Profiling the diocotron instability

[spossann]

We want to understand the costs of the different algorithms used in the diocotron example.

The relevant files are here: https://github.com/struphy-hub/struphy/tree/devel/examples/ToyGyrokinetic/diocotron_instability

We have to install Struphy from source in development mode:

git clone git@github.com:struphy-hub/struphy.git
cd struphy
pip install -e .[dev]
struphy compile

To run the diocotron example:

mpirun -n 1 python examples/ToyGyrokinetic/diocotron_instability/params_diocotron.py

For testing we can run just 10-100 time steps.

Turn on profiling with

export LINE_PROFILE=1

and run again. Increase the number of MPI ranks and see what happens to the time per time step.

We can try different algorithms in the propagator model.propagators.push_gc_bxe.options, for instance explicit algos with various Butcher tableaus. Just change the file params_diocotron.py for this.

See some plots with

python examples/ToyGyrokinetic/diocotron_instability/pproc_diocotron.py

Our tasks are:

• check the strong scaling by increasing the number of MPI ranks, for different pusher algorithms
• identify bottlenecks via line_profiler

Based on these results, we will try to alleviate some of the bottlenecks.