openPMD: Use Steps if != BTD (#2157)

* openPMD: Use Steps if != BTD

For all but back-transformed diagnostics, we can use efficient,
temporally sequentially increasing writes to iteration numbers
for iterations.

This allows us to give a guarantee to HPC I/O libraries on how
to arrange data, e.g., we can use ADIOS2 BeginStep() and
EndStep().

This enables paths to:
- streaming workflows, such as ADIOS2 SST or SSC, where we stage data
  over the network instead of using files
  https://openpmd-api.readthedocs.io/en/0.14.0/usage/streaming.html

This mitigates:
- host-side memory aggregation for ADIOS2 with openPMD `groupBased`
  iteration encoding
  https://openpmd-api.readthedocs.io/en/0.14.0/backends/adios2.html#memory-usage

* openPMD: Open Iterations Explicitly

Explicitly open iterations. Usually, file-open operations are delayed
until the first load/storeChunk operation is flush-ed. In parallel
contexts where we might want to do only particle writes from a few
ranks but no field wries in the future, this will avoid that we run
into hangs from the non-collective nature of the first flush.

The Streaming API (i.e., Series::writeIterations()) will call this
method implicitly as well, but back-transformed diags (particles)
might still need this.

* openPMD: 0.13.0+

Needed for streaming API and collective open.

0 files changed, 0 insertions, 0 deletions