Storage of Ensemble Solutions

When running large ensemble simulations, storing the full high-precision solution can consume a significant amount of memory and disk space. This example demonstrates two strategies to reduce storage requirements:

1. Using strip_solution to remove dense interpolation data before saving.

2. Converting trajectory data to Float32 before saving.

We will use JLD2.jl for saving the data.

using TestParticle
using OrdinaryDiffEq
using SciMLBase
using StaticArrays
using JLD2
using Random

Random.seed!(1234);

Setup: Ensemble Simulation

First, we set up a simple ensemble problem: protons in a uniform magnetic field.

B_uniform(x) = SA[0, 0, 1e-8]
E_zero = TestParticle.ZeroField()

# Initialize a proton with some perpendicular velocity
x0 = [0.0, 0.0, 0.0]
u0 = [1.0, 0.0, 0.0]
stateinit = [x0..., u0...]

param = prepare(E_zero, B_uniform, species = Proton)
tspan = (0.0, 10.0)

# Define the problem
prob = ODEProblem(trace!, stateinit, tspan, param)

# Define a prob_func to vary initial velocity slightly
prob_func(prob, i, repeat) = remake(prob, u0 = prob.u0 .* (1 + 0.1 * rand()))

ensemble_prob = EnsembleProblem(prob, prob_func = prob_func)

# Solve for a small ensemble
# We use a high-order solver (Vern9) which produces dense interpolation data by default
trajectories = 2
sols = solve(ensemble_prob, Vern9(), EnsembleThreads(); trajectories, saveat = 0.2);

Case 1: Converting to Float32

For visualization or statistical analysis, Float64 precision is often unnecessary. We can convert the position and time data to Float32 before saving.

Note: Since sols is an EnsembleSolution containing a vector of ODESolutions, we iterate through them.

filename_f32 = tempname() * ".jld2"

# Extract and convert data
# We structure the data as a vector of named tuples or a dictionary for saving
output_data = map(sols) do s
   (t = Float32.(s.t), u = [Float32.(state) for state in s.u])
end

# Save to JLD2
jldsave(filename_f32; data = output_data)

# Verify loading
loaded_f32 = load(filename_f32)["data"]
println("Loaded Float32 data type: ", eltype(loaded_f32[1].t))
println("Float32 conversion successfully saved and loaded.")

Loaded Float32 data type: Float32
Float32 conversion successfully saved and loaded.

Cleanup

Remove the temporary files.

rm(filename_f32, force = true)