Mittelmann interior-point and simplex benchmarks from http://plato.asu.edu/bench.html
Keywords, tags: linear programming, benchmark, Mittelmann, HiGHS, python, scipy, sparse-matrix
This plot shows the runtimes of various LP interior-point and simplex programs on Mittelmann's benchmark problems.
To plot similar rows / similar columns near each other,
I sort rows and columns by their median times.
(Medians are scale-invariant --
the sort order is the same for x
, x^2
, log x
.)
Comments on better ways of doing this would be welcome.
Everyone knows that commercial solvers such as Gurobi are waaay faster than any opensource solver. Mittelmann's data supports this. But I don't know which of these dozen solvers
- are really opensource (not "academic opensource")
- have doc and support that I like (subjective)
- talk to python scipy.sparse .
In real-world optimization, an optimizer engine is part of a flow, a cycle, a process:
- input: map a problem to a sea of numbers
A b c ...
- run the optimizer -> a sea of numbers
x[...]
- map back: make the solution
x[...]
understandable, with plots and talks - check that there are no mistakes along the way.
Amdahl's law certainly applies here.
Interactive Visualizations of Mittelmann benchmarks
lpgen34.py
generates LP problems 4n^3 x n^4, with 4 1s in each column and n in each row.
cheers
-- denis 26 Jan 2021
Theory and practice are closer in theory than they are in practice.