13 unstable releases (5 breaking)

0.8.0 May 3, 2024
0.7.0 Oct 21, 2023
0.6.0 Aug 5, 2022
0.5.4 Mar 27, 2018
0.2.0 Nov 10, 2015

#25 in Concurrency

Download history 288/week @ 2024-07-20 165/week @ 2024-07-27 165/week @ 2024-08-03 272/week @ 2024-08-10 131/week @ 2024-08-17 219/week @ 2024-08-24 319/week @ 2024-08-31 498/week @ 2024-09-07 1018/week @ 2024-09-14 446/week @ 2024-09-21 276/week @ 2024-09-28 274/week @ 2024-10-05 260/week @ 2024-10-12 288/week @ 2024-10-19 738/week @ 2024-10-26 819/week @ 2024-11-02

2,187 downloads per month
Used in 10 crates

MIT/Apache

300KB
5.5K SLoC

MPI bindings for Rust

GitHub Actions Documentation Crates.io License: Apache License 2.0 or MIT

The Message Passing Interface (MPI) is a specification for a message-passing style concurrency library. Implementations of MPI are often used to structure parallel computation on High Performance Computing systems. The MPI specification describes bindings for the C programming language (and through it C++) as well as for the Fortran programming language. This library tries to bridge the gap into a more rustic world.

Requirements

An implementation of the C language interface that conforms to MPI-3.1. rsmpi is currently tested with these implementations:

Users have also had success with these MPI implementations, but they are not tested in CI:

For a reasonable chance of success with rsmpi with any MPI implementation, you must have one of:

  • export MPI_PKG_CONFIG to be the name or path for pkg-config for your implementation
    • rsmpi automatically uses CRAY_MPICH_DIR on Cray environments so the above need not be set
    • mpich and ompi are tried by default as a last resort
    • Tip: test with a command like pkg-config --cflags --libs mpich
  • The implementation provides a C compiler wrapper mpicc
    • export MPICC=/path/to/mpicc to specify fully
    • otherwise tries mpicc in $PATH
    • mpicc -show should print the full command line that is used to invoke the wrapped C compiler in gcc-compatible syntax (e.g., -lmpi, -I/usr/local/include, ...)
  • On Windows, the variables MSMPI_INC and either MSMPI_LIB32 or MSMPI_LIB64 should be set

Since the MPI standard leaves some details of the C API unspecified (whether to implement certain constants and even functions using preprocessor macros or native C constructs, the details of most types, etc.) rsmpi takes a two step approach to generating functional low-level bindings.

First, it uses a thin static library written in C (see rsmpi.h and rsmpi.c) that tries to capture the underspecified identifiers and re-exports them with a fixed C API. This library is built from build.rs using the gcc crate.

Second, to generate FFI definitions tailored to each MPI implementation, rsmpi uses rust-bindgen which needs libclang. See the bindgen project page for more information.

Furthermore, rsmpi uses the libffi crate which installs the native libffi which depends on certain build tools. See the libffi project page for more information.

Usage

Add the mpi crate as a dependency in your Cargo.toml:

# "features" is optional
[dependencies]
mpi = { version = "0.8.0", features = ["user-operations", "derive"] }

Then use it in your program like this:


use mpi::request::WaitGuard;
use mpi::traits::*;

fn main() {
    let universe = mpi::initialize().unwrap();
    let world = universe.world();
    let size = world.size();
    let rank = world.rank();

    let next_rank = (rank + 1) % size;
    let previous_rank = (rank - 1 + size) % size;

    let msg = vec![rank, 2 * rank, 4 * rank];
    mpi::request::scope(|scope| {
        let _sreq = WaitGuard::from(
            world
                .process_at_rank(next_rank)
                .immediate_send(scope, &msg[..]),
        );

        let (msg, status) = world.any_process().receive_vec();

        println!(
            "Process {} got message {:?}.\nStatus is: {:?}",
            rank, msg, status
        );
        let x = status.source_rank();
        assert_eq!(x, previous_rank);
        assert_eq!(vec![x, 2 * x, 4 * x], msg);

        let root_rank = 0;
        let root_process = world.process_at_rank(root_rank);

        let mut a;
        if world.rank() == root_rank {
            a = vec![2, 4, 8, 16];
            println!("Root broadcasting value: {:?}.", &a[..]);
        } else {
            a = vec![0; 4];
        }
        root_process.broadcast_into(&mut a[..]);
        println!("Rank {} received value: {:?}.", world.rank(), &a[..]);
        assert_eq!(&a[..], &[2, 4, 8, 16]);
    });
}

Features

The bindings follow the MPI 3.1 specification.

Currently supported:

  • Groups, Contexts, Communicators:
    • Group and (Intra-)Communicator management from section 6 is mostly complete.
    • no Inter-Communicators
    • no process topologies
  • Point to point communication:
    • standard, buffered, synchronous and ready mode send in blocking and non-blocking variants
    • receive in blocking and non-blocking variants
    • send-receive
    • probe
    • matched probe/receive
  • Collective communication:
    • barrier
    • broadcast
    • (all) gather
    • scatter
    • all to all
    • varying counts operations
    • reductions/scans
    • blocking and non-blocking variants
  • Datatypes: Bridging between Rust types and MPI basic types as well as custom MPI datatypes which can act as views into buffers.

Not supported (yet):

  • One-sided communication (RMA)
  • MPI parallel I/O
  • A million small things

Optional Cargo Features

These optional features can be enabled in your cargo manifest. See the Usage section above.

user-operations enables capturing lambdas and safe creation in UserOperation. This feature requires the libffi system library, which is not available on all systems out-of-the-box.

let mut h = 0;
comm.all_reduce_into(
    &(rank + 1),
    &mut h,
    &UserOperation::commutative(|x, y| {
        let x: &[Rank] = x.downcast().unwrap();
        let y: &mut [Rank] = y.downcast().unwrap();
        for (&x_i, y_i) in x.iter().zip(y) {
            *y_i += x_i;
        }
    }),
);

derive enables the Equivalence derive macro, which makes it easy to send structs over-the-wire without worrying about safety around padding, and allowing arbitrary datatype matching between structs with the same field order but different layout.

#[derive(Equivalence)]
struct MyProgramOpts {
    name: [u8; 100],
    num_cycles: u32,
    material_properties: [f64; 20],
}

Documentation

Every public item of rsmpi should at least have a short piece of documentation associated with it. Documentation can be generated via:

cargo doc

Documentation for the latest version of the crate released to crates.io is hosted on Github pages.

Examples

See files in examples/. These examples also act as integration tests.

Python integration

It is possible to use rsmpi with a communicator provided by mpi4py. An example project demonstrating this is mpi4py_with_rsmpi.

License

Licensed under either of

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

Dependencies

~0.4–3.5MB
~75K SLoC