Introduction

The OSHMPI project provides an OpenSHMEM 1.4 implementation on top of MPI-3. MPI is the standard communication library for HPC parallel programming. OSHMPI provides a lightweight OpenSHMEM implementation on top of the portable MPI-3 interface and thus can utilize various high-performance MPI libraries on HPC systems.

Getting Started

1. Installation

The simplest way to configure OSHMPI is using the following command-line:

./configure --prefix=/your/oshmpi/installation/dir CC=/path/to/mpicc CXX=/path/to/mpic++

Once you are done configuring, you can build the source and install it using:

make && make install

2. Test

OSHMPI contains some simple test programs under the tests/ directory. To check if OSHMPI built correctly:

cd tests/ && make check

For comprehensive test, we recommend using external test suites such as SOS Test Suite.

3. Compile an OpenSHMEM program with OSHMPI

OSHMPI supports C and C++ programs. Example to compile the C program test.c:

/your/oshmpi/installation/dir/bin/oshcc -o test test.c

Example to compile the C++ program test.cpp:

/your/oshmpi/installation/dir/bin/oshc++ -o test test.cpp

4. Execute an OSHMPI program

OSHMPI relies on the MPI library's startup command mpiexec or mpirun to execute programs. Example to run the binary test:

/path/to/mpiexec -np 2 ./test

For more information about the MPI startup command, please check the MPI library's documentation.

Configure Options

Below are some commonly used configure options. For detailed explanation please check ./configure --help.

• Default configuration

./configure --prefix=/your/oshmpi/installation/dir CC=/path/to/mpicc CXX=/path/to/mpic++

• With fast build, no multi-threading, disable all active messages.

./configure --prefix=/your/oshmpi/installation/dir CC=/path/to/mpicc CXX=/path/to/mpic++ \
    --enable-fast --enable-threads=single --enable-amo=direct --enable-rma=direct

• With MPICH/OFI version 3.4b1 or newer, enable dynamic window enhancement for scalable memory space implementation.

./configure --prefix=/your/oshmpi/installation/dir CC=/path/to/mpicc CXX=/path/to/mpic++ \
    --enable-win-type=dynamic_win

• With CUDA GPU memory space support

./configure --prefix=/your/oshmpi/installation/dir CC=/path/to/mpicc CXX=/path/to/mpic++ \
    --with-cuda=/path/to/cuda/installation

Examples

Examples to use memory space extension and CUDA GPU memory kind can be found at the built-in test suite:

  tests/space.c
  tests/space_int_amo.c
  tests/space_int_put.c
  tests/space_ctx_int_put.c

Environment Variables

OpenSHMEM Standard Environment Variables

• SHMEM_SYMMETRIC_SIZE (default 128 MiB) Number of bytes to allocate for symmetric heap. The size value can be either number of bytes or scaled with a suffix of:

  • "K" or "k" for kilobytes (B * 1024)
  • "M" or "m" for Megabytes (KiB * 1024)
  • "G" or "g" for Gigabytes (MiB * 1024)
  • "T" or "t" for Terabytes (GiB * 1024)

• SHMEM_DEBUG (0|1, default 0) Enables debugging messages from the OpenSHMEM runtime. It is invalid when configured with --enable-fast (see --enable-fast).

• SHMEM_VERSION (0|1, default 0) Prints OpenSHMEM library version at start_pes(), shmem_init(), or shmem_init_thread().

• SHMEM_INFO (0|1, default 0) Prints the list of OpenSHMEM environment variables at stdout.

OSHMPI Environment Variables

• OSHMPI_VERBOSE (0|1, default 0) Prints both the standard and OSHMPI environment variables at stdout.

• OSHMPI_AMO_OPS (Comma-separated operations, default "all") Defines the AMO operations used in the OpenSHMEM program. If all PEs issue concurrent AMOs only with the same operation, or with the same operation and fetch, then OSHMPI can directly utilize MPI accumulate operations. This is because, MPI grantees the atomicity of accumulate operations only with same_op or same_op_no_op. The default value of OSHMPI_AMO_OPS is "cswap,finc,inc,fadd,add,fetch,set,swap,fand,and,for,or,fxor,xor" (identical to "all"). Arbitrary combination of the above operations can be given at execution time.

  The variable can be adjusted only with configure --enable-amo=auto.

• OSHMPI_MPI_GPU_FEATURES (Comma-separated features, default "all") Defines the list of GPU-aware communication functions provided by the underlying MPI library. The default value is "pt2pt,put,get,acc" (identical to "all"). Arbitrary combination of the above features can be given at execution time. If none of the features is supported, specify "none".

• OSHMPI_ENABLE_ASYNC_THREAD (0|1, default 0) Runtime control of asynchronous progress thread when MPI point-to-point based active messages are used internally. Both AMO and RMA may use the active message based approach:

  • When AMO cannot be directly translated to MPI accumulates (see OSHMPI_AMO_OPS), each AMO operation is issued via active messages.
  • When GPU buffer is used in an RMA operation but the MPI library does not support GPU awareness in the RMA mode (see OSHMPI_MPI_GPU_FEATURES), each RMA operation is issued via active messages. The variable can be adjusted only with configure --enable-async-thread=auto and --enable-threads=multiple.

Debugging Options

• Enable debugging flag by setting the configure option --enable-g.
• Set environment variable SHMEM_DEBUG=1 to print OSHMPI internal debugging message.

Test Framework

OSHMPI uses SOS Test Suite for correctness test.

• Tested platforms: CentOS, MacOS (compilation-only)
• Tested MPI implementations:
  • MPICH-3.4rc1 (with CH3/TCP, CH4/OFI, or CH4/UCX)
  • MPICH main branch (with CH3/TCP, CH4/OFI, or CH4/UCX)
  • OpenMPI 4.0.5 (with UCX)

Known Issues

1. Some OpenSHMEM features are not supported in OSHMPI.

   • Context: OSHMPI cannot create separate or shared context on top of MPI interfaces. This feature may be enabled if the MPI end-point concept is accepted at MPI forum and implemented in MPI libraries. Current version always returns SHMEM_NO_CTX error at shmem_ctx_create(); shmem_ctx_destroy() is a no-op.

2. OSHMPI does not provide Fortran APIs.

3. OSHMPI may not work on 32-bit platforms. This is because some internal routines rely on 64-bit integer, e.g., shmem_set_lock(), shmem_clear_lock(), shmem_test_lock.

4. OSHMPI may not correctly initialize symmetric data on OSX platform.

Support

If you have problems or need any assistance about the OSHMPI installation and usage, please contact [email protected] mailing list.

Bug Reporting

If you have found a bug in OSHMPI, please contact [email protected] mailing list, or create an issue ticket on github: https://github.com/pmodels/oshmpi. If possible, please try to reproduce the error with a smaller application or benchmark and send that along in your bug report.