Awesome
ParaStation MPI
[TOC]
Introduction
ParaStation MPI is an implementation of the Message-Passing Interface (MPI) Standard. Being an MPICH-derivate it bases on the MPICH-4.1.1 implementation.
As shown in the figure above, ParaStation MPI relies on the low-level communication layer pscom for the implementation of the PSP Device on the ADI-Layer of the MPICH architecture. ParaStation MPI provides full MPI-4 compliance including support for the MPI Sessions model. Furthermore, the Process Management Interface for Exascale (PMIx) is supported as interface to the process manager.
Requirements
Mandatory
- C compiler with C99 support
- ParaStation Communication library (pscom)
Optional
- ParaStation Management (psmgmt) process manager if you do not want to use the Hydra process manager that comes with MPICH
- C++ compiler for applications written in C++
- Fortran compiler for applications written in Fortran
- PMIx library if you want to use PMIx as interface to the process manager (use of PMIx and Hydra is mutual exclusive)
- hcoll library for hcoll support
- hwloc library for hwloc support in MPICH/hydra
- CUDA runtime library for CUDA awareness
Installation
Configure ParaStation MPI
Download the source code from GitHub:
$ git clone https://github.com/ParaStation/psmpi.git psmpi
Run Autotools:
$ cd /path/to/psmpi
$ ./autogen.sh
We encourage you not to build ParaStation MPI directly within the main source folder:
$ mkdir build && cd build
ParaStation MPI relies on Autotools as the build system requiring a configure step. The build system already provides so-called "confsets" that pass the necessary configuration arguments for a particular installation type to the underlying MPICH build system. It is strongly recommended to rely on these confsets for a proper installation! Currently, the following confsets are provided:
default : Like 'gcc'
none : Do not configure mpich. Prepare only for tar, rpm and srpm build
gcc : Use Gnu compiler (gcc)
intel : Use Intel compiler (icx)
icx : Like 'intel'
pgi : Portland group compiler (pgcc)
nvhpc : Nvidia HPC compiler (nvc)
llvm : llvm-based toolchains (e.g., AOCC)
devel : With error checking and debug info (gcc)
debug : Like 'devel' but without any optimizations
user : No predefined options
ch3 : original mpich ch3 device (no parastation)
ch4 : original mpich ch4 device (no parastation)
The following example configures ParaStation MPI for using the gcc compiler:
$ ../configure --prefix=/path/to/installation/dir --with-confset=default
We recommend developers to use the devel
confset for more error checking and debug infos at compile time.
Optional configure arguments
Argument | Default | Description |
---|---|---|
--with-pscom[=path] | yes | Use pscom as communication transport [path to pscom installation] |
--with-cuda[=path] | no | Use CUDA awareness [path to CUDA installation] |
--with-hydra | no | Build and install MPICH's process manager Hydra |
--with-hcoll[=path] | no | Use hcoll support [path to hcoll installation] |
--with-hwloc[=path] | no | Use hwloc in MPICH/Hydra [built-in or path to hwloc installation] |
--with-pmix[=path] | no | Use PMIx as process manager interface [path to PMIx installation] |
--enable-threading | disabled | Enable multi-thread support |
--enable-msa-awareness | disabled | Enable MSA awareness like hierarchical collectives |
--enable-statistics | disabled | Enable the collection of statistical information |
--enable-confset-overwrite | disabled | Enable overwriting of compilers selected by confset via environment variables CC , CXX and FC |
Build ParaStation MPI
For a successful build, you have to provide a path to your pscom installation via the
--with-pscom[=path]
configure option (defaults to /opt/parastation
) or alternatively by setting
the LIBRARY_PATH
and CPATH
environment variables:
$ export LIBRARY_PATH=/path/to/pscom/installation/lib[64]:${LIBRARY_PATH}
$ export CPATH=/path/to/pscom/installation/include:${CPATH}
Now, ParaStation MPI can be built and installed in accordance with the configuration arguments:
$ make -j8 && make install
Prepare the environment
To use ParaStation MPI for building and running MPI applications, it is advisable to adjust the environment properly. This can be done, e.g., by using the following bash script:
#!/bin/bash
if [ $# -eq 0 ]; then
echo "ERROR: Please provide the path to ParaStation MPI. Abort!"
exit -1
fi
PARASTATION_MPI=`realpath ${1}`
export PATH="${PARASTATION_MPI}/bin${PATH:+:}${PATH}"
export CPATH="${PARASTATION_MPI}/include${CPATH:+:}${CPATH}"
export LD_LIBRARY_PATH="${PARASTATION_MPI}/lib${LD_LIBRARY_PATH:+:}${LD_LIBRARY_PATH}"
export LIBRARY_PATH="${PARASTATION_MPI}/lib${LIBRARY_PATH:+:}${LIBRARY_PATH}"
Alternative Installation
Instead of relying on the pscom as a shared-library, ParaStation MPI can be optionally compiled as a single shared-object by directly using the pscom sources. For doing so, the pscom source files are required:
$ git clone https://github.com/ParaStation/pscom.git pscom
After downloading the psmpi sources, the same
configuration parameters apply as discussed above (see Configure).
Additionally, you will need to add the --with-pscom-allin
flag, e.g.:
$ ../configure --prefix=/path/to/installation/dir --with-confset=default --with-pscom-allin=/path/to/pscom/sources
By default, the pscom4ucp as well as the pscom4psm plugins are included
firmly into ParaStation MPI if --with-pscom-allin
is set and the related
low-level drivers are found. For specifying the plugins to be built-in
explicitly, the --with-pscom-builtin[=list]
option can be used.
Runtime Parameters
Debugging
Environment Variable | Description |
---|---|
PSP_DEBUG=0 | only fatal conditions (like detected bugs) |
PSP_DEBUG=1 | fatal conditions + errors (default) |
PSP_DEBUG=2 | + warnings |
PSP_DEBUG=3 | + information |
PSP_DEBUG=4 | + debug |
PSP_DEBUG=5 | + verbose debug |
PSP_DEBUG=6 | + tracing calls |
PSP_DEBUG_VERSION=1 | Show always the pscom version (info) |
PSP_DEBUG_CONTYPE=1 | Show connection types (info) |
PSP_HARD_ABORT=0 | Process termination in MPI_Abort() via exit() (default) |
PSP_HARD_ABORT=1 | Process termination in MPI_Abort() via PMI_Abort() |
PSP_HARD_ABORT=2 | Process termination in MPI_Abort() via _exit() |
PSP_HARD_ABORT=3 | Process termination in MPI_Abort() via abort() |
PSP_CUDA_ENFORCE_STAGING=1 | Enforce staging of CUDA buffers via host memory (with a significant performance penalty!) |
Feature Activation
Environment Variable | Description | Required build config options |
---|---|---|
PSP_CUDA=1 | Enable/Disable CUDA awareness (default = 0) | --with-cuda |
PSP_HCOLL=1 | Enable/Disable HCOLL support (default = 0) | --with-hcoll[=path] |
PSP_SMP_AWARENESS=1 | Take locality information into account (default = 1) | |
PSP_SMP_AWARE_COLLOPS=1 | Enable/Disable SMP-aware collectives (default = 0) | --enable-msa-awareness |
PSP_MSA_AWARENESS=1 | Take topology information into account (default = 0) | --enable-msa-awareness |
PSP_MSA_AWARE_COLLOPS=1 | Enable/Disable MSA-aware collectives (default = 0) | --enable-msa-awareness |
Statistical Analysis
Environment Variable | Description | Required build config options |
---|---|---|
PSP_HISTOGRAM=1 | Enable the collection of statistical data | --enable-statistics |
PSP_HISTOGRAM_MIN=x | Set the lower message size limit for the histogram | --enable-statistics |
PSP_HISTOGRAM_MAX=y | Set the upper message size limit for the histogram | --enable-statistics |
PSP_HISTOGRAM_SHIFT=z | Bit shift for the number of bins of the histogram | --enable-statistics |
PSP_HISTOGRAM_CONTYPE=con | Limit the histogram to a particular connection type | --enable-statistics |
PSP_HCOLL_STATS=1 | Enable the collection of HCOLL usage statistics | --enable-statistics --with-hcoll[=path] |
Universe size
The attribute MPI_UNIVERSE_SIZE
is determined via the environment variable MPIEXEC_UNIVERSE_SIZE
which can be set explicitly by the user or implicitly through the process manager. The following cases are considered:
- If
MPIEXEC_UNIVERSE_SIZE
is set, then its value is returned as theMPI_UNIVERSE_SIZE
attribute. This holds even if the value is smaller than the world size. Users are expected to set the universe size appropriately for the needs of their application. - If
MPIEXEC_UNIVERSE_SIZE
is not set, then theMPI_UNIVERSE_SIZE
attribute is not available.
Test Suite
MPICH has a test suite that can also be used and is even extended by ParaStation MPI.
$ make test
...will run the complete test suite comprising tests in most of the subfolders in mpich2/test/mpi/
.
For getting an overview of these folders, one can use the help
wildcard
together with the TESTDIR
variable:
$ make test TESTDIR=help
Please note that some of these test folders in turn contain further subfolders
with tests. For example, for listing the subfolders with tests within the folder
threads
, type:
$ make test TESTDIR=threads/help
However, you can also have only tests of a certain subdirectory to be executed by specifying a TESTDIR
, for example:
$ make test TESTDIR=pt2pt
or for another example:
$ make test TESTDIR=threads/pt2pt
Selecting Test Sets
In addition, a TESTSET
can be specified to further restrict the number of tests.
The tests that belong to a certain test set are stated as a list in a file with the same name within the above listed subdirectories.
Currently, only ps-test-minimal
(minimal list of tests) and testlist
(containing all tests) are valid test sets.
(Use the wildcard TESTSET=help
to list all valid test sets.)
So, for example, the following invocation runs all tests belonging to ps-test-minimal
within all subdirectories:
$ make test TESTSET=ps-test-minimal
And if, for example, the test set ps-test-minimal
should only be executed for tests within the subfolder pt2pt
, then the following invocation is the means of choice:
$ make test TESTSET=ps-test-minimal TESTDIR=pt2pt
In the case that several directories and/or several test sets should be used for a run, one can do this by using comma-separated lists and the variables TESTDIRS
and TESTSETS
.
(Please note the additional S
at the end of the variable names!)
However, in this case no explicit check for validity of the passed test sets or folder names is performed.
Example:
$ make test TESTDIRS=pt2pt,coll,info
Furthermore, there is a "meta" set of tests that are specifically intended for large numbers of processes and/or large messages and data sets.
This additional (and somehow orthogonal) set can be chosen via largetest
as the
make
target:
$ make largetest
...which can also be used in combination with the TESTSET
/TESTSETS
and the TESTDIR
/TESTDIRS
options.
Using Test Configurations
When running the tests, some settings like the mpiexec
command to be used are pre-configured and depend on the configuration chosen for psmpi during its top-level configure call.
However, there is still the possibility to overwrite some of such configurations by explicitly setting the following environment variables:
MPIEXEC
: choose the path and the executable to be used for process startMPIEXECARG
: additional arguments to be passed to thempiexec
command
For example:
$ make test MPIEXEC=/opt/parastation/bin/mpiexec MPIEXECARG="-e MPIEXEC_UNIVERSE_SIZE=128"
Furthermore, there is a small set of such overriding configurations that can also be chosen via the TESTCONF
option.
The following configurations are available and can be shown by using TESTCONF=help
:
* hydra # Enforce the use of mpiexec as provided by MPICH/Hydra
* psmgmt # Enforce the use of mpiexec as provided by ParaStation/psmgmt
* verbose # Enable verbose output during the test run
* prrte # Use PRRTE's prterun to start the MPI processes
* patient # Increase the timeout for all tests by a factor of 10
* psmgmt/memcheck # Run the tests on Valgrind (and psmgmt's mpiexec) with leak-check=full
* psmgmt/pmix # Run the tests using PMIx as interface to the process manager (and psmgmt's mpiexec)
* psmgmt/valgrind # Run the tests on Valgrind (and psmgmt's mpiexec)
If multiple of such configurations are to be selected at the same time, then the TESTCONFS
(mind the additional S
at the end!) can be used.
Example:
$ make test TESTCONFS=hydra,verbose
Contributing
To ensure that all commits conform to the coding style, the pre-commit hook should be activated. Therefore, you have to link this from the top-level source directory:
$ ln -s ../../scripts/hooks/pre-commit .git/hooks/pre-commit