Sub-MIP recombiner and B&B global variable changes by akifcorduk · Pull Request #259 · NVIDIA/cuopt

akifcorduk · 2025-08-06T12:03:19Z

This PR adds a new recombiner: sub-MIP. We use the B&B solver to solve the subproblem with the time limit of the recombiner. This PR also runs both local searches (line segment and FJ) one after another, instead of running either of them with 50% probability.

This PR also changes all the global variables in B&B to class member variables. This way, multiple B&B instances could run in parallel in the same process, including the sub-MIP recombiner.

hlinsen · 2025-08-06T15:18:15Z

+    CUOPT_LOG_DEBUG(
+      "n_vars_from_guiding %d n_vars_from_other %d", n_vars_from_guiding, n_vars_from_other);
+    this->compute_vars_to_fix(offspring, vars_to_fix, n_vars_from_other, n_vars_from_guiding);
+    auto [fixed_problem, fixed_assignment, variable_map] = offspring.fix_variables(vars_to_fix);


Do you know if it would cause size issues if we are to remove fixed variable or apply presolve on the fixed problem?

What do you mean by "removing fixed variable"?

I don't think it would cause any problems if we applied presolve as long as the conversions are done correctly.

If we can change the size of the CSR matrix

We are changing the size of the CSR matrix already. That's the fixed_problem.

chris-maes · 2025-08-06T16:13:09Z

-  global_variables::mutex_upper.unlock();
-  return upper_bound;
+  mutex_upper.lock();
+  const f_t upper_bound_ = upper_bound;


Thanks for making these changes @akifcorduk !

One nitpick. We have local variables for lower bound, upper bound, and gap. As well as the member variables that are shared between threads.

I think you used the convention of adding an underscore suffix to local variables (i.e. upper_bound_ or lower_bound_). I'm used to the exact opposite convention---that member variable have an underscore suffix. So I read lower_bound_ as a member variable of the class.

To avoid confusion, would you be ok if we used the underscore prefix for the member variables?

I agree with Chris :) To me at a first glance, prefix or suffix underscores suggest a member variable

Actually, I wanted to get rid of this local variables for reading. I wanted to convert all global/shared variables into atomic. I can do it in this PR, or later in another PR. What's your preference @chris-maes ?

Handled it, let's do the atomic changes in another PR.

Sounds great. Thanks

chris-maes · 2025-08-06T16:23:41Z

+      branch_and_bound_settings.integer_tol = context.settings.tolerances.integrality_tolerance;
+      // disable B&B logs, so that it is not interfering with the main B&B thread
+      branch_and_bound_settings.log.log = false;
+      dual_simplex::branch_and_bound_t<i_t, f_t> branch_and_bound(branch_and_bound_problem,


I think you should add a callback here. Since any feasible solution in the sub-MIP should be a feasible solution in the original problem. So you want to propagate solutions out of the sub-MIP.

Good catch. I wanted to implement getting the best primal solution, this way we can get all intermediate solutions.

chris-maes · 2025-08-06T16:24:04Z

+      fixed_problem.get_host_user_problem(branch_and_bound_problem);
+      branch_and_bound_solution.resize(branch_and_bound_problem.num_cols);
+      // Fill in the settings for branch and bound
+      branch_and_bound_settings.time_limit = sub_mip_recombiner_config_t::sub_mip_time_limit;


Probably instead of, or in addition to, a time limit, we should add a node limit to branch and bound

Why is this a case? Does returning early provide any value? Or does the B&B have diminishing returns in submip context when we run it more than certain number of nodes?

As we discussed offline, I think using a node limit just makes things more deterministic. We can switch later when the MIP heuristics are deterministic. Fine to use time limit for now.

aliceb-nv · 2025-08-06T12:28:46Z

+namespace diversity_config_t {
+static double time_ratio_on_init_lp              = 0.1;
+static double max_time_on_lp                     = 30;
+static double time_ratio_of_probing_cache        = 0.10;
+static double max_time_on_probing                = 60;
+static size_t max_iterations_without_improvement = 15;
+static int max_var_diff                          = 256;
+static size_t max_solutions                      = 32;
+static double initial_infeasibility_weight       = 1000.;
+static double default_time_limit                 = 10.;
+static int initial_island_size                   = 3;
+static int maximum_island_size                   = 8;
+static bool use_avg_diversity                    = false;
+static double generation_time_limit_ratio        = 0.6;
+static double max_island_gen_time                = 600;
+static size_t n_sol_for_skip_init_gen            = 3;
+static double max_fast_sol_time                  = 10;
+static double lp_run_time_if_feasible            = 15.;
+static double lp_run_time_if_infeasible          = 1;
+static bool halve_population                     = true;
+};  // namespace diversity_config_t


Was the loss of the const/constexpr qualifier intended?
Mutable global variables may cause issues

Yes this was intended. We are changing those constants depending on a environment config variable. They are wrapped in a namespace, so i doubt they will cause issues.

Okay, makes sense!
Although, ideally to make the semantics clearer I'd see this as a separate struct type, marked as a 'const' member and properly initialized at solver construction
Even if in this context it is unlikely to cause issues, global mutable variables are just inherently alarming especially in a large multithreaded codebase

If you keep them as global: please mark them as 'static inline' or 'extern'. As it stands, a static variable declared in a header will be instantiated with internal linkage in each separate translation unit; so if one of them is modified in a given .cu file, other .cu files won't see this change

Okay, i like the idea of member, i will do that.

aliceb-nv · 2025-08-06T16:17:04Z

-  global_variables::mutex_upper.unlock();
-  return upper_bound;
+  mutex_upper.lock();
+  const f_t upper_bound_ = upper_bound;


I agree with Chris :) To me at a first glance, prefix or suffix underscores suggest a member variable

chris-maes

Changes to branch and bound and pseudocosts look good to me.

Awesome work Akif. I'm excited to have sub-MIPping in cuOpt!

akifcorduk · 2025-08-22T16:21:42Z

/merge

This PR adds a new recombiner: sub-MIP. We use the B&B solver to solve the subproblem with the time limit of the recombiner. This PR also runs both local searches (line segment and FJ) one after another, instead of running either of them with 50% probability. This PR also changes all the global variables in B&B to class member variables. This way, multiple B&B instances could run in parallel in the same process, including the sub-MIP recombiner. Authors: - Akif ÇÖRDÜK (https://github.com/akifcorduk) Approvers: - Chris Maes (https://github.com/chris-maes) URL: NVIDIA#259

akifcorduk added 14 commits July 23, 2025 05:27

fix bug on initial solution size in the check

1fdd7d6

from dir baseline

55de7fd

make MAB 2 armed and move diversity into namespace

703f653

Merge branch 'initial_sol_bug' into paper_tests

4c82e8a

with mab

3abcbe4

with mab

76c4e2f

without mab

76e379c

run both LS methods together

8c70101

more parameters

d308303

implement submip

f261895

only do optimal ones

83ffe42

get best incumbent, add initial guess, reduce number of vars

5d46036

make global vars per class, set logging to false in mip recombiner

449b79b

remove lp check

3266646

akifcorduk added this to the 25.10 milestone Aug 6, 2025

akifcorduk requested a review from a team as a code owner August 6, 2025 12:03

akifcorduk added non-breaking Introduces a non-breaking change improvement Improves an existing functionality labels Aug 6, 2025

akifcorduk requested review from kaatish and rg20 August 6, 2025 12:03

make config log info

3c7cea9

rgsl888prabhu changed the base branch from branch-25.08 to branch-25.10 August 6, 2025 14:25

hlinsen reviewed Aug 6, 2025

View reviewed changes

chris-maes reviewed Aug 6, 2025

View reviewed changes

Comment thread cpp/src/dual_simplex/pseudo_costs.hpp Outdated

chris-maes reviewed Aug 6, 2025

View reviewed changes

Comment thread cpp/src/dual_simplex/pseudo_costs.hpp Outdated

chris-maes reviewed Aug 6, 2025

View reviewed changes

Comment thread cpp/src/dual_simplex/pseudo_costs.cpp

chris-maes reviewed Aug 6, 2025

View reviewed changes

aliceb-nv reviewed Aug 6, 2025

View reviewed changes

akifcorduk added 4 commits August 7, 2025 07:48

handle review comments

152e24f

add submip callback and add intermediate solutions to the population

2aa28a2

add suffix to member vars

50f60d0

make diversity config a member var

b5dbb5d

akifcorduk requested review from a team as code owners August 8, 2025 13:42

akifcorduk requested review from tmckayus and removed request for a team August 8, 2025 13:42

akifcorduk added 2 commits August 8, 2025 06:43

Merge branch 'branch-25.10' of github.com:NVIDIA/cuopt into paper_tests

1221aad

remove cmake comments

d2f1d98

akifcorduk removed request for a team and tmckayus August 8, 2025 13:44

akifcorduk added 6 commits August 8, 2025 09:45

fix B&B name issues

798ad68

Merge branch 'branch-25.10' of github.com:NVIDIA/cuopt into paper_tests

89f98cc

add presolve and remove an assert from factorization

4748007

move initialize

368f940

Merge branch 'branch-25.10' of github.com:NVIDIA/cuopt into paper_tests

21a4471

fix presolve data issue

22f7ac1

chris-maes approved these changes Aug 21, 2025

View reviewed changes

akifcorduk added 4 commits August 21, 2025 02:05

Merge branch 'branch-25.10' of github.com:NVIDIA/cuopt into paper_tests

4d2935a

remove clearing of the queue

85af648

Merge branch 'branch-25.10' of github.com:NVIDIA/cuopt into paper_tests

d7cb346

fix rounding bug and add a tesT

07a2189

rapids-bot bot merged commit 2283fd5 into NVIDIA:branch-25.10 Aug 22, 2025
142 of 144 checks passed

akifcorduk mentioned this pull request Aug 26, 2025

[BUG] Python test failure in diversity_manager_t::set_simplex_solution #236

Closed

Conversation

akifcorduk commented Aug 6, 2025

Uh oh!

hlinsen Aug 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

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!

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!

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!

aliceb-nv Aug 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

Uh oh!

chris-maes left a comment

Choose a reason for hiding this comment

Uh oh!

akifcorduk commented Aug 22, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

hlinsen Aug 6, 2025 •

edited

Loading

aliceb-nv Aug 7, 2025 •

edited

Loading