Benchmarks

ripopt is benchmarked against Ipopt (native C++ with MUMPS linear solver) on three standard test suites and several domain-specific problem sets. All benchmarks run on the same hardware; both solvers receive identical problem data through the same Rust trait interface.

Hock-Schittkowski Suite (120 problems)

The HS suite is the classic test set for NLP solvers, covering small problems (n ≤ 15) with mixed equality/inequality constraints.

On 116 commonly solved problems (strict Optimal status required):

Run: make hs-run

CUTEst Benchmark Suite (727 problems)

CUTEst covers a wide range of problem types, sizes, and structures. Problems range from n=2 to n=10,000+.

On 525 commonly solved problems:

Run: make benchmark (full suite, ~2 hours) or individual problems:

cargo run --bin cutest_suite --features cutest,ipopt-native --release -- PROBLEM1 PROBLEM2

Where ripopt is faster

1. Small problems (n < 50). Stack allocation and cache-efficient dense BK factorization avoid sparse overhead. 2–5x per-iteration speedup over Ipopt.
2. Tall-narrow problems (m >> n, n ≤ 100). Dense condensed KKT via Schur complement reduces an (n+m)×(n+m) sparse factorization to n×n dense. 100–800x speedup on EXPFITC (n=5, m=502), OET3 (n=4, m=1002).
3. Better iteration counts. Mehrotra predictor-corrector with Gondzio centrality corrections (enabled by default) cuts iterations by 20–40% on many problems.
4. Fallback cascade. NE-to-LS reformulation, two-phase restoration, and multi-solver fallback recover problems Ipopt cannot solve.

Where Ipopt is faster

1. Large sparse problems (n+m > 5,000). Ipopt's Fortran MUMPS is 10–15x faster per factorization than rmumps on large systems.
2. Some medium constrained problems. A handful of problems (CORE1, HAIFAM, NET1) have higher per-iteration cost in ripopt's line search or fallback cascade.

Large-Scale Benchmarks

Both solvers use the same Rust trait interface; ripopt uses rmumps, Ipopt uses Fortran MUMPS.

Numbers above are fresh (v0.7.0). The larger problems (Poisson 2.5K, Rosenbrock 5K, Bratu 10K, OptControl 20K, Poisson 50K, SparseQP 100K) are not re-run for v0.7.0 — the stricter KKT backward-error probe causes Poisson 2.5K to exhaust max_iter and the full sweep is gated behind a separate investigation. Historical timings from v0.6.2 remain in benchmarks/large_scale/large_scale_results.txt snapshots.

Run: make benchmark

Domain-Specific Benchmarks