MIP Worker Refactor by Opt-Mucca · Pull Request #2886 · ERGO-Code/HiGHS

Opt-Mucca · 2026-03-02T16:12:36Z

No description provided.

…Worker

…main

… restart.

…a deque

…todos

fwesselm

Wow, @Opt-Mucca, looks like a lot of work! Good stuff.

I have added some minor comments (for the first 10 files). I will continue soon.

highs/mip/HighsCutGeneration.cpp

highs/mip/HighsCutGeneration.h

highs/mip/HighsCutPool.cpp

highs/mip/HighsCutPool.h

highs/mip/HighsCliqueTable.cpp

Opt-Mucca · 2026-03-03T09:38:48Z

@fwesselm I deleted the line in HighsCliqueTable and added const following all your suggestions

fwesselm

Some more unqualified comments from me. I will continue in the afternoon.

highs/mip/HighsLpRelaxation.cpp

highs/mip/HighsLpRelaxation.h

highs/mip/HighsLpRelaxation.cpp

cmatraki · 2026-03-03T22:08:09Z

I can't reproduce it with the latest version, but commit 4e29a0d with default options, finds app1-2 from miplib infeasible:

Running HiGHS 1.12.0 (git hash: 4e29a0d95): Copyright (c) 2025 HiGHS under MIT licence terms
Set option log_file to "HiGHS.log"
Number of PL entries in BOUNDS section is 13571
Number of BV entries in BOUNDS section is 13300
MIP app1-2 has 53467 rows; 26871 cols; 199175 nonzeros; 13300 integer variables (13300 binary)
Coefficient ranges:
  Matrix  [1e-05, 1e+00]
  Cost    [1e+00, 1e+00]
  Bound   [1e+00, 1e+00]
  RHS     [1e-05, 1e+00]
WARNING: Problem has some excessively small row bounds
Presolving model
53461 rows, 26859 cols, 199133 nonzeros  0s
39849 rows, 13559 cols, 275633 nonzeros  0s
34782 rows, 11265 cols, 206079 nonzeros  18s
Presolve reductions: rows 34782(-18685); columns 11265(-15606); nonzeros 206079(+6904) 
Objective function is integral with scale 1

Solving MIP model with:
   34782 rows
   11265 cols (11000 binary, 0 integer, 0 implied int., 265 continuous, 0 domain fixed)
   206079 nonzeros

Src: B => Branching; C => Central rounding; F => Feasibility pump; H => Heuristic;
     I => Shifting; J => Feasibility jump; L => Sub-MIP; P => Empty MIP; R => Randomized rounding;
     S => Solve LP; T => Evaluate node; U => Unbounded; X => User solution; Y => HiGHS solution;
     Z => ZI Round; l => Trivial lower; p => Trivial point; u => Trivial upper; z => Trivial zero

        Nodes      |    B&B Tree     |            Objective Bounds              |  Dynamic Constraints |       Work      
Src  Proc. InQueue |  Leaves   Expl. | BestBound       BestSol              Gap |   Cuts   InLp Confl. | LpIters     Time

         0       0         0   0.00%   -238            inf                  inf        0      0      0         0    19.0s
         0       0         0   0.00%   -231.7553454    inf                  inf        0      0      2       817    20.2s
         0       0         0   0.00%   -215.3802548    inf                  inf    12422    709     99      3956    25.6s
         0       0         0   0.00%   -206.5307028    inf                  inf    11112    924    141      6686    30.7s
         0       0         0   0.00%   -198.3939772    inf                  inf    11604   1131    194      9142    36.0s
         0       0         0   0.00%   -186.7867751    inf                  inf    12614   1501    242     11571    41.1s
         0       0         0   0.00%   -182.4413041    inf                  inf    11793   1307    396     97428   233.9s

0.3% inactive integer columns, restarting
Model after restart has 34702 rows, 11228 cols (10963 bin., 0 int., 0 impl., 265 cont., 0 dom.fix.), and 205562 nonzeros

         0       0         0   0.00%   -182.4413041    inf                  inf     1307      0      0     97428   237.3s
         0       0         0   0.00%   -182.4368136    inf                  inf     1307   1152      4     99837   239.6s
         0       0         0   0.00%   -178.2681812    inf                  inf    13238   2059     69    101864   244.8s
         0       0         0   0.00%   -172.3725509    inf                  inf    19141   2120    129    105295   249.9s
         0       0         0   0.00%   -161.889637     inf                  inf    19369   2349    186    107863   254.9s
         0       0         0   0.00%   -156.2504424    inf                  inf    24243   2471    255    109553   260.2s
         0       0         0   0.00%   -156.2504424    inf                  inf    24257   2239    328    161721   368.1s
        31       5        13  99.77%   -156.2504424    inf                  inf    24015   2239    341    165970   373.3s
        48       6        21  99.82%   -156.2504424    inf                  inf    24015   2239    349    169966   378.6s
        69       7        31  99.82%   -156.2504424    inf                  inf    24018   2239    359    174356   384.2s
       100       0        46 100.00%   inf             inf                  inf     2326   2421    376    183365   395.2s
       100       0        46 100.00%   inf             inf                  inf     2326   2421    376    183365   395.2s

Solving report
  Model             app1-2
  Status            Infeasible
  Primal bound      inf
  Dual bound        inf
  Gap               inf
  P-D integral      0
  Solution status   -
  Timing            395.23
  Max sub-MIP depth 4
  Nodes             100
  Repair LPs        0
  LP iterations     183365
                    0 (strong br.)
                    25562 (separation)
                    136761 (heuristics)

Note the the presolved model is different than what is obtained with the latest code, but is consistent with the numbers shown by 1.12.0.

Opt-Mucca · 2026-03-03T22:21:36Z

I can't reproduce it with the latest version, but commit 4e29a0d with default options, finds app1-2 from miplib infeasible:
Note the the presolved model is different than what is obtained with the latest code, but is consistent with the numbers shown by 1.12.0.

@cmatraki christ you are on it. Did you test over the entire MIPLIB? If so can you share any observations of the change? (I haven't really focused on performance, only focusing on correctness, but am properly interested in any opinions and user insights)

For app1-2: I've got that instance written down and can reproduce the incorrect status on an older commit locally (before the v1.13.1 merge). It was causing me enough headaches that I've just decided to push ahead for now and hope it pops up on a smaller instance. If you're thinking of looking into it: There's an incorrect bound change from propagating a cut (cut is completely valid), which I suspect is coming from stale activities, although I've failed to prove this (history of issues with this and multiple cutpools, but I thought I'd finally managed to fix it).

cmatraki · 2026-03-03T22:44:13Z

Did you test over the entire MIPLIB? If so can you share any observations of the change? (I haven't really focused on performance, only focusing on correctness, but am properly interested in any opinions and user insights)

Unfortunately no, I was just following along during development doing random tests for correctness. At some point everything started working but then I found this issue (and only this one) - just didn't know how to report it until now.

If you're thinking of looking into it

I'm not familiar enough with the code yet, I can only offer my own suspicion that the presolve changes since 1.12.0 mask the issue.

highs/mip/HighsMipSolver.cpp

highs/mip/HighsMipSolverData.cpp

fwesselm

Some more comments from me.

fwesselm · 2026-03-05T07:30:41Z

highs/mip/HighsMipWorker.cpp

+  return std::make_pair(feasible, transformed_solobj);
+}
+
+bool HighsMipWorker::trySolution(const std::vector<double>& solution,


Is this a copy of HighsMipSolverData::trySolution?

Yes...... It is just calling a different addIncumbent at the end of feasibility check (which requires the HighsMipWorker object). I could pass HighsMipWorker* worker = nullptr as a default argument, and call the global HighsMipSolverData::addIncumbent vs HighsMipWorker::addIncumbent depending on whether or not it's a nullptr. Any strong opinion for or against (or a different approach)?

Maybe another thing to look at after the initial submission.

highs/mip/HighsMipWorker.cpp

fwesselm · 2026-03-05T07:33:15Z

highs/mip/HighsMipWorker.cpp

+  return true;
+}
+
+std::pair<bool, double> HighsMipWorker::transformNewIntegerFeasibleSolution(


This looks like a modified version of HighsMipSolverData::transformNewIntegerFeasibleSolution. I guess it is not possible to re-use the existing code? Should this live in HighsMipSolverData?

After looking into this again, I guess it is now possible to reuse some of the existing code (I think I may have gotten lazy at this point). It would need:

Some nullptr default argument like in the above point

Some new statistics added to HighsMipWorker (repair LP iterations)

Some timer magic (delaying fixing all of this for now)

Would you want this to be unified into one function?

We can make a note and look at it later.

highs/mip/HighsPrimalHeuristics.cpp

fwesselm · 2026-03-05T07:46:54Z

highs/mip/HighsPrimalHeuristics.cpp

  HighsDomain& localdom = heur.getLocalDomain();
  heur.setHeuristic(true);

+  std::vector<HighsInt> intcols_;


I don't understand this code segment (intcols_, intcols, this->intcols). What is it doing?

I have now done this a few times: I don't want to add boolean switches at all points in the code, i.e., wherever intcols is accessed, so I simply define it once as a reference, which involves creating the thing it potentially is referencing, e.g., intcols_.
There might be some obvious C++ mechanic I could be using here.

fwesselm · 2026-03-05T07:49:19Z

highs/mip/HighsPrimalHeuristics.cpp

  if (heur.getCurrentDepth() > targetdepth) {
    if (!heur.backtrackUntilDepth(targetdepth)) {
-      lp_iterations += heur.getLocalLpIterations();
+      worker.heur_stats.lp_iterations += heur.getLocalLpIterations();


Not for now, but maybe worth having a getHeurLpIterations method at some point.

That would return a reference to heur_stats.lp_iterations?

Yes, I just saw similar updates in different places and wanted to ask.

highs/mip/HighsPseudocost.cpp

highs/mip/HighsPseudocost.h

highs/mip/HighsTransformedLp.h

highs/mip/HighsTransformedLp.cpp

fwesselm · 2026-03-09T12:56:07Z

highs/presolve/HighsPostsolveStack.h

  void undo(const HighsOptions& options, HighsSolution& solution,
-            HighsBasis& basis, const HighsInt report_col = -1) {
-    reductionValues.resetPosition();
+            HighsBasis& basis, const HighsInt report_col = -1,


Why is this needed? Don't the workers have separate postsolve stacks?

The workers don't have separate postsolve stacks. The current undo function is changing some global data structures, so you can't call undo on two different threads simultaneously.
This change is adding an argument so it does everything on copied data.

fwesselm · 2026-03-09T12:57:41Z

highs/presolve/HPresolve.cpp

    mipsolver->mipdata_->implications.rebuild(model->num_col_, newColIndex,
                                              newRowIndex);
-    mipsolver->mipdata_->cutpool =
+    // TODO: Find a sensible way to do this


Are you deleting the cut pool?

Yup! This was done before and I didn't see a need to change the workflow.

I had to get creative here because the std::atomic in HighsCutPool. I didn't want to change cutpool to a ptr in HighsMipData either (wanted to keep it as a reference). I'm happy with any ideas here as I think the current thing looks ugly (even if it works).

jajhall · 2026-03-15T14:52:53Z

@Opt-Mucca I've merged latest into this branch to see whether all the CI tests pass

galabovaa and others added 30 commits February 13, 2025 06:11

adding HighsMipWorker and HighsSearchWorker classes for a clean start

2490b29

undo changes from HighsSearchWorker from parallel-MIP

e3a4ec3

methods in HighsSearchWorker clean up

057f653

new classes initializing in HighsMipSolver

54d338f

added solution members, const mipdata ref, global cutpool to HighsMip…

95c1711

…Worker

added most getters

13f11c4

initialize lps and workers in mipdata

2d3de4d

using new search

267f01f

clean up highssearchworker

f43d22f

clean up

2073b0c

heuristics on, all tests passing

f22d9e1

WIP

14ac63d

go to refactor more!

bc04e0c

trying to init mipworker deque

3caa540

WIP

76f2b8a

primal heuristics, mipworker clean, init segfault on mipdata, highsdo…

89ee858

…main

comment out offending code, find another way to init first worker

0d4d922

master worker initialized in mipdata deque from mipsolver just before…

32bfbf2

… restart.

lprelaxation changed to ref in mipworker pointing to member of mipdat…

ec595c2

…a deque

highscliquetable check

18bc713

Merged parallel-MIP into this branch

69cbecf

solution added to highsmipworker, format

8b4485b

addIncumbent added in mipworker, not tested yet.

719253e

HighsSeparation done

d8abdc7

mipdata domain propagation off in HighsDomain

7f3e851

globaldom const in HighsDomain

90de58f

more on HighsRedcostFixing, HighsSeparation, HighsLpRelaxation more …

b17d159

…todos

HighsCutGeneration

8ed2844

initializing worker array OK: addIncumbent WIP

a55efe6

clean up

b9ad017

fwesselm reviewed Mar 3, 2026

View reviewed changes

Add comments from Franz

56b8567

fwesselm reviewed Mar 3, 2026

View reviewed changes

Opt-Mucca added 3 commits March 3, 2026 11:58

Comments from Franz v2

24a3726

Remove forward declaration and potential circualr dependency

08ad869

Formatting

a06cf54

Remove empty line

9314a91

fwesselm reviewed Mar 4, 2026

View reviewed changes

Comments from Franz v3

832ebd8

fwesselm reviewed Mar 5, 2026

View reviewed changes

Move assert outside loop

9863be4

jajhall assigned Opt-Mucca Mar 8, 2026

fwesselm reviewed Mar 9, 2026

View reviewed changes

Opt-Mucca and others added 2 commits March 13, 2026 15:03

Make globaldom const in translp

f28a935

Merge branch 'latest' into hmw-mt

9bedd5a

Opt-Mucca and others added 10 commits March 16, 2026 12:24

Fix cutpool aging bugs. Enable MIP timers in non-parallel

93b6992

Double backtrack plunge allowance

1e6813a

Add deterministic test

eaae314

Make thread sanitizer happy

0e069b0

Remove references

d2c4528

Revert std::atomic for parallel_lock

1637356

Format. Fix pseudocost bug. Move cutpool init

bdba2bd

Add HighsMipWorker.cpp to meson.build"

5f85038

Add const getter functions

8d81865

Merge branch 'hmw-mt-2' into hmw-mt

323a03f

Conversation

Opt-Mucca commented Mar 2, 2026

Uh oh!

fwesselm left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Opt-Mucca commented Mar 3, 2026

Uh oh!

fwesselm left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cmatraki commented Mar 3, 2026

Uh oh!

Opt-Mucca commented Mar 3, 2026

Uh oh!

cmatraki commented Mar 3, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

fwesselm left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment