# MPI: an introduction <div class="center"> [Introductory notes](/notes/mpi/) </div>

## Main idea - _Processes_ (not threads), with no shared memory - Explicit communication between processes - No "special" process

## Picture Threads share memory  Processes don't 

- Most [MPI](https://www.mpi-forum.org/) programs are [single program, multiple data](https://en.wikipedia.org/wiki/SPMD) (SPMD) - Every process gets its own copy of _the same_ program - ... but each process operates on _its own_ data

## Hello world ```c [8,12|9,10|11] #include <stdio.h> #include <mpi.h> int main(int argc, char **argv) { int rank, size, len; MPI_Comm comm; MPI_Init(&argv, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &rank); MPI_Comm_size(MPI_COMM_WORLD, &size); printf("Hello, World! [%d/%d]\n", rank, size); MPI_Finalize(); return 0; } ```

## On two processes <div class="container"> <div class="col"> ### Process zero ```c [6,7] int main(int argc, char **argv) { int rank, size, len; MPI_Comm comm; MPI_Init(&argv, &argv); rank = 0; size = 2; printf("Hello, World! [%d/%d]\n", rank, size); MPI_Finalize(); return 0; } ``` </div> <div class="col"> ### Process one ```c [6,7] int main(int argc, char **argv) { int rank, size, len; MPI_Comm comm; MPI_Init(&argv, &argv); rank = 1; size = 2; printf("Hello, World! [%d/%d]\n", rank, size); MPI_Finalize(); return 0; } ``` </div> </div>

## Communicators - Messages are passed in context of a [_communicator_](/notes/mpi/#communicators) - Represent a group of processes - Processes identified by `(communicator, rank)` pair - Communicators available on startup 1. `MPI_COMM_WORLD`: All processes 2. `MPI_COMM_SELF`: Every process "on its own"

## Example with eight processes 

## Programming - MPI is a _library_. No special compiler support - Need to `#include <mpi.h>` - Large API, we will touch on some of it

## Compiling - Most MPI libraries come with _compiler wrappers_ - `mpicc` (for C) - `mpicxx` (for C++) - `mpif90` (for Fortran) - Take care of include and link flags ### On Hamilton ``` module load gcc/9.3.0 module load intelmpi/gcc/2019.6 ```

## Running - Amount of parallelism specified at _runtime_ - Use MPI library's "runner", usually `mpirun` or `mpiexec` - On Hamilton use `mpirun` ``` $ mpirun -n 10 ./myapp ``` Run 10 copies of `myapp` in same MPI context

## Important concepts - No process special - Communicators are key - Focus on decomposing _data_ and _work_ on that data Next: [point to point messaging](mpi-ptp.html)