Plumed Flagship meeting: Lecture: Parallel and GPUs programming in PLUMED

We have some simple examples for using parallelism in plumed:

• SerialCoordination.cpp is a vandalized simplified version of the cv Coordination
• OMPCoordination.cpp is an example on how openMP can be used in Plumed
• MPICoordination.cpp is an example on how MPI can be used in Plumed
• CUDACoordination.cpp is an example on trying to user Cuda for solving this problem (it needs an ad hoc compiler queue and the kernel CUDACoordinationkernel.cu )

(The original Coordination combines the use of both openMP and MPI)

Plumed helps the developer with some tools for parallelism:

• tools/OpenMP.h contains some function that are useful in working with openMP. In the example we are using OpenMP::getNumThreads() to get the number of threads from the environmental variable PLUMED_NUM_THREADS

• tools/Communicator.h is present as the variable comm that is inherited through PLMD::Action. PLMD::Communicator is an interface to some of the functionalities of the C API of mpi.h. In the example we are using PLMD::Communicator::Get_size() to get the number of the processes spawned by mpirun, PLMD::Communicator::Get_rank() to get the id of the process, and PLMD::Communicator::Sum() to sum the result of the coordination and make the correct value avayable for further calculations.

Intro

The exercise uses Base.hpp to give a very base version of the COORDINATION CV, in this case it is returning the sum of the number of atoms that are within R_0 from each atom. For simplicity it the pbcs will be ignored throught all of the examples. MyCoordinationBase in Base.hpp do not have the calculate method and so the example will have more or less the following structure:

#include "Base.hpp"
namespace PLMD {
class MyCoordination : public MyCoordinationBase {
public:
  explicit MyCoordination(const ActionOptions &ao)
      : Action(ao), MyCoordinationBase(ao) {}
  ~MyCoordination() = default;
  // active methods:
  void calculate() override;
};
PLUMED_REGISTER_ACTION(MyCoordination, "MYCOORDINATION")
void MyCoordination::calculate() {
...code goes here...
}
} // namespace PLMD

For automation purposes (see the Makefile) we are using the same key PLUMED_REGISTER_ACTION(MyCoordination, "MYCOORDINATION") for all of the examples.

Prerequisites

• Plumed 2.9.0 configured and installed with --enable-modules=all and MPI
  • in my workstation I am using gcc 9.4.0 and openmpi 4.1.1
• For the GPU offloading the example is written with Nvdia’s Cuda
  • In my workstation I am using cuda 11.7 with a T1000 card

The serial code

Threading: openMP

Processes: MPI

EXTRA: GPU offloading with Cuda

Closing information

The file in the Solution directory give the correct results