nordisk/nornir

mimir-room bench — 2026-10-13

Box: dwarf (AMD Ryzen 5 7535HS, 12 threads, 61 GiB), load average 11–25 while running (shared box). SuiteSparse:GraphBLAS 7.4.0 (Ubuntu), release build. Runner: mimir-room <lake> bench [--universe] (src/bin/mimir-room.rs) and mimir-room <lake> diff --repo R --from N --to M. Lakes are COPIES under ~/d-scip/tmp/room-bench/ (EventLog::open creates missing tables).

Lakes:

• zz — znippy-zoomies, lane 9b run (--with-binary --std): 1 seq, 594 crates, 8 bins, std/core/alloc in the ring (~/d-scip/tmp/lgzm/state/lake).
• heimdal — heimdal's last 10 first-parent commits, re-consumed with the binwire producers (--with-binary; heimdal links no bin, so its binary stream is empty — the evidence says so). The older e2e-heimdal-v3 lake predates universe_manifest.rustc_release and does not open.

Build (fold → node table → matrices → provenance)

(Round 1. The ring-on rows are superseded by round 2 below: the ring is now scoped to what the seq depends on / links.)

¹ ring off but the binary stream links ring items: the provenance lookup folds the ring's 1.1 M node events once to find which linked ids the ring defines (same rule as EventLog::provenance). zz's derived provenance equals the ingest's own print exactly: {SourceOnly: 15982, BinaryOnly: 79593, Both: 13656}.

Every kind took the sorted fast path (ImportPath::SortedPack): edges are packed (from << 32 | to), sorted, deduplicated before the import. "open cached" = scope resolution + log identity + stamp read + cache hit.

Edges per kind (zz, ring on)

Node kinds (zz, ring on): param 342 024 · method 180 844 · generic 124 028 · impl 121 017 · mono_instance 77 308 · field 48 228 · const 43 393 · fn 26 077 · assoc_type 26 915 · variant 15 998 · struct 15 480 · ffi 4 346 · ffi_import 35 · …

Queries (Jikes kinds CALLS | USES_TYPE | IMPLEMENTS | FOR_TYPE)

Single seed = the most-called node (zz ring on: core … Try#branch(), fan-in 12 890). ×1000 = defined fns/methods at a fixed stride.

Time travel: diff seq N−1 → N (heimdal)

Seq 3 of this re-consume is an error row (the workspace did not load at that commit in the consumer's clone: a path dependency outside the repo), so seq 3 = seq 2's state, by the log's own rules.

Reading

• Cold open is the fold (skade scan + per-identity last event); matrices are ~1–2 % of it. Cached reopen: 3–16 ms.
• Keep Aᵀ cached (lane 7): the first dependents walk pays for the kind union and its transpose once (585 ms at 2.5 M edges), later walks 0.7–128 ms.
• Ring off + binary stream: the ring lookup for provenance dominates the cold open (1.2 s of 2.1 s on zz). Caching the ring's definitions per universe manifest would remove it (open item).

Round 2 (same day): scoped ring, ring definitions cache, Jikes down CONTAINS

Load average 9–10. Ring on now holds only the ring crates the scope's local crates DEPENDS_ON at their seqs, plus every ring crate its binaries link (panic_unwind / panic_abort are linked with no DEPENDS_ON naming them). zz has 594 dumped crates; dependency/build-only crates leave the room. Ring on and ring off now give the same provenance for every linked item (zz: BinaryOnly 79 593 · Both 13 656 either way; before the linked-owner rule 6 panic-runtime items differed).

The ring definitions cache is per process (per log identity + manifest snapshot): the first ring-off open with linked ring items still folds the ring once (1.3 s on zz); later cold builds skip it (2.0 s → 0.7 s).

Dependents now walk Jᵀ ∨ diag(impls)·CONTAINS (down from reached impl blocks to their items); "first" includes building the union, its transpose and that matrix.

Diff with changed (payload hash of nodes in both), heimdal:

changed 0 checked against the log: no + node event at seq b for a defined id open at seq a in these pairs (heimdal's edits there add/remove items or change bodies/consts, which are not node payload).