Multi-GPU MCMC Implementation

Implementation of Metropolis-MCMC using multi-GPUs approach on OpenACC as part of the MSc HPC with Data Science Dissertation in EPCC.

Requirements:

• OpenBLAS To download and install:

$  git clone https://github.com/xianyi/OpenBLAS.git
$  cd OpenBLAS
$  make USE_THREAD=0 USE_OPENMP=0
$  make install PREFIX=your_installation_directory

Build and Run on CIRRUS:

Copy the config file in root dir:

$ cp config/linux-pgi-default.mk ./config.mk

Remember to specify the paths in the files

Load PGI compilers:

$ module load pgi

For serial and threaded versions you will need to build mpi_s:

$ cd mpi_s
$ make clean
$ make libc

For MPI Runs you will need mpt:

$ module load mpt

For GPU Runs you will need cuda:

$ module load cuda

To compile the code:

$ cd src
$ make verbose

This will display the options and select one of the options.

To run serial or threaded executable:

$ cd src
$ ./mcmc.exe input.ref

This will run the executable with the contents of input.ref. You can change the threads from there.

To run mpi executable submit in a pbs:

$ cd src
$ mpiexec_mpt -ppn N -n M ./mcmc.exe input.ref

or for hybrid

$ cd src
$ mpiexec_mpt -ppn N -n M omplace -nt K ./mcmc.exe input.ref

Remember to load the necessary modules.

To run on GPUs compile one of the gpu options and submit on the gpu node