Getting Started with MPI

Objectives

We would like to :

Understand the basics of MPI
Setup our MPI Environment on euler
Run our first hello world program.

Notes

Discuss MPI
- MPI stands for the message passing interface.
- It defines a library of routines for performing distributed computation.
- Primarily c and FORTRAN, but works with C++, C#, Java and Python, R, ruby, ....
MPI is not a substitution for OpenMP but a different layer.
- OpenMP code runs on a multiprcoessor
- MPI code runs on a network.
- The two form the backbone for distributed parallel computation.
MPI runs on distributed systems
- On a laptop, but why?
- On a SMP, but I would use OpenMP
- On a Beowulf Cluster
- On a supercompueter like Bridges-2 at PSC
- On El Capitain the fastest scientific computer as of 11/24
MPI is network based
- Probably built on sockets.
- But can incorporate other high speed networks protocols.
The book says that people call mpi "the assembly language of parallel programming"
- There are a large number of routines with very explicit control over how they are run.
- This is because:
  - It runs everywhere, as above.
  - It is still adaptable to bleading edge technology.
  - And it runs on the top scientific computing machines.
    - Real time on these machines is hard to come by.
    - And they expect optimized code.
    - And they provide expert assistance in writing this code.
- and writing parallel code at this level is difficult.
Frequently it runs in a batch environment
- Again, processing time at supercomputer centers is expensive.
There are multiple implementations of mpi.
- MPICH was developed at ANL.
- Open MPI
- Most hardware vendors have their own versions, tuned to their hardware.
Discuss MPICH
- Wikipedia states "MPICH is one of the most popular implementations of MPI. It is used as the foundation for the vast majority of MPI implementations, including IBM MPI (for Blue Gene), Intel MPI, Cray MPI, Microsoft MPI, Myricom MPI, OSU MVAPICH/MVAPICH2, and many others."
Setting up your environment
- On fedora, mpi setup is controlled by the environment module system.
  - There are multiple implementations of MPI
  - You need to select the one you want.
- Using the module system:
  - If it is not setup for use by default.
    - source /etc/profile.d/modules.sh
    - At the command line
    - Running this once seems to permanently set up the module system in your environment.
- Then load the mpi module you will be using.
  - module load mpi/mpich-x86_64
- This will
  - modify your path to include the mpicc binaries
  - Modify your environment to search for libraries in the right place.
  - printenv | grep MPI after you run this.

Hello world.

A standard c++ program but

#include <mpi.h>
include argc, and argv.
Initialize MPI first.
- MPI_Init(&argc, & argv);

Get the size of the computation

int size;
MPI_Comm_size(MPI_COMM_WORLD, &size);

Get the rank of the process (ie process id)

int rank;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);

These can now be used to identify what we want the process to do .

cout << "Hello world from " << rank << " of " << size << "!" << endl;

Finally close down the MPI process
- ```
MPI_Finalize();
```
The full code:

#include 
#include 

using namespace std;

int main(int argc, char * argv[]) {
    MPI_Init(&argc, &argv);

    int size;
    MPI_Comm_size(MPI_COMM_WORLD, &size);

    int rank;
    MPI_Comm_rank(MPI_COMM_WORLD, &rank);

    cout << "Hello world from " << rank << " of " << size << "!" << endl;

    MPI_Finalize();
    return 0;
}

For your makefile
- You want to redefine the C++ compiler
  - CXX = mpic++
  - This could also be mpicxx
  - Makefile
Run you code
- mpiexec -n 5 hello
- This could also be mpirun
- -n indicates the number of processors
- This is a complex command, but -n will do for now.