/ishmem

Intel® SHMEM - Device initiated shared memory based communication library

Primary LanguageC++OtherNOASSERTION

Intel® SHMEM

Installation   |   Usage   |   Release Notes   |   Documentation   |   How to Contribute   |   License

Intel® SHMEM provides an efficient implementation of GPU-initiated communication on systems with Intel GPUs.

Table of Contents

Prerequisites

  • Linux OS
  • Intel® oneAPI DPC++/C++ Compiler 2024.0 or higher.

SYCL support

Intel® oneAPI DPC++/C++ Compiler with Level Zero support.

To install Level Zero, refer to the instructions in Intel® Graphics Compute Runtime for oneAPI Level Zero and OpenCL™ Driver repository or to the installation guide for oneAPI users.

Installation

Intel® SHMEM requires a host OpenSHMEM back-end to be used for host-sided operations support. In particular, it relies on a collection of extension APIs (shmemx_heap_create, shmemx_heap_preinit, and shmemx_heap_postinit) to coordinate the Intel® SHMEM and OpenSHMEM heaps. We recommend Sandia OpenSHMEM v1.5.3rc1 or newer for this purpose.

Building Sandia OpenSHMEM (SOS)

Download the SOS repo to be configured as a back-end for Intel® SHMEM.

git clone --recurse-submodules https://github.com/Sandia-OpenSHMEM/SOS.git SOS

Build SOS following instructions below. FI_HMEM support in the provider is required for use with Intel® SHMEM. To enable FI_HMEM with a supported provider, we recommend a specific set of config flags. Below are two examples for configuring and building SOS with two providers supporting FI_HMEM. To configure SOS with the verbs;ofi_rxm provider, use the following instructions:

cd SOS
./autogen.sh
./configure --prefix=<sos_dir> --with-ofi=<ofi_installation> --enable-pmi-simple --enable-ofi-mr=basic --disable-ofi-inject --enable-ofi-hmem --disable-bounce-buffers --enable-hard-polling
make -j
make install

To configure SOS with the HPE Slingshot provider cxi, please use the following instructions:

cd SOS
./autogen.sh
./configure --prefix=<sos_dir> --with-ofi=<ofi_installation> --enable-pmi-simple --enable-ofi-mr=basic --disable-ofi-inject --enable-ofi-hmem --disable-bounce-buffers --enable-ofi-manual-progress --enable-mr-endpoint --disable-nonfetch-amo --enable-manual-progress
make -j
make install

To configure SOS with the psm3 provider, please use the following instructions:

cd SOS
./autogen.sh
./configure --prefix=<sos_dir> --with-ofi=<ofi_installation> --enable-pmi-simple --enable-manual-progress --enable-ofi-hmem --disable-bounce-buffers --enable-ofi-mr=basic --enable-mr-endpoint
make -j
make install

Please choose an appropriate PMI configure flag based on the available PMI client library in the system. Please check for further instructions on SOS Wiki pages. Optionally, users may also choose to add --disable-fortran since fortran interfaces will not be used.

Building Intel® SHMEM

Check that the SOS build process has successfully created an <sos_dir> directory with include and lib as subdirectories. Please find shmem.h and shmemx.h in include.

Build Intel® SHMEM using the following instructions:

cd ishmem
mkdir build
cd build
cmake .. -DSHMEM_INSTALL_PREFIX=<sos_dir> -DCMAKE_INSTALL_PREFIX=<ishmem_install_dir>
make -j

Usage

Launching Example Application

Validate that Intel® SHMEM was built correctly by running an example program.

  1. Add the library path for SOS to the environment:
export LD_LIBRARY_PATH=<sos_dir>/lib:$LD_LIBRARY_PATH
  1. Run the example program or test on an allocated node using a process launcher:
mpiexec.hydra -n 2 -hosts <allocated_node_id> ./scripts/ishmrun ./test/unit/SHMEM/int_get_device

  • Note: Current supported launchers include: MPI process launchers (i.e. mpiexec, mpiexec.hydra, mpirun, etc.), Slurm (i.e. srun, salloc, etc.), and PBS (i.e. qsub).

  • Note: Intel® SHMEM execution model requires applications to use a 1:1 mapping between PEs and GPU devices. Attempting to run an application without the ishmrun launch script may result in undefined behavior if this mapping is not maintained.

  1. Validate the application ran succesfully; example output:
Selected device: Intel(R) Data Center GPU Max 1550
Selected vendor: Intel(R) Corporation
Selected device: Intel(R) Data Center GPU Max 1550
Selected vendor: Intel(R) Corporation
No errors
No errors

Launching Example Application w/ CTest

ctest can be used to run Intel® SHMEM tests that are generated at compile-time. To see a list of tests available via ctest, run:

ctest -N

To launch a single test, execute:

ctest -R <test_name>

Alternatively, all the tests in a directory (such as test/unit/SHMEM/) can be run with the following command:

ctest --test-dir <directory_name>

By default, a passed or failed test can be detected by the output:

    Start 69: sync-2-gpu
1/1 Test #69: sync-2-gpu .......................   Passed    2.29 sec

100% tests passed, 0 tests failed out of 1

To have a test's output printed to the console, add either the --verbose or --output-on-failure flag to the ctest command

Available Scheduler Wrappers for Jobs Run via CTest

The following values may be assigned to CTEST_SCHEDULER at configure-time (ex. -DCTEST_SCHEDULER=mpi) to set which scheduler will be used to run tests launched through a call to ctest:

  • srun (default)
    • Launches CTest jobs on a single node using Slurm's srun.
  • mpi
    • Uses mpirun to launch CTest jobs with the appropriate number of processes.
  • qsub
    • Launches CTest jobs on a single node using qsub. If this option is being used on a system where a reservation must be made (i.e. via pbsresnode) prior to running a test, assign the JOB_QUEUE environment variable to the queue associated with your reservation:
    export JOB_QUEUE=<queue>

Additional Resources

OpenSHMEM Specification

Intel® SHMEM Specification