GitHub

Parallel execution

LikelihoodProfiler.jl supports parallel computation to accelerate profile likelihood calculations. All implemented methods support independent computations across parameters and profile branches ("left" and "right"). Depending on your system configuration and workload, you can leverage either multithreading or distributed processing (workers) for parallel execution.

Available parallel modes

Parallelization is controlled via the parallel_type keyword argument in the profile function.

profile(plprob, method; parallel_type = :none)

Supported values:

  • :none (default): Run all profiling computations sequentially.
  • :threads: Use Julia's multithreading (i.e. Base.Threads.@threads) across available CPU threads. To enable multithreading, set the JULIA_NUM_THREADS environment variable.
  • :distributed: Use Distributed.jl to run tasks across multiple Julia processes (e.g., started via addprocs() or julia -p N).

Distributed workers example

The following example illustrates how to compute profile likelihoods using distributed parallelism across Julia processes in a multi-core (or multi-node) environment.

using Distributed

addprocs(2)

@everywhere using LikelihoodProfiler, Optimization, ForwardDiff

@everywhere rosenbrock(x,p) = (1.0 - x[1])^2 + 100.0*(x[2] - x[1]^2)^2

x0 = [1., 1.]
optf = OptimizationFunction(rosenbrock, AutoForwardDiff())
optprob = OptimizationProblem(optf, x0)

plprob = PLProblem(optprob, x0, (-5.,5.); threshold = 1.0)
meth = OptimizationProfiler(optimizer = Optimization.LBFGS(), stepper = FixedStep(; initial_step=0.1))

sol = profile(plprob, meth; parallel_type=:distributed)