Cache optimization #2339
Merged
Cache optimization #2339
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
Contributor
There was a problem hiding this comment.
Pull request overview
This PR introduces significant performance optimizations to the MultifrontalSolver, focusing on cache locality, memory efficiency, and parallel execution granularity. The changes implement QR factorization for leaf cliques, a fused load-and-eliminate path, cached factorization results, and improved memory management, delivering substantial speedups on larger datasets while introducing architectural improvements for precomputed symbolic data.
Key changes:
- Implements QR elimination mode for high aspect-ratio leaf cliques without prior Hessian factors
- Introduces fused
eliminateInPlace(graph)that interleaves loading and elimination in a single traversal - Caches factorization results in separate
RSd_matrix for optimized back-substitution - Adds precomputed symbolic data support via
PrecomputedDatastruct andPrecompute()method - Increases parallel task threshold from 10 to 4096 to reduce scheduling overhead for small cliques
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated 10 comments.
Show a summary per file
| File | Description |
|---|---|
| gtsam/linear/MultifrontalSolver.h | Adds PrecomputedData struct, new constructor overload, Precompute() static method, eliminateInPlace(graph) overload, and state tracking flags |
| gtsam/linear/MultifrontalSolver.cpp | Refactors constructor to use precomputed data, implements fused elimination path, adds row counting for VBM sizing, updates parallel thresholds |
| gtsam/linear/MultifrontalClique.h | Adds QR mode, damping methods, RSd_ caching, prepareForElimination/factorize separation, IndexedSymbolicFactor row tracking |
| gtsam/linear/MultifrontalClique.cpp | Implements lazy SBM allocation, QR leaf factorization, damping support, cached factorization, separator-only SBM for QR updates |
| gtsam/base/SymmetricBlockMatrix.h | Adds utility methods for diagonal damping (addToDiagonalBlock, addScaledIdentity) |
| gtsam/linear/tests/testMultifrontalSolver.cpp | Updates tests to explicitly call load(), adds test for new eliminateInPlace(graph) API |
| timing/timeSFMBAL.h | Adds createOrderings helper function, minor formatting fixes |
| timing/timeMultifrontalSolver.cpp | Refactors benchmarks into separate functions, uses fused elimination path, adds BAL135 test |
Member
Author
Linux Timing (TBB)Similar conclusions. Chain improvement is massive, BAL less so, and speedups against GTSAM are less pronounced than on Mac, but this might just be because Linux is a 20 core machine with 32G RAM and a nice cache: Analysis
AfterBefore |
Member
Author
|
Unfortunately, without TBB on Linux, for BAL datasets, we have worse than GTSAM results - which stumps me. The overall picture seems to be: Chains:
BAL:
Some detailed profiling on Linux, single-thread, might at least bring that on par. |
ProfFan
reviewed
Jan 5, 2026
Member
Author
|
@ProfFan the |
Member
Author
|
PS for chains even single-threaded Linux is 4-6 times faster, it's really something about the massive fan-in for BAL cliques. |
ProfFan
approved these changes
Jan 6, 2026
Collaborator
ProfFan
left a comment
ProfFan
left a comment
There was a problem hiding this comment.
I'll do more profiling with this PR merged
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Overview
This PR introduces significant performance improvements and architectural updates to the MultifrontalSolver, focusing on cache locality, memory efficiency, and parallel granularity.
Key Optimizations:
QR Elimination for Leaves:
Cache Locality & "Fused" Path:
Improved Memory Management:
RSd_): The solver now explicitly caches the elimination resultParallel Task Tuning:
Architectural Changes
Precomputed Symbolic Structure:
Damping Support:
Timing and Analysis
Below are the timings in Mac, with TBB enabled. M1 Macbook with only 16G Ram.
The optimizations significantly scalability, at the cost of a small regression on very small problems.
Chain benchmarks
BAL benchmarks
Timings after MultifrontalSolver optimizations, and before:
After
Before