FOCAPO_ExaProcess

Scripts to reproduce the flash benchmark results for the paper "Harnessing GPU-Acceleration in Large-Scale Process Optimization".

The main benchmark runs JuMP/Ipopt, ExaModels/MadNLP on CPU, and ExaModels/MadNLP on GPU for several scenario counts. The output is written to benchmark_table.csv.

Requirements

• Julia 1.12.1. The committed Manifest.toml was generated with this version.
• An NVIDIA GPU with a working CUDA driver.
• The HSL MA57 linear solver. The scripts use MA57 through HSL_jll for JuMP/Ipopt and through MadNLPHSL for MadNLP CPU solves.
• This repository, including the committed data/inlet_data_*_scenarios.jls files.

The full benchmark can take about 4 hours, depending on CPU, GPU, and solver configuration.

Step-by-step reproduction

1. Clone the repository and enter it.

   git clone <repository-url>
   cd FOCAPO_ExaProcess

2. Start Julia in the project environment.

   julia --project=.

3. Instantiate the exact package environment from Manifest.toml.

   using Pkg
   Pkg.instantiate()

4. Verify that CUDA is available to Julia.

   using CUDA
   CUDA.functional()

   This should print true. If it prints false, fix the NVIDIA driver/CUDA setup before running the GPU benchmark.

5. Run the full benchmark table.

   include("flash_benchmark_table.jl")

   The script runs the default scenario counts:

   [1, 10, 100, 1000, 5000]

   For each scenario count, it solves:

   • JuMP/Ipopt
   • ExaModels/MadNLP CPU
   • ExaModels/MadNLP GPU
   • ExaModels/MadNLP GPU with the two-stage Schur-complement KKT system

6. Confirm the output.

   After the script finishes, it writes:

   benchmark_table.csv
   

   The CSV contains one row per scenario count and these columns:

   omega,jump_build_s,jump_solve_s,jump_iter,jump_status,exa_cpu_build_s,exa_cpu_solve_s,exa_cpu_iter,exa_cpu_status,exa_gpu_build_s,exa_gpu_solve_s,exa_gpu_iter,exa_gpu_status,exa_gpu_schur_build_s,exa_gpu_schur_solve_s,exa_gpu_schur_iter,exa_gpu_schur_status
   

   Successful ExaModels/MadNLP runs should report SOLVE_SUCCEEDED. JuMP/Ipopt runs may report LOCALLY_SOLVED; for some large cases, the status can depend on solver version, hardware, and runtime limits.

Notes

• The benchmark script checks CUDA.functional() before running and stops if CUDA is not functional.
• benchmark_table.csv is overwritten each time flash_benchmark_table.jl completes.
• Timing values include machine-dependent build and solve times, so exact numbers are not expected to match across different systems.