Improve performance¶

The biggest cost in a radCAD (and cadCAD) simulation is deep-copying state to protect it from accidental mutation. These options trade safety or detail for speed. Measure before and after; the right choice depends on your model.

Disable `deepcopy`¶

By default radCAD deep-copies State Variables before each state update so your functions can't mutate state outside the framework. Disabling this is the single biggest performance win, at the cost of mutability safety:

from radcad import Engine

experiment.engine = Engine(deepcopy=False)

Warning

With deepcopy=False, mutating a nested object in your state (e.g. a list or dict) will leak across timesteps. Only disable it if your state update functions treat state as immutable. See State mutation & performance.

Choose a faster `deepcopy` method¶

When you do need copying, the method matters. radCAD defaults to pickle (faster than copy.deepcopy, but limited to picklable types). To use a different method, subclass SimulationExecution and override deepcopy_method:

import copy
from radcad.core import SimulationExecution

class CustomSimulationExecution(SimulationExecution):
    @staticmethod
    def deepcopy_method(obj):
        """Use copy.deepcopy instead of the default Pickle dumps/loads."""
        return copy.deepcopy(obj)

experiment.engine.simulation_execution = CustomSimulationExecution

The same subclassing technique works for overriding any other step of the simulation loop.

Drop substeps¶

If you don't need intermediate substeps in post-processing, drop them to cut both runtime and dataset size. Only the final substep of each timestep is kept:

experiment.engine = Engine(drop_substeps=True)

Tune parallelism¶

Make sure you're on a parallel backend (PATHOS is the default).
For very large ensembles, scale onto Ray.
More processes isn't always faster: process startup and data serialisation have overhead. Benchmark with your real model.

Benchmark your changes¶

The repository includes time and memory benchmarks under benchmarks/:

# Time profiling
uv run pytest benchmarks/benchmark_radcad.py
uv run pytest benchmarks/benchmark_single_process.py

# Memory profiling
uv run mprof run benchmarks/benchmark_memory_radcad.py
uv run mprof plot