Profiling

Where the accessors come from

The examples on this page use output of a Struphy Output, which loads plasma-plots and its .plasma accessors. For other xarray data, run import plasma_plots first (see Getting started).

Struphy records timing regions (propagators, compiled kernels, linear solves, …) when a run is started with sim.run(profiling_activated=True), written to profiling_data.h5. Output.profile reads it back as a scope-profiler ProfilingResults object, and array.plasma.plot.profile wraps three of scope-profiler’s own plotting functions directly on it. Install the optional extra to use them:

pip install "plasma-plots[profiling]"

Every method here is a thin pass-through: it returns exactly what scope-profiler itself returns (a (fig, axes) pair for the default matplotlib backend, or a Plotly figure with backend="plotly"), and forwards every other keyword straight through (ranks, include/exclude, filepath, …) — see scope-profiler’s own docs for the full set.

Gantt chart

out.plot.profile.gantt()

A timeline of every recorded region, one row per rank — the first thing to look at for “where did the time go”:

The same plot with Plotly

out.plot.profile.gantt(backend="plotly")

Loading interactive chart…

Flame chart

out.plot.profile.flame()

Reconstructs the call stack from region timings, one column per call, stacked by nesting depth — good for spotting which specific call of a hot region was slow, not just the region as a whole:

Call graph

out.plot.profile.callgraph(compact=True)

The explicit call graph — which region calls which — without timings. compact=True collapses every invocation of a region into one node, so repeated calls (e.g. once per time step) don’t sprawl the graph; leave it False to see each individual call as its own node: