Skip to content

MPI-IO Support

The Message Passing Interface (MPI) Standard, maintained by the MPI Forum, includes a chapter on MPI-IO.

ROMIO is a well-known implementation of MPI-IO and is included in many MPI implementations. DAOS provides its own MPI-IO ROMIO ADIO driver. This driver has been merged in the upstream MPICH repository, see https://github.com/pmodels/mpich/tree/main/src/mpi/romio/adio/ad_daos for details.

Note

Starting with DAOS 1.2, the --svc parameter (number of service replicas) is no longer needed, and the DAOS API has been changed accordingly Patches have been contributed to MPICH that detect the DAOS API version to gracefully handle this change, but those patches have not yet been picked up in the MPI releases below. For details check the latest commits here.

Supported MPI Version

MPICH

The DAOS ROMIO ADIO driver has been accepted into MPICH. It is included in mpich-3.4.1 (released Jan 2021).

To build MPICH, including ROMIO with the DAOS ADIO driver:

export MPI_LIB=""

git clone https://github.com/pmodels/mpich

cd mpich

./autogen.sh

./configure --prefix=dir --enable-fortran=all --enable-romio \
 --enable-cxx --enable-g=all --enable-debuginfo --with-device=ch3:nemesis \
 --with-file-system=ufs+daos --with-daos=/usr

make -j8; make install

This assumes that DAOS is installed into the /usr tree, which is the case for the DAOS RPM installation. Other configure options can be added, modified, or removed as needed, like the network communicatio device, fortran support, etc. For those, please consule the mpich user guide.

Set the PATH and LD_LIBRARY_PATH to where you want to build your client apps or libs that use MPI to the path of the installed MPICH.

Intel MPI

The Intel MPI Library includes DAOS support since the 2019.8 release.

Note that Intel MPI uses libfabric (both 2019.8 and 2019.9 use libfabric-1.10.1-impi). Care must be taken to ensure that the version of libfabric that is used is at a level that includes the patches that are critical for DAOS. DAOS 1.0.1 includes libfabric-1.9.0, and the DAOS 1.2 and 2.0 releases includes libfabric-1.12.

To use DAOS 1.1 with Intel MPI 2019.8 or 2019.9, the libfabric that is supplied by DAOS (and that is installed into /usr/lib64 by default) needs to be used by listing it first in the library search path:

export LD_LIBRARY_PATH="/usr/lib64/:$LD_LIBRARY_PATH"

There are other environment variables that need to be set on the client side to ensure proper functionality with the DAOS MPIIO driver and those include:

export I_MPI_OFI_LIBRARY_INTERNAL=0
export FI_OFI_RXM_USE_SRX=1

Open MPI

Open MPI 4.0.5 does not yet provide DAOS support. Since one of its MPI-IO implementations is based on ROMIO, it will likely pick up DAOS support in an upcoming release.

MVAPICH2

MVAPICH2 2.3.4 does not yet provide DAOS support. Since its MPI-IO implementation is based on ROMIO, it will likely pick up DAOS support in an upcoming release.

Testing MPI-IO with DAOS

Build any client (HDF5, ior, mpi test suites) normally with the mpicc command and mpich library installed above.

Using the UNS

DAOS UNS allows encoding pool and container information into a path on the filesystem, so one can easily access that container using that path instead of using explicit addressing using the pool and container uuids/labels.

Create a container with a path on dfuse or lustre, or any file system that supports extended attributes:

daos cont create mypool --path=/mnt/dfuse/ --type=POSIX

Then using that path, one can start creating files using the DAOS MPIIO driver by just appending daos: to the filename/path. For example: daos:/mnt/dfuse/file daos:/mnt/dfuse/dir1/file

Using a Prefix Environment Variable

Another way to use the DAOS MPIIO driver is using an environment variable to set the prefix itself for the file:

export DAOS_UNS_PREFIX="path"

That prefix path can be: 1. The UNS prefix if that exists (similar to the UNS mode above): /mnt/dfuse 2. A direct path using the pool and container label (or uuid): daos://pool/container/

Then one can specify the path to the file relative to the root of the container being set in the prefix. So in the example above, if the file to be accessed is under /dir1 in the container, one would pass `daos:/dir1/file' to MPI_File_open().

Using Pool and Container Environment Variables

This mode is meant just for quick testing to use the MPIIO DAOS driver bypassing the UNS and setting direct access with pool and container environment variables. At the client side, the following environment variables need to be set: export DAOS_POOL={uuid/label}; export DAOS_CONT={uuid/label}; export DAOS_BYPASS_DUNS=1. The user still need to append the daos: prefix to the file being passed to MPI_File_open().

Known limitations

Limitations of the current implementation include:

  • No support for MPI file atomicity, preallocate, or shared file pointers.
Back to top