Problem
After the #1448 shaping lanes (#1599), one complete S18 ingest on 8 cores reaches 1.19× its 1-core throughput and 0.97 effective cores. The agreed #1448 core-use criterion is at least 2.0 on both.
Canonical encoding is now the largest region still on one core. Median region wall at 8 cores (three runs, candidate e0746156, evidence docs/development/evidence/core-scaling-1448/candidate/):
| Region |
Wall |
CPU / wall |
| complete ingest |
29.0 s |
0.97 |
| validate/seal/canonical_encoding |
8.30 s |
0.90 |
| …/canonical_encoding/adjacency_encoding |
4.46 s |
0.85 |
| validate/seal/shaping/shape_routing |
4.52 s |
0.83 |
| validate/append |
4.19 s |
0.86 |
| commit/publish |
2.50 s |
0.53 |
Two structural limits apply:
- Admission. ADR 0047 gives construction
compute_threads − reserve lanes. At 8 usable cores that is min(8, ceil(8 / 2)) − 1 = 3 lanes plus the coordinator.
- Model, not measurement. Even perfect 4-way encoding would save at most about 6 s at 8 cores, which gives about 1.3×. Reaching 2× also needs routing and append off the coordinator. Encoding is the largest single step and the next bounded one.
Scope
Encode the canonical topology and the adjacency index on admission lanes. The coordinator keeps the order of every published artifact and of every digest over it.
Published artifacts stay byte-identical (ADR 0038). The determinism suite's encoded digests do not move at any lane count.
Acceptance criteria
Relationships
Native sub-issue and blocker of #1448 (its core-use criterion). #1465 evaluated a DataFusion seam for encoding and found it 7.6% slower; this issue uses the retained machinery with lanes (ADR 0046). Routing and append parallelism are further candidates if this does not reach the criterion. Do not add them here.
Problem
After the #1448 shaping lanes (#1599), one complete S18 ingest on 8 cores reaches 1.19× its 1-core throughput and 0.97 effective cores. The agreed #1448 core-use criterion is at least 2.0 on both.
Canonical encoding is now the largest region still on one core. Median region wall at 8 cores (three runs, candidate
e0746156, evidencedocs/development/evidence/core-scaling-1448/candidate/):Two structural limits apply:
compute_threads − reservelanes. At 8 usable cores that ismin(8, ceil(8 / 2)) − 1 = 3lanes plus the coordinator.Scope
Encode the canonical topology and the adjacency index on admission lanes. The coordinator keeps the order of every published artifact and of every digest over it.
Published artifacts stay byte-identical (ADR 0038). The determinism suite's encoded digests do not move at any lane count.
Acceptance criteria
DETERMINISM_SAME_INPUTencoded digests are byte-identical with one lane, the maximum lane count and a permuted lane schedule. Query answers are identical at S18 and S20.Relationships
Native sub-issue and blocker of #1448 (its core-use criterion). #1465 evaluated a DataFusion seam for encoding and found it 7.6% slower; this issue uses the retained machinery with lanes (ADR 0046). Routing and append parallelism are further candidates if this does not reach the criterion. Do not add them here.