[Custom Issue]: Post-processing: Ensure calliper data are ordered by MPI rank

[oakleybrunt]

Contact Details

oakley.brunt@metoffice.gov.uk

Are there any linked Issues or Pull Requests?

No response

Brief description

Data read in by VernierReader.load() are not ordered in any way. This is especially apparent when reading in multiple files using the ThreadPoolExecutor.

Further details of the issue.

MPI rank is important information when examining processes executing on specific ranks, so this information should be added.

The most likely route to retrieving this information is by scraping the header section of each MPI rank's calliper data and storing this information alongside the timing data:

Task 1 of 2 : MPI rank ID 0

There are are two options for using this information after scraping:

1. Order the calliper data by MPI rank and then convert the VernierCalliper attributes to ordered constructs (e.g. Python's deque)
2. Keep the MPI rank data as another VernierCalliper attribute but ensure that "rows" are kept intact during data operations

Both methods have drawbacks, suggestions on implementation are welcome.

[oakleybrunt]

Since the ordering of the VernierCalliper data is not manipulated anywhere (currently), the best option is to just include the MPI rank as a new data attribute.

[EdHone]

I had a think about this and tried a few ways to preserve the rank order - I actually think the current implementation will preserve the order but the best way to ensure that the information is retained is to add a mpi_rank field to the VernierData class. That way a user also doesnt have to convert the 0-indexed arrays into a 1-indexed task number - not exactly onerous but still an additional inference that may not be needed

[oakleybrunt]

That's easy enough. Thanks Ed.