[BUG] Flaky test: c_api/TimeLimitTestFixture.time_limit/2 (MIP) overshoots time budget on constrained CI runners

[rgsl888prabhu]

Summary

c_api/TimeLimitTestFixture.time_limit/2 (the MIP parametric case using /mip/supportcase22.mps, time_limit=15s, CUOPT_METHOD_DUAL_SIMPLEX) fails intermittently on CI when the reported solve_time exceeds target_solve_time + excess_allowed_time (15s + 3s = 18s).

Failing assertion

cpp/tests/linear_programming/c_api_tests/c_api_tests.cpp:64

EXPECT_NEAR(solve_time, target_solve_time, excess_allowed_time);  // excess = 3.0

Observed CI failure

• solve_time = 18.398s, target 15s, tolerance 3.0s → over by ~0.4s.
• Solver itself respected the limit: log shows Explored 0 nodes in 15.01s. and termination status Time limit (5).
• CI environment per log: AMD EPYC 9554, only 12/12 threads visible (cgroup-limited on a 64-core host), 43 GiB RAM, NVIDIA H100 NVL, CUDA 13.1.

Root cause

Test reads cuOptGetSolveTime, which for MIP returns mip_solution_t::get_total_solve_time() → stats_.total_solve_time. That field is recorded at cpp/src/mip_heuristics/solver.cu:489 (and similar check_time_limit guards) using a timer_t that was started in cpp/src/mip_heuristics/solve.cu:306 — before Papilo presolve.

Therefore solve_time includes:

• Papilo presolve (OpenMP, ~2.59s in this run)
• cuOpt presolve (~1.59s)
• run_mip (LP root relaxation, primal heuristics, B&B)
• Post-B&B serial wind-down inside mip_solver_t::solve(): branch_and_bound_status_future.get(), sol.compute_feasibility(), sol.test_variable_bounds(...) + device sync

It does NOT include Papilo postsolve, the full_sol rebuild, or C-API wrapping (those run after total_solve_time is finalized).

The 3.4s gap between B&B's internal stop (15.01s) and total_solve_time (18.40s) is the post-join serial wind-down. On CPU-thread-constrained runners:

• Papilo presolve takes a larger share of the 15s budget (it's OpenMP-parallel — see cpp/src/mip_heuristics/solve.cu:412 passing settings.num_cpu_threads).
• B&B uses omp_get_max_threads() - 1 worker threads (cpp/src/mip_heuristics/solver.cu:360).
• The post-B&B compute_feasibility over the ~260K-row presolved problem becomes proportionally more expensive vs the budget.

Memory is not a factor at this problem size (~260K rows / 2.2M nz, peak well under 1 GiB).

Reproduction

• CI: failure observed on H100 NVL CI runner with cgroup-limited 12-thread visibility.
• Local (RTX 8000, compute 7.5): not reproducible — full c_api/TimeLimitTestFixture.time_limit/* set passes.

Suggested fixes (ranked)

1. Measure solve_time from inside the time-limited region only so the field reflects what time_limit actually controls. Excludes presolve/post-solve serial work that the user's time_limit setting cannot bound.
2. Per-case excess_allowed_time — keep 3s for LP cases, raise (e.g. 7s) for the MIP case to absorb host-side wind-down variance.
3. Increase target_solve_time for the MIP case so post-solve overhead becomes a smaller % of the budget.

(1) is the principled fix; (2)/(3) are pragmatic unblocks.

Mitigation in progress

Disabling the /2 case in a follow-up PR while this is investigated; the LP /0 and /1 cases remain enabled.

Related

There is a stale TODO at cpp/src/dual_simplex/presolve.cpp:1517 mentioning this test name, but it is about a different (numerical assertion) failure mode, not this timing one.