Skip to content

Latest commit

 

History

History
108 lines (88 loc) · 5.08 KB

README.md

File metadata and controls

108 lines (88 loc) · 5.08 KB
Raw

MPItrampoline

  • MPI wrapper library: GitHub CI
  • MPI trampoline library: GitHub CI
  • MPI integration tests: GitHub CI
  • DOI

MPI is the de-facto standard for inter-node communication on HPC systems, and has been for the past 25 years. While highly successful, MPI is a standard for source code (it defines an API), and is not a standard defining binary compatibility (it does not define an ABI). This means that applications running on HPC systems need to be compiled anew on every system. This is tedious, since the software that is available on every HPC system is slightly different.

This project attempts to remedy this. It defines an ABI for MPI, and provides an MPI implementation based on this ABI. That is, MPItrampoline does not implement any MPI functions itself, it only forwards them to a "real" implementation via this ABI. The advantage is that one can produce "portable" applications that can use any given MPI implementation. For example, this will make it possible to build external packages for Julia via Yggdrasil that run efficiently on almost any HPC system.

A small and simple MPIwrapper library is used to provide this ABI for any given MPI installation. MPIwrapper needs to be compiled for each MPI installation that is to be used with MPItrampoline, but this is quick and easy.

Successfully Tested

  • Debian 11.0 via Docker (MPICH; arm32v5, arm32v7, arm64v8, mips64le, ppc64le, riscv64; C/C++ only)
  • Debian 11.0 via Docker (MPICH; i386, x86-64)
  • macOS laptop (MPICH, OpenMPI; x86-64)
  • macOS via Github Actions (OpenMPI; x86-64)
  • Ubuntu 20.04 via Docker (MPICH; x86-64)
  • Ubuntu 20.04 via Github Actions (MPICH, OpenMPI; x86-64)
  • Blue Waters, HPC system at the NCSA (Cray MPICH; x86-64)
  • Graham, HPC system at Compute Canada (Intel MPI; x86-64)
  • Marconi A3, HPC system at Cineca (Intel MPI; x86-64)
  • Niagara, HPC system at Compute Canada (OpenMPI; x86-64)
  • Summit, HPC system at ORNL (Spectrum MPI; IBM POWER 9)
  • Symmetry, in-house HPC system at the Perimeter Institute (MPICH, OpenMPI; x86-64)

Workflow

Preparing an HPC system

Install MPIwrapper, wrapping the MPI installation you want to use there. You can install MPIwrapper multiple times if you want to wrap more than one MPI implementation.

This is possibly as simple as

cmake -S . -B build -DMPIEXEC_EXECUTABLE=mpiexec -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_INSTALL_PREFIX=$HOME/mpiwrapper
cmake --build build
cmake --install build

but nothing is ever simple on an HPC system. It might be necessary to load certain modules, or to specify more cmake MPI configuration options.

The MPIwrapper libraries remain on the HPC system, they are installed independently of any application.

Building an application

Build your application as usual, using MPItrampline as MPI library.

Running an application

At startup time, MPItrampoline needs to be told which MPIwrapper library to use. This is done via the environment variable MPITRAMPOLINE_LIB. You also need to point MPItrampoline's mpiexec to a respective wrapper created by MPIwrapper, using the environment variable MPITRAMPOLINE_MPIEXEC.

For example:

env MPITRAMPOLINE_MPIEXEC=$HOME/mpiwrapper/bin/mpiwrapper-mpiexec MPITRAMPOLINE_LIB=$HOME/mpiwrapper/lib/libmpiwrapper.so mpiexec -n 4 ./your-application

The mpiexec you run here needs to be the one provided by MPItrampoline.

Current state

MPItrampoline uses the C preprocessor to create wrapper functions for each MPI function. This is how MPI_Send is wrapped:

FUNCTION(int, Send,
         (const void *buf, int count, MT(Datatype) datatype, int dest, int tag,
          MT(Comm) comm),
         (buf, count, (MP(Datatype))datatype, dest, tag, (MP(Comm))comm))

Unfortunately, MPItrampoline does not yet wrap the Fortran API. Your help is welcome.

Certain MPI types, constants, and functions are difficult to wrap. Theoretically, there could be MPI libraries where it is not possible to implement the current MPI ABI. If you encounter this, please let me know -- maybe there is a work-around.