[Mac agent] CC engine default flip: tiebreak + sparse rows + -cc-roworder=auto

Round linked from #15 (hub). Priority 1 of the current plan: flip the CC engine defaults to the opt-in features merged on 2026-10-31.

Candidates

A: CC default rules excl,pack,orbf,parity,leaf,overlap,charge,tiebreak,sparse,sparseov,sparsemrv and -cc-roworder=auto as the default row order.
B (fallback): excl,pack,orbf,parity,leaf,overlap,charge,tiebreak and -cc-roworder=auto (no sparse tokens).

The previous default stays available as -cc-rules=excl,pack,orbf,parity,leaf,overlap,charge -cc-roworder=index; with those flags the output must be byte-identical to main. The MaxSAT default path is untouched (checked byte-identical).

Features: tiebreak (GH-162, #180), sparse/sparseov/sparsemrv (GH-181, #203), -cc-roworder=auto (GH-206, #215). The fixed mrv + -cc-roworder=bfs configuration (GH-183) is NOT part of either candidate (it fails the Tier 5 1.2x rule).

Gates (docs/OPTIMIZATION_LOOP_POLICY.md)

Tier 0: make test-cc, make test-cc-flags, answers equal the stems, every witness GF(2)-verified; old-default flags byte-identical to main.
Tiers 1-4: 5 seeds, main and candidate concurrently on pinned yfclab2 cores; per-tier geomean < 1 and no cell > 1.25. Tier 1 start-up floor reported to the PI, not decided here.
Tier 5: 6 h, default seed; no code solved by main's CC default may become more than 1.2x slower (watch LP_2610, LP_1428, LP_1122, LP_1768).
Results are grouped by the tier layout on main at judging time (a re-tiering data PR is in progress).

Merge main (134d756: #214 AMC data) into GH-217

GH-217: runners and Tier 0-5 evidence

Round result: PR #220, candidate A (tiebreak,sparse,sparseov,sparsemrv added to the default rules, -cc-roworder=auto as the default). The PR is open and not merged: the Tier 1 start-up floor needs the PI's decision.

Tier 0: PASS.
- make test-cc and make test-cc-flags (551 checks) pass.
- With the old-default flags, every CC log is identical to main; the MaxSAT default path is unchanged.
- 253 answer runs match their stems; 152 GF(2) witnesses check out.
Tiers 2-4 (5 seeds, concurrent): PASS. Geomeans T2 0.727, T3 0.769, T4 0.081; worst cell 1.023 (LP_714).
Tier 1:
- Concurrent panel: 0.969.
- Quiet interleaved panel: 1.031 (LP_442 1.076, about +2.5 ms from building the second auto half search).
- This start-up floor is the only item outside the rules.
- B (no sparse) has the same floor (1.035) and is weaker on T2/T4, so it is not an alternative.
Tier 5: no code that main's CC default solves is more than 1.2x slower. Worst is LP_1122 at 1.17x single run, 1.137x interleaved; LP_2610 1.14x / 1.078x; LP_1428 1.02x; LP_1768 1.04x. BB_360 takes 0.17x.
- BB_756 / BB_864 / BB_1080 reach lb 32 (main reaches 30); TN_496 reaches lb 27 (main reaches 23).
Bounds above the stem names, reached by both arms: TN_468 23, TN_612 24, TN_648_14 21, TN_684 22. Relabelled reproductions (-seed=1 / -seed=2) are running before any rename.

Evidence: optimization/experiments/GH-217/ on the PR branch.

Merge main (972b28c: #219 erasure opt-in, #221 re-tier) into GH-217

Merge pull request #223 from guluchen/data/gh217-lb-stems

GH-217: post-merge evidence (Tier 0 on b7b4126, new-layout panel, int…

Merge main (fecee8c: GH-217 stem renames #223) into GH-217

Update (2026-11-05). PR #220 is still open and not merged; it waits for the PI.

Merged main 972b28c (erasure opt-in, re-tier) and fecee8c into the branch.

Tier 0 on the merged binary passes:
- make test-cc: PASS;
- make test-cc-flags: 572/572;
- CC logs with the old-default flags are identical to main on the new tier0+1 (35 codes);
- answers: 550/550;
- GF(2) witnesses: 290 OK.

New tier layout, 5 seeds, concurrent (A geomean / worst cell):

T1: 0.855 / 1.111;
T2: 0.648 / 1.222 (LP_544, a 36 ms cell);
T3: 0.573 / 1.024;
T4 is empty.

Tier 1 (quiet interleaved): A geomean 0.898; LP_442 alone is 1.467.

That cell's wall time is quantized by the 10 ms parent poll (usleep(10000)): runs land at about 23 ms or about 33 ms.
A's +9 % user time puts every A run in the 33 ms bucket.
This is the only item outside the rules and needs the PI's decision.

Tier 5:

No solved code is more than 1.2x slower. On the merged binaries, the interleaved short LP cells are at most 1.112 (LP_1122).
New tier5 members: COLOR_331 0.12x; HGP_1105 0.06x.

TN stem lbs not reached in 6 h (coordinator question):

Not a code regression: main runs within 1 % of the [Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157 binary c3c315d in a same-session run, with identical nodes.
The machine ran about 10-20 % slower during the Tier 5 batch, and [Mac agent] CC engine: fractional-charge / edge-cover node lower bounds #157 had reached those lbs only at 21 100-21 560 s.
Forcing charge on (-cc-lb-auto=0) does not recover them: TN_648_10 and TN_648_14 both stay at lb 21 (lb times within 3 %); TN_496 becomes 1.35x slower.
So the flip does not change lb-auto.

Renames merged: DistQLDPC #223 (fecee8c) and QDistSAT #17 (7ea1aa96):

TN_468_4_23to29;
TN_612_2_24to40;
TN_648_14_21to28;
TN_684_2_22to29.

Each was reached by both main's and the candidate's CC default and reproduced with a relabelled run (same engine).

Tier 1 re-check on top of the GH-224 fix (#224, PR #226, main `82ca41d`)

#226 removes the 10 ms polling floor in the parent's wait loop (poll() on a SIGCHLD self-pipe instead of usleep(10000)).

Setup

Candidate A = this PR's head 42d7d1d with main 82ca41d merged in locally (d8cf2b39; not pushed). Compared with this PR, its src/ differs only by GH-224: parent waits for the solver child in poll() (removes the 10 ms timing floor) #226.
Reference = main 82ca41d.
CC defaults of each binary; 5 seeds x 9 repetitions.
Arms interleaved with rotating order on one pinned core per (code, seed), on yfclab2 cores 72-103. Each core and its SMT sibling were checked idle before every unit.
Ratio = median over seeds of the repetition medians, A / main.

Tier 1 (9 codes): geomean 0.870, worst cell 1.082 (LP_442), no cell > 1.25.

Code	main wall (s)	A wall (s)	A/main	A/main CPU
AMC_438_6_17	0.0746	0.0595	0.798	0.769
BB_186_10_14	0.0379	0.0332	0.875	0.988
BB_288_12_18	0.3464	0.2628	0.759	0.760
COLOR_127_1_13	0.0672	0.0332	0.494	0.450
GB_144_12_12	0.0744	0.0721	0.969	0.995
HGP_265_1_12	0.0331	0.0343	1.037	0.978
HGP_610_64_6	0.1482	0.1437	0.970	1.000
LP_340_56_8	0.0244	0.0252	1.033	1.058
LP_442_68_10	0.0271	0.0293	1.082	1.064

LP_442

The 1.467 in the earlier quiet panel was the 10 ms bucket.
With the fix, both arms' wall times are contiguous: main 24-37 ms, A 25-39 ms, 1 ms bins.
What remains is A's real extra CPU, about 6 % (the -cc-roworder=auto set-up noted earlier). It gives a 1.08x wall ratio.

Tier 0 (26 codes, information only): geomean 0.991, worst cell 1.186 (BB_144_12_12, 15.6 ms vs 18.5 ms), no cell > 1.25.

Correctness: 3150 runs, all rc 0, all answers equal to the stem. stdout differs only in timing-dependent line order on GB_144_12_12, and main varies the same way across repetitions.

The PR is not merged. The merge decision stays with the PI.

Merge branch 'main' into experiment/gh-217-cc-default-flip

Merge pull request #220 from guluchen/experiment/gh-217-cc-default-flip

Merge main (1239e9a: #220 GH-217 default flip, #226 polling fix) into G…

Navigation Menu

Repository navigation

[Mac agent] CC engine default flip: tiebreak + sparse rows + -cc-roworder=auto #217

Candidates

Gates (docs/OPTIMIZATION_LOOP_POLICY.md)

Tier 1 re-check on top of the GH-224 fix (#224, PR #226, main `82ca41d`)

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

[Mac agent] CC engine default flip: tiebreak + sparse rows + -cc-roworder=auto #217

Description

Candidates

Gates (docs/OPTIMIZATION_LOOP_POLICY.md)

Activity

guluchen commented on Nov 4, 2026

guluchen commented on Nov 4, 2026

guluchen commented on Nov 4, 2026

Tier 1 re-check on top of the GH-224 fix (#224, PR #226, main 82ca41d)

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

Tier 1 re-check on top of the GH-224 fix (#224, PR #226, main `82ca41d`)