JuMP developers' response to benchmarks by GAMS

miles.lubin · July 22, 2023, 1:09pm

davidanthoff · July 22, 2023, 1:45pm

I’d love to see how the DataFrame JuMP formulation compares to the GAMS solution in terms of performance Did anyone run that comparison?

miles.lubin · July 22, 2023, 1:54pm

I’m also curious, but that would require a GAMS license.

tbeason · July 22, 2023, 2:10pm

With all of the university-affiliated people on this board, I’m sure there is a GAMS license to be found. I just checked and do not have one, unfortunately.

mtanneau · July 22, 2023, 2:32pm

I have access to a GAMS license and link to a couple commercial solvers.
If you give me code I can execute it.

odow · July 24, 2023, 2:19pm

The GAMS code is:

The main entry point is:

github.com

justine18/performance_experiment/blob/master/main_IJKLM.py

import pandas as pd
import numpy as np

# import submodules
import IJKLM.data_generation as data
import visualization
from help import (
    create_directories,
    create_data_frame,
    incremental_range,
    save_to_json,
    below_time_limit,
    process_results,
    print_log_message,
    save_results,
)
from IJKLM.run_gurobipy import run_gurobi, run_fast_gurobi
from IJKLM.run_gams import data_to_gams, run_gams
from IJKLM.run_pyomo import run_pyomo, run_fast_pyomo
from IJKLM.run_jump import run_julia

This file has been truncated. show original

The JuMP code is

github.com

jump-dev/jump-dev.github.io/blob/master/assets/tutorials/2023-07-20-gams-blog.jl

# A [recent blog post](https://www.gams.com/blog/2023/07/performance-in-optimization-models-a-comparative-analysis-of-gams-pyomo-gurobipy-and-jump)
# by GAMS demonstrated a significant performance difference between JuMP and
# GAMS on a model they call IJKLM. We respond to this blog post by explaining
# the difference in performance and presenting an alternative JuMP
# implementation with asymptotically better performance. We also identify that
# differences in the input data format—not anything intrinsic to the
# respective libraries—explain the performance difference between the Pyomo
# model and the JuMP model. Finally, we respond to general claims about domain
# specific languages and the trade-offs of developing optimization models within
# a fully-featured programming language like Julia or Python.

# ## The model and input data

# Read the GAMS post for background on the model, as well as its mathematical
# formulation.

# Input data for the model is provided in three JSON files, labeled IJK, JKL,
# and KLM.

import JSON

This file has been truncated. show original