Repository navigation
perf(multibody): full Newton for cables in the monolithic step - #113
Conversation
|
@codex review |
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 29a62360e4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
With OPTIONS modified_newton True every finite-EI cable reused its tangent within a step. A cable in the monolithic step of the multibody march is restored and stepped again from the same start at every coupling iteration, and there the reuse cost more than it saved: mixed_body_rods_points.dat ran in 7.0 s with the option against 5.8 s without (5.7-6.3 s with the cables alone kept on full Newton). Such a cable now keeps full Newton; the other solvers of the deck keep the option. The deck runs in 5.6 s with it. The documented trade-off of the option is measured again on the examples (one thread): 4-20 % faster on finite-EI cables, free-hanging or with a touchdown; 12 % slower on an EI = 0 cable laid on a frictional seabed in a current (a 600 s march), within a few per cent elsewhere on EI = 0 systems. The former "about 1.5 times slower on finite-EI lines on the seabed" no longer holds: the touchdown decks run 4-12 % faster with it. Results without the option are unchanged. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
|
@codex review |
29a6236 to
7d89bd2
Compare
|
Codex Review: Didn't find any major issues. 👍 Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
The [Unreleased] entries of the open changes are added together in one change, so that they do not conflict with each other. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
|
@codex review |
|
Codex Review: Didn't find any major issues. Chef's kiss. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
The staggered march (bodyScheme staggered) steps each cable once per step, without the repeated coupling iterations that make tangent reuse dearer, so a cable attached to a moving object keeps OPTIONS modified_newton there. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
|
@codex review |
|
Codex Review: Didn't find any major issues. You're on a roll. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
With
modified_newton True, every finite-EI cable reused its tangent within a step. In the monolithic step of the multibody march, a cable is restored and stepped again from the same start at every coupling iteration, and there reuse cost more than it saved.mixed_body_rods_points.datmodified_newton True, beforemodified_newton True, afterSuch a cable now keeps full Newton. The other solvers in the deck keep the option.
The option's documented trade-off is re-measured on the examples, one thread:
The former claim, "about 1.5 times slower on finite-EI lines on the seabed", no longer holds: the touchdown decks run 4–12 % faster with the option.
Results without the option are unchanged. Full CTest passes on Release, 365/365 on the default kernel and on Prescott.