Simple diving for Branch-and-Bound #305
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: nicolas <nguidotti@nvidia.com>
Contributor
Author
|
/ok to test f9fb5ad |
Contributor
Author
|
/ok to test 041e104 |
Contributor
Author
|
/ok to test dd7e340 |
chris-maes
reviewed
Sep 18, 2025
| ? (lower_bound == 0.0 ? 0.0 : std::numeric_limits<f_t>::infinity()) | ||
| : std::abs(obj_value - lower_bound) / std::abs(obj_value); | ||
| if (user_mip_gap != user_mip_gap) { return std::numeric_limits<f_t>::infinity(); } | ||
| // Handle NaNs (i.e., NaN != NaN) |
Contributor
There was a problem hiding this comment.
The test user_mip_gap != user_mip_gap is equivalent to std::is_nan
chris-maes
reviewed
Sep 18, 2025
| i_t nodes_explored = 0; | ||
|
|
||
| while (node_stack.size() > 0) { | ||
| repair_heuristic_solutions(lower_bound, solution); |
Contributor
There was a problem hiding this comment.
Why do you need to repair a heuristic solution in the dive?
I would let the best-first search thread handle repairs.
chris-maes
reviewed
Sep 18, 2025
| lower_bound = lower_bound_ = root_node.lower_bound; | ||
| mutex_lower.unlock(); | ||
| gap = get_upper_bound() - lower_bound; | ||
| if (settings_.bnb_search_strategy == search_strategy_t::DEPTH_FIRST) { |
Contributor
There was a problem hiding this comment.
We are in a dive. Don't we know what the search strategy is?
chris-maes
reviewed
Sep 18, 2025
| stats_.nodes_unexplored = 0; | ||
| stats_.num_nodes = 1; | ||
|
|
||
| if (settings_.bnb_search_strategy == search_strategy_t::DEPTH_FIRST) { |
Contributor
There was a problem hiding this comment.
Why would the search strategy ever be depth first here?
I think we only want best first or multithreaded best first with diving
Contributor
Author
There was a problem hiding this comment.
@akifcorduk wanted to have a diving only option for submip (for this reason, I have the checks above). However, for simplicity sake, we have just a single search strategy: bfs + diving. Is that OK?
chris-maes
reviewed
Sep 18, 2025
|
|
||
| enum class search_strategy_t { | ||
| BEST_FIRST = 0, | ||
| DEPTH_FIRST = 1, |
chris-maes
reviewed
Sep 18, 2025
| mutex_gap_.unlock(); | ||
| } | ||
| } | ||
|
|
Contributor
There was a problem hiding this comment.
To confirm: in a multithreaded bfs + dives, the nodes explored and unexplored only comes from the BFS thread?
chris-maes
reviewed
Sep 18, 2025
Contributor
chris-maes
left a comment
chris-maes
left a comment
There was a problem hiding this comment.
Thanks for simplifying the PR. I left a few more comments. These are mostly nitpicks.
The only change I would suggest making before merging is to remove repairing solutions from the diving threads.
You should make sure to run the MIPLIB benchmarks and put the geomean performance improvement in the commit message. You should also run this on a MIP problem that is infeasible from the LP relaxation and one that is infeasible due to no integer solution and verify the code correctly handles these cases.
Signed-off-by: nicolas <nguidotti@nvidia.com>
Contributor
Author
|
/ok to test 413ee24 |
Signed-off-by: nicolas <nguidotti@nvidia.com>
Contributor
Author
|
/ok to test 53127f5 |
Contributor
Author
|
/ok to test 772f49e |
Contributor
Author
|
I got the following results for the MIPLIB2017 benchmark (arithmetic means of the primal gap): i.e., a 1.6% improvement over the main branch. The |
chris-maes
approved these changes
Sep 19, 2025
Contributor
|
/merge |
copy-pr-bot bot
pushed a commit
that referenced
this pull request
Sep 22, 2025
This PR introduces the following changes: - Implements a simple diving procedure - Allows the branch-and-bound to switch between different search strategies: `BEST_FIRST`, `DEPTH_FIRST` and `MULTITHREADED_BEST_FIRST_WITH_DIVING` - Refactor the branch-and-bound code such that the `solve` function is now organized into separated methods - Moved some commonly used variables to be member variables in the branch-and-bound solver. Authors: - Nicolas L. Guidotti (https://github.com/nguidotti) - Ramakrishnap (https://github.com/rgsl888prabhu) - https://github.com/ahehn-nv Approvers: - Gil Forsyth (https://github.com/gforsyth) - Akif ÇÖRDÜK (https://github.com/akifcorduk) - Trevor McKay (https://github.com/tmckayus) - Chris Maes (https://github.com/chris-maes) URL: #305
rapids-bot bot
pushed a commit
that referenced
this pull request
Oct 3, 2025
This PR implement a parallel branch-and-bound procedure, which is split into two phases. In the first phase, the algorithm will greedily expand the search tree until a certain depth and then add the bottom nodes to a global heap. The parallel expansion is implemented using `omp task`. In the second phase, some threads will explore the tree using best first search with plunging, i.e., they take the first node from the global heap and then explore the entire branch that starts on this node. Any unexplored node are insert into the heap. The remaining threads will perform deep dives in order to find feasible solutions. The solver keep a small heap contains the most promising nodes to perform the dives, which is keep in sync with the global heap. This PR also - Replace the `std::thread`-based parallelization in the strong branching with OpenMP in order to use dynamic scheduling. This ensures that all threads have similar amount of work and improve parallel performance. - Fixed invalid memory access when trying to access the status of a fathomed node. - Replaced `std::mutex` with `omp atomic` whatever applicable. - Added dedicated classes `dive_queue_t` and `search_tree_t` to store the diving heap and the search tree, respectively. This is an extension of #305. Closes #320. Closes #417. ## Benchmark results (MIPLIB2017): master branch (53d6e74) ``` Average Gap: 0.2174861712 ``` This PR: ``` Average Gap: 0.1989485546 ``` i.e., a `1.8%` improvement. In terms of the geomean of the gap ratio, this is equal to `1.62x`. Authors: - Nicolas L. Guidotti (https://github.com/nguidotti) Approvers: - Rajesh Gandham (https://github.com/rg20) - Chris Maes (https://github.com/chris-maes) URL: #412
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
9 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR introduces the following changes:
BEST_FIRST,DEPTH_FIRSTandMULTITHREADED_BEST_FIRST_WITH_DIVINGsolvefunction is now organized into separated methods