Thin wrapper to a Message Passing Interface (MPI)

Contents

• Introduction
• Installation on Debian/Ubuntu/Linux
• Installation on macOS
• Setting Cargo.toml
• Examples
• Todo list

Introduction

MsgPass (Message Passing) is a thin Rust wrapper to MPI. We consider a small subset of MPI functions. This subset will grow as our projects require more functionality. We implement (by hand) C functions that Rust can easily call using the FFI (in the c_code directory).

We try to test all functions as much as possible, but test coverage could be better. The tests must be called with mpiexec, thus it is easy to use the run-tests.bash script.

[!NOTE] Unlike the MPI standard, we use mpi_init to initialize the simulation with multiple threads. Thus, our mpi_init function calls MPI_Init_thread with MPI_THREAD_MULTIPLE. On the other hand, our mpi_init_single_thread calls MPI_Init.

Documentation:

Installation on Debian/Ubuntu/Linux

On Ubuntu/Linux, install OpenMPI or MPICH (requires an environment variable). For instance,

sudo apt install libopenmpi-dev

or

sudo apt install libmpich-dev

For MPICH, the following environment variable is required:

export MSGPASS_USE_MPICH=1

Installation on macOS

On macOS, install the following packages:

brew install llvm@13 open-mpi

Also, export the following environment variable:

export echo TMPDIR=/tmp

Setting Cargo.toml

👆 Check the crate version and update your Cargo.toml accordingly:

[dependencies]
msgpass = "*"

Examples

See also:

• examples

The example below (available in the examples directory) will send an array from ROOT to all the other processors.

use msgpass::*;

fn main() -> Result<(), StrError> {
    mpi_init()?;

    let rank = mpi_world_rank()?;
    let size = mpi_world_size()?;
    let mut comm = Communicator::new()?;

    const ROOT: i32 = 0;
    const TAG: i32 = 70;

    if rank == ROOT as usize {
        let x = vec![1.0, 2.0, 3.0];
        for to in 1..size {
            comm.send_f64(&x, to, TAG)?;
        }
        println!("{}: x = {:?}", rank, x);
    } else {
        let mut y = vec![0.0, 0.0, 0.0];
        comm.receive_f64(&mut y, ROOT, TAG)?;
        println!("{}: y = {:?}", rank, y);
    }

    mpi_finalize()
}

Running the code above with mpiexec -np 4 ex_send_receive (see run-examples.bash), we get an output similar to the one below:

### ex_send_receive ######################################################
2: y = [1.0, 2.0, 3.0]
0: x = [1.0, 2.0, 3.0]
3: y = [1.0, 2.0, 3.0]
1: y = [1.0, 2.0, 3.0]

Todo list

• Implement basic functionality
  • Initialize and finalize
  • Abort and barrier
• Wrap more MPI functions
  • Implement reduce
  • Implement allreduce
  • Implement send/receive
  • Implement scatter/gather