MPI

Collective Behavior

openPMD-api is designed to support both serial as well as parallel I/O. The latter is implemented through the Message Passing Interface (MPI).

A collective operation needs to be executed by all MPI ranks of the MPI communicator that was passed to openPMD::Series. Contrarily, independent operations can also be called by a subset of these MPI ranks. For more information, please see the MPI standard documents, for example MPI-3.1 in “Section 2.4 - Semantic Terms”.

Functionality	Behavior	Description
Series	collective	open and close
::flush()	collective	read and write
Iteration [1]	independent	declare and open
::open() [4]	collective	explicit open
Mesh [1]	independent	declare, open, write
ParticleSpecies [1]	independent	declare, open, write
::setAttribute [3]	backend-specific	declare, write
::getAttribute	independent	open, reading
RecordComponent [1]	independent	declare, open, write
::resetDataset [1] [2]	backend-specific	declare, write
::makeConstant [3]	backend-specific	declare, write
::storeChunk [1]	independent	write
::loadChunk	independent	read
::availableChunks [4]	collective	read, immediate result

[1] (1,2,3,4,5,6)

Individual backends, i.e. parallel HDF5, will only support independent operations if the default, non-collective (aka independent) behavior is kept. Otherwise these operations are collective.

[2]

Dataset declarations in parallel HDF5 are only non-collective if chunking is set to none (auto by default). Otherwise these operations are collective.

[3] (1,2)

HDF5 only supports collective attribute definitions/writes; ADIOS2 attributes can be written independently. If you want to support all backends equally, treat as a collective operation. Note that openPMD represents constant record components with attributes, thus inheriting this for ::makeConstant.

When treating attribute definitions as collective, we advise specifying the ADIOS2 JSON/TOML key adios2.attribute_writing_ranks for metadata aggregation scalabilty, typically as adios2.attribute_writing_ranks = 0.

[4] (1,2)

We usually open iterations delayed on first access. This first access is usually the flush() call after a storeChunk/loadChunk operation. If the first access is non-collective, an explicit, collective Iteration::open() can be used to have the files already open. Alternatively, iterations might be accessed for the first time by immediate operations such as ::availableChunks().

Tip

Just because an operation is independent does not mean it is allowed to be inconsistent. For example, undefined behavior will occur if ranks pass differing values to ::setAttribute or try to use differing names to describe the same mesh.

Efficient Parallel I/O Patterns

Note

This section is a stub. We will improve it in future versions.

Write as large data set chunks as possible in ::storeChunk operations.

Read in large, non-overlapping subsets of the stored data (::loadChunk). Ideally, read the same chunk extents as were written, e.g. through ParticlePatches (example to-do).

See the implemented I/O backends for individual tuning options.