Extremely different results for same model

ユーザーの入力を待っています。

anonymous123456788

2025年06月15日 04:23

• 
• 
• 

Hi all,

I’m seeing huge performance swings when I run the same minimization MIP (~20 M vars, ~1.4 M rows) on two similar cloud nodes.

Run    Cores    Key parameters    Best incumbent    Time to that incumbent    Total wall‑time (job killed)
1    32    TimeLimit≈1.2e7, Seed=123456    4 ,841 ,408    <1 min (heuristic)    48 h, no improvement
2    24    same    2 ,205.8    ~1 h    2 h

Notes

Same model file and parameter set; only hardware differs slightly (Run 1 has more and faster cores).

Presolved sizes and root barrier solutions are identical.

All other parameters (including ConcurrentMIP and Threads) are default.

Questions

Could the extreme gap really be due solely to the non‑determinism from multithreading/default ConcurrentMIP, even with a very large TimeLimit?

Full solver logs for both runs are available if needed.

Thanks!

関連

• Branch-and-Cut

1

• 

  Riley Clement

  • Gurobi Staff

  2025年06月25日 00:58

  Hi Bobak,

  Could the extreme gap really be due solely to the non‑determinism from multithreading/default ConcurrentMIP, even with a very large TimeLimit?

  Even with no-multithreading and deterministic solves Gurobi will likely run differently on non-identical machines https://support.gurobi.com/hc/en-us/articles/360045849232-Why-does-Gurobi-perform-differently-on-different-machines

  The performance variation seems extreme here though.  What do the coefficient stats in the log look like?  Are there signs of numerical issues?

  If you have not already I would see how our NoRel heuristic performs.

  - Riley

  0

サインインしてコメントを残してください。