Serial-phase profile; increment 21 design (parallel refine) - #101
Merged
Merged
Conversation
…ominate it, scan stops near 5x Measurement only, battery, d6d6beb, 1 m benchmark on the quarter circle. - Phase timers: scan 0.303 s, split 0.127 s, rest 0.029 s of 0.462 s at 1 thread; Amdahl ceiling 2.9x; measured best 2.03x at 16 threads. - sample(1) + atos on an -O3 -g -flto build: flip test 13.2 % of refine (exact incircle 5.0 %: 9.1 % of tests, all exactly cocircular), legalise loop 6.3 %, topology writes 7.3 %, active sort 3.6 %, allocation 1.5 %, no locking. - Scan 4.95x at 8 threads: chunk imbalance 16.5 ms of 61.8 ms. - Per round (local instrumentation patch, not applied to the tree): 95 % of insertions share a footprint slot with another in the same round, 65 % a written slot; 51 % of marked triangles are deferred by the loop. Also the first stored bench.py battery run for the quarter at d6d6beb. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed to the measured third Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…options) Design only, no code or tests. Records Ola's 2026-09-27 ruling verbatim and what it leaves open (thread-count independence, one path or a size switch). - Prior art: legacy grep (nothing), literature from recall only, flagged for checking; no novelty claimed; the searches to run are listed. - Quick wins, all bit-identical: dynamic scan scheduling, an int64 incircle for all-node quads under an exact-frame condition, merge instead of sort for `active`; the tie classification to measure first. - Determinism levels L0-L3; recommends L1. - Serial-phase options: deterministic reservations, block colouring, batch insert plus parallel Lawson, DD (deferred to the mosaic work). - @Perf's six questions answered or placed; 21c measurements specified. - LOC per increment 21a-21d and invariant-critical suites; five questions for Ola. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… insertion prior art for option D Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… DD deferred) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n contract, conflict degree, frame orientation; profile README denominators and power records Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…1x, block-count provenance; 09-26 README points at the profile Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
What
No production code: evidence and design.
docs/benchmarks/2026-09-27/serial-profile/: the serial part is about a third of 1-thread refine (34 %), mostly Lawson legalisation; the scan stops near 5x from load imbalance (16.5 of 61.8 ms at 8 threads); ceiling 2.02-2.03x. Exact incircle fallback on 9.1 % of tests, all Cocircular. Insertions in the big rounds are ~3 nodes apart and each conflicts with 3-6 others. Raw data, scripts and an unapplied instrumentation patch are committed. Also the first storedtools/bench.pyrun (serial-profile-bench/, battery).docs/increments/21-parallel-refine.md: prior art (three core citations verified by web search; tile-parallel greedy insertion added for option D), determinism levels, quick wins 21a/21b (bit-identical), 21c measurements, 21d options (C batch-then-flip, or A1 deterministic reservations)..claude/agents/perf.mdand the 2026-09-26 README corrected to the measured figures (a third, not half; 2.1x battery / 2.2x AC).Review
@Reviewer: CHANGES REQUESTED twice (prose only), APPROVED at
267fc74.🤖 Generated with Claude Code