ADIOS2 File Creation

[rfbird]

Hey,

I swapped from ADIOS1 to ADIOS2 as my backend, and my writes started failing because the output dir (my_filename.bp) was not created before the contents were written. I assume this not expected behavior?

Where is the equivalent of

openPMD-api/src/IO/ADIOS/ADIOS1IOHandler.cpp

Line 100 in 03cd55a

createFile(i.writable, *dynamic_cast< Parameter< Operation::CREATE_FILE >* >(i.parameter.get()));

for ADIOS2?

Axel said to spam issues just Incase something was afoot ;)

Best,
Bob

[ax3l]

Thank you for the report!
I haven't experienced this in my tests yet, but it might be related to the fact that you wrote a particle series and some ranks might have not written particles. This is a valid use case that might not fully covered by our CI tests, so we have to dig into this. Can you confirm this or have slightly more info when it occured?

Generally, ADIOS2 does not need a meta-file anymore since the .bp.dir directory is all one needs to open the data again with ADIOS2 tools. Did you see a crash due to the missing directory at write time?

Let me ping @franzpoeschel on this, he'll be back at work around next week and just successfully defended his MA thesis (congratulations!! 🚀 ✨ ).

[rfbird]

It died trying to write to ./hydro_electron.bp/md.0, because the dir was not present at the time it tried to write the file, and it was fixed if we manually create the folder for it and re ran the job

All ranks should write particles in this case. I guess it could be something screwy our end, but from a quick look at the code it seems like you expect OpenPMD to successfully create the bp dir

[ax3l]

Thanks for the details. In openPMD-api, we actually just create all missing dirs recursively that are before the filename (e.g. ./some/path/well/oh/). The hydro_electron.bp/` dir is done on the ADIOS2 side.
Could be a non-collective call on our side when opening the file, or something upstream. We'll investigate.

Can you please provide the error message on which that writer died, too? The closer we get this for us to write a reproducer to debug the better.
Just to ask the obvious: you do not clean up your simulation output dir ./* at some point during app initialization, right?

[ax3l]

After further investigation, I got a similar error today.

This error only appears when you first wrote an ADIOS1 file, which creates a my_filename.bp for you.
I guess you then re-tested your app and switched to ADIOS2, which tries to create a my_filename.bp/ directory.

Simple solution: clear the output directory from pre-existing .bp files.
This will avoid errors of the kind:

  what():  ERROR: couldn't open file my_filename.bp/md.0, check permissions or path existence, in call to POSIX open