:

python compilation.py mpi_clean

Currently the particular dispersion curve data and model you wish to run must be set up in the main function. A future enhancement will be to add file IO and command line arguments to the executable for usability.

Cuda version:

To compile Cuda MASWA, first make sure all of the required software is downloaded. Then enter this command:

python compilation.py cuda

to compile the cuda code. From there you can run it with this command:

./maswa_cuda

To delete the .o files generated by compilation, run:

python compilation.py cuda_clean

Like the MPI version, the Cuda version requires the desired tests and data to be placed in the main function. It is possible the Cuda tags in compilation.py may not work for your version of Cuda, so they may need to be changed.

Citation:

This code was based off of MASWaves, an implementation of MASW written in Matlab.

Ólafsdóttir, Elín Ásta & Erlingsson, Sigurdur & Bessason, Bjarni. (2018). OPEN SOFTWARE FOR ANALYSIS OF MASW DATA.