AppendAfterSteps

[adrianjhpc]

I've a file that I want to overwrite the dataset inside each time I do I/O. i.e. a checkpoint where I only want the latest to survive. I've tried using AppendAfterSteps=0 and opening the file in append mode but it still creates multiple steps, i.e. creates new versions of the variable every time a put is called, is this expected?

I'm programming in Fortran and do the following:

...
 
 call adios2_set_engine(io, 'BP5', ierr)

  call adios2_set_parameter(io, "AppendAfterSteps", "0", ierr)

...

  call adios2_open (bp_writer, io, filename, adios2_mode_append, ierr)

Should I be doing something different?

[pnorbert]

adios2_mode_write should create a new file as you intend for checkpoint files.

[adrianjhpc]

The problem is I can do that, but if I loop on the put it creates a copy of the variable, doesn't overwrite it.

[pnorbert]

I don't understand your statement. What does "loop on the put" mean?

I assume after calling put() for all variables that go into the checkpoint, you close the file, and next time you open again (with adios2_write_mode). Can you show a sketch of the adios calls?

[adrianjhpc]

I don't close the file, just keep it open.

The workflow is like this:

Open file

Define variable

Loop to run the program

Do some work

Put the variable

When finished doing the work close the file.

i.e. We don't bother closing the checkpoint file until then end of the program. We do sync after writing the variable to persist it.

[pnorbert]

That's not how anyone else is doing checkpointing. Close would write out the current data. And whenever you want a new checkpoint, you open, write and close again.

The open(mode) argument as well as the AppendAfterSteps parameter only play a role when opening the file. They are acting for the moment of opening, they will not have any effect later.

By "sync" you mean calling end_step() or what?

[adrianjhpc]

So AppendAfterSteps will only work between file closes? i.e. only the steps that are visible on file open will be considered?

[pnorbert]

The purpose of AppendAfterSteps is to reset data files, other than the checkpoint file, to the restarting point, when restarting the application. Imagine you write a checkpoint every 100 iterations and you write a data file every 10 iterations. You may end up with a checkpoint written at iteration 600, while the data file may have 64 steps (iteration 0, 10, 20, ..630). If you simply open the data file for appending when restarting the application, the next step will be the 65th in the data file, even though you restart from iteration 600 and the next data step is ought to be 610. By setting AppendAfterSteps=60 for the data file, adios will truncate it after 60 steps and the next output will be the 61st (with data from iteration 610).

For the checkpoint file itself, you just should use write mode, and open/close at every checkpoint step.

The examples/simulation/gray-scott/ application demonstrates how this works. It has a ckpt.bp at a given step and a gs.bp data output, with many steps. At restart (see options for that in settings-files.json), it will read the step value from the checkpoint, calculate how many steps should be in gs.bp based on the output frequency and the restart step, and opens it up for appending with truncation.

[adrianjhpc]

Thanks, understood.