Skip to main content

Number of Threads affects MIP solution

Answered
Adrian Lipow
Gurobi-versary
First Comment
First Question

Hi,

I am solving a MIP using Gurobi 9.5. There seems to be an issue that causes Gurobi to find a drastically different objective value when a certain number of threads is set. In particular, I got the following objective values for different thread counts:

1 thread: -0.576121
2 threads: -0.576121
6 threads: -0.57609
7 threads: -0.576121
8 threads: -0.0217499
9 threads: -0.576121
16 threads: -0.576121
28 threads: -0.576121
56 threads: -0.576118
For all these solutions, all parameters were at their default settings, except for NonConvex, which was set to 2. When setting the MIPFocus parameter to any value other than 0, I was also able to get the objective value of -0.576121 while running with 8 threads.
The model that caused this behaviour for me is here: https://rwth-aachen.sciebo.de/s/crVyPB2DkPqfF7a (unfortunately, I can't provide any better MRE as I have no idea what is causing this behaviour).
Is there any (known) reason why this might be happening? I can't think of anything why the number of threads should be affecting the solution this much.
Cheers,
Adrian
0

Comments

5 comments

Date Votes
  • Eli Towle
    Gurobi Staff Gurobi Staff

    The result for Threads=8 is certainly suspicious. Can you please post the log output you see when solving the model with 8 threads? Also, are you using the latest version of Gurobi (9.5.2)?

    0
  • Adrian Lipow
    Gurobi-versary
    First Comment
    First Question

    Eli,

    The Gurobi version is 9.5.0, I unfortunately have no control over software updates at my institution, so switching to 9.5.2 on that machine isn't easily done.

    Here is the log:

    Gurobi 9.5.0 (win64) logging started Thu Jul 21 09:27:10 2022

    Set parameter LogFile to value "gurobi_debug.log"
    Read MPS format model from file it6.mps
    Reading time = 0.05 seconds
    minmax_output_bnf/: 840 rows, 899 columns, 3550 nonzeros
    Set parameter NonConvex to value 2
    Set parameter Threads to value 8
    Gurobi Optimizer version 9.5.0 build v9.5.0rc5 (win64)
    Thread count: 28 physical cores, 56 logical processors, using up to 8 threads
    Optimize a model with 840 rows, 899 columns and 3550 nonzeros
    Model fingerprint: 0x445e2538
    Model has 45 quadratic constraints
    Variable types: 434 continuous, 465 integer (465 binary)
    Coefficient statistics:
      Matrix range     [1e-02, 1e+02]
      QMatrix range    [1e-02, 1e-02]
      QLMatrix range   [1e+00, 1e+02]
      Objective range  [1e+00, 1e+00]
      Bounds range     [3e-01, 5e+04]
      RHS range        [3e-03, 9e+01]
    Presolve removed 150 rows and 150 columns
    Presolve time: 0.03s
    Presolved: 915 rows, 884 columns, 3875 nonzeros
    Presolved model has 135 bilinear constraint(s)
    Variable types: 419 continuous, 465 integer (465 binary)

    Root relaxation: objective -9.366221e+01, 1063 iterations, 0.04 seconds (0.02 work units)

        Nodes    |    Current Node    |     Objective Bounds      |     Work
     Expl Unexpl |  Obj  Depth IntInf | Incumbent    BestBd   Gap | It/Node Time

         0     0  -93.66221    0  224          -  -93.66221      -     -    0s
    H    0     0                       0.0123268  -93.66221      -     -    0s
    H    0     0                       0.0031427  -93.66221      -     -    0s
         0     0  -93.66221    0  354    0.00314  -93.66221      -     -    0s
         0     0  -93.66221    0  370    0.00314  -93.66221      -     -    0s
    H    0     0                      -0.0217499  -92.47724      -     -    0s
         0     0  -92.47508    0  354   -0.02175  -92.47508      -     -    0s
         0     0  -70.33434    0  357   -0.02175  -70.33434      -     -    0s
         0     0   -9.34311    0  149   -0.02175   -9.34311      -     -    0s
         0     0   -9.27170    0  147   -0.02175   -9.27170      -     -    0s
         0     0   -7.54211    0  153   -0.02175   -7.54211      -     -    0s
         0     0   -6.09764    0  153   -0.02175   -6.09764      -     -    0s
         0     0   -6.09764    0  152   -0.02175   -6.09764      -     -    0s
         0     0   -5.49716    0  153   -0.02175   -5.49716      -     -    0s
         0     0   -5.32572    0  154   -0.02175   -5.32572      -     -    0s
         0     0   -4.23841    0  155   -0.02175   -4.23841      -     -    0s
         0     0   -4.23841    0  154   -0.02175   -4.23841      -     -    0s
         0     0   -4.23841    0  154   -0.02175   -4.23841      -     -    0s
         0     0   -4.09206    0  150   -0.02175   -4.09206      -     -    1s
         0     0   -3.98261    0  152   -0.02175   -3.98261      -     -    1s
         0     0   -3.95741    0  150   -0.02175   -3.95741      -     -    1s
         0     0   -3.95741    0  154   -0.02175   -3.95741      -     -    1s
         0     0   -3.87786    0  150   -0.02175   -3.87786      -     -    1s
         0     0   -3.76887    0  150   -0.02175   -3.76887      -     -    1s
         0     0   -3.76887    0  325   -0.02175   -3.76887      -     -    1s
         0     0   -3.76887    0  163   -0.02175   -3.76887      -     -    1s
         0     0   -3.76887    0  163   -0.02175   -3.76887      -     -    1s
         0     0   -3.76887    0  164   -0.02175   -3.76887      -     -    1s
         0     0   -3.76887    0  166   -0.02175   -3.76887      -     -    1s
         0     0   -3.76887    0  168   -0.02175   -3.76887      -     -    1s
         0     0   -1.25973    0  166   -0.02175   -1.25973  5692%     -    1s
         0     0   -1.25973    0  164   -0.02175   -1.25973  5692%     -    1s
         0     0   -1.25973    0  168   -0.02175   -1.25973  5692%     -    1s
         0     0   -1.21637    0  166   -0.02175   -1.21637  5493%     -    1s
         0     0   -1.21632    0  166   -0.02175   -1.21632  5492%     -    1s
         0     0   -1.11653    0  166   -0.02175   -1.11653  5033%     -    1s
         0     0   -1.09918    0  165   -0.02175   -1.09918  4954%     -    1s
         0     0   -0.97465    0  165   -0.02175   -0.97465  4381%     -    1s
         0     0   -0.81027    0  160   -0.02175   -0.81027  3625%     -    2s
         0     0   -0.81027    0  161   -0.02175   -0.81027  3625%     -    2s
         0     0   -0.81027    0  161   -0.02175   -0.81027  3625%     -    2s
         0     0   -0.81027    0  161   -0.02175   -0.81027  3625%     -    2s
         0     0   -0.81027    0  161   -0.02175   -0.81027  3625%     -    2s
         0     0   -0.81027    0  161   -0.02175   -0.81027  3625%     -    2s
         0     0   -0.81027    0  161   -0.02175   -0.81027  3625%     -    2s
         0     2   -0.81027    0  161   -0.02175   -0.81027  3625%     -    2s

    Cutting planes:
      Implied bound: 40
      MIR: 19
      RLT: 50
      Relax-and-lift: 11

    Explored 2820 nodes (17647 simplex iterations) in 2.98 seconds (1.29 work units)
    Thread count was 8 (of 56 available processors)

    Solution count 3: -0.0217499 0.00314266 0.0123268 

    Optimal solution found (tolerance 1.00e-04)
    Best objective -2.174989470392e-02, best bound -2.174989470392e-02, gap 0.0000%
    0
  • Gwyneth Butera
    Gurobi Staff Gurobi Staff

    Also see the forum post MIP bound is greater than reported incumbent solution and reported solution is suboptimal.

    0
  • Eli Towle
    Gurobi Staff Gurobi Staff

    Thanks for posting the log file and for taking the time to report the issue. This inconsistency is caused by a bug that was fixed in Gurobi 9.5.2. Upgrading to the latest version of Gurobi should resolve the issue. Unfortunately, I'm not aware of any workarounds for Gurobi 9.5.0 that can avoid the issue (on this particular model, at least).

    1
  • Adrian Lipow
    Gurobi-versary
    First Comment
    First Question

    Ok, thanks for the help.

    0

Please sign in to leave a comment.

Didn't find what you were looking for?

New post