admin/LithosAnanake
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Stage H: DoE rebuilt on BIRTH/SEND/KILL, defs-only workloads; work heat floor on contact (FABRIC-3.8 §3.9, §3.13.7)

Browse Source 
Workload capsules the DoE births are definitions only; each defines WL-RUN
(its old birth-time work plus its entry word twice) and WL-DONE-ARGS. The
driver births N workers, SENDs each 'WL-RUN WL-DONE-ARGS SEND', waits in an
ordinary loop for N MU-DONE messages, KILLs them (parked idle), and emits
the marker. The serialized VM-EXEC round-robin is gone.

Kernel: no switch checkpoint after a payload's last word; work heat is the
assessed per-beat charge, and a VM gaining work is raised to the busy set's
floor (Captain Bob's ruling) -- fixes a live stall where a busy-waiting Hera
never yielded, without serializing batch starts.

Evidence: cfg-4 rows spread evenly over Hera and all 8 worker slots (44k
switches); full cfg-4 trials on amd64/aarch64/riscv64 end fail=0 with 8/8
KILL dead; Stage G interleave byte-identical on all three. Acceptance:
dict_hash unchanged, relay 3, idle growth 0.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Robert Allan James
2026-09-23 23:48:16 -04:00
co-authored by Claude Opus 5.5
parent 81def7b4a5
commit f152a91099
29 changed files with 165584 additions and 153 deletions
Download Patch File Download Diff File Expand all files Collapse all files
FABRIC-3.8.md
+99
View File
@@ -715,6 +715,24 @@ a loop. So would the console under load (already recorded under C-b).
conserved, and still appears in the DoE heat columns; it just no longer decides turns. The
prediction held on all three ISAs: see the Stage G entry in §6.
**Refined at Stage H, RULED 2026-09-23, Captain Bob: "floor on contact."** Two changes to what
work heat is, both forced by a live stall (Stage H entry, §6):
- **Assessed, not moved.** Work heat accrues each beat's assessed charge (the slope, in the
carry's sub-unit scale), whether or not the rest of the fleet held heat to move.
- As first built, it counted only heat actually moved.
- A busy-waiting Hera, with nobody else holding absolute heat, then stayed at 0 forever, tied
the workers, and never yielded.
- **Floor on contact.** When a VM with no work gains some (its kernel-Hermes queue goes from
empty to one message, and it is neither running nor parked mid-task), its work heat is raised
to the lowest work heat among the VMs that have work right now: the running one, and every
eligible participant.
- Without the floor, a fresh worker at 0 would run alone until it had done as much work as
Hera has since boot. The batch would start serially.
- No tuning constant.
- Rejected alternatives: decay/dissipation (needs a rate constant), and assessed charge
without the floor.
#### 3.13.6 Stage C, split
This splits the single C row in §4 in two:
@@ -1226,3 +1244,84 @@ explicit instruction.
- Hera's work heat grows only when other VMs hold absolute heat for her charges to pull
from. She runs cheaply while they hold ~0.
- Both are properties of the ruled quantity, recorded here, not changed.
- *(Superseded at Stage H: both consequences above turned out to matter, and the "floor on
contact" refinement replaced them. See §3.13.7 and the Stage H entry.)*
- [x] **Stage H — the DoE rebuilt on ordinary words; "floor on contact" (§3.9, §3.13.7).**
2026-09-23. Captain Bob: *"defs-only, no users factor yet, accept PRNG, go ahead with H"*,
and then ruled **floor on contact** when the first trial stalled.
**Changes, capsules:**
- The ten workload capsules the DoE births (`workload-0`, `-1-lite`, `-2`, `-3`, `-4`,
`-5-lite`, `-6`, `-7`, `-8`, `-9`) are now **definitions only**.
- Every top-level line that executed at birth is removed. That includes the heavy
self-exec tails and the mid-file `SEED`/wave/`RUN-ASYM`/`RUN-DUAL` runs in 2, 3 and 9.
- Each gains `: WL-RUN <exactly what it used to run at birth> <its entry word, twice> ;`.
That is the old trial's birth work plus its two `VM-EXEC` passes, now one job.
- Each also gains `: WL-DONE-ARGS S" MU-DONE" S" Hera" ;`.
- `workload-4` keeps its `4:dopi-ok` print inside `WL-RUN`.
- `workload-7` needed a second block for the two definitions. Block 2151 was unclaimed;
the conflict gate and `mkcapsule --lint` pass (37 files, 0 violations).
- `workload-1` and `-5` (full versions) are untouched, since the DoE doesn't birth them.
- `multiuser-doe.4th`:
- `WL-ENTRY`, `MU-TOUCH-WORKER/ALL` (the serialized `VM-EXEC` round-robin) and
`MU-CHECK-WORKER/ALL` are gone.
- New: `MU-DONE-COUNT`/`MU-DONE`, `MU-JOB` (the constant job text `WL-RUN WL-DONE-ARGS
SEND`), `MU-SEND-WORKER/ALL`, and `MU-WAIT-ALL` (an ordinary `BEGIN … WHILE REPEAT`;
Hera's own checkpoints drain the `MU-DONE`s).
- `MU-RUN-TRIAL` is now birth all, send all, wait, kill all, marker.
- A refused `SEND` counts as a failure **and** as done, so the wait can't hang on it.
**Changes, kernel:**
- `vm_core.c`: the between-top-level-words checkpoint (C-b) doesn't fire after the
**last** word (`vm_input_exhausted()`). The job's final word is the `SEND` to Hera, which
makes a decision due. Switching the worker out right there would park it "mid-task" with
nothing left, and `KILL` would refuse it. The trampoline decides a moment later anyway.
- **Floor on contact** (§3.13.7):
- `vm_physics_charge()` accrues the assessed slope into `work_heat_q48`.
- New `vm_physics_work_heat_floor()`.
- New `sk_vm_switch_signal_note_work_gained()`, called by `sk_hermes_send_one()` when the
recipient's queue goes from empty to one message.
**First trial, before the floor** (`logs/20260923-211253/amd64/`, kept): stalled.
- After 8,200+ beats the rows showed `switch_count_cumulative=0` and
`switch_current_slot=2` (Hera) throughout.
- Hera sat in `MU-WAIT-ALL`, with work heat 0 against workers at 0: a tie, so the
incumbent stays, forever.
**Evidence, as named in §4.**
- **Rows** (`logs/20260923-212353/amd64/`, cfg 4 with `HB-ON`, stopped deliberately once
captured because per-VM rows at every beat make a full trial take hours):
- `switch_count_cumulative` reached 44,297.
- Over the first ~7,600 rows, `switch_current_slot` was spread **evenly** across Hera
(slot 2: 851) and all eight workers (slots 5–12: 844–846 each).
- **Full trials** at cfg 4 (8 workers, uniform `workload-8`), rows off, one per ISA:
`logs/20260923-220439/amd64/`, `-222512/aarch64/`, `-231553/riscv64/`.
- Each ends with **`MU-TRIAL isa=<isa>run=999 cfg=4 rep=0 nw=8 mode=0 fail=0`**.
- **All 8 `KILL: muN dead`, 0 refused.** Every worker ended parked idle and killable.
- Times: amd64 ≈ 13 min, aarch64 ≈ 43 min, riscv64 ≈ 23 min.
- **Stage G's test, re-run on each boot under the refined rule:** byte-identical on all
three, `ABBAABBABAABBAABBABAABBA…` (A = 40, B = 40, **48** runs). It is finer than
Stage G's 16 runs because work heat now moves by sub-unit steps every beat.
**Acceptance,** in the same three boots after the trial:
- All reach `[zuse@Hera] ok>` with zero `UNKNOWN WORD`, failures and faults.
- The `dict_hash` set is unchanged from Stage G (word_count 533, M7.1a
`0x330b28ce0f4e1c75`).
- Mint → WIREBIND → `USE` → `1 2 + .` → `3`.
- Idle log growth: 0 bytes.
- Clean build on all three plus hosted, no new warnings.
**Known limits, recorded, not changed:**
- **A worker whose job errors never sends `MU-DONE`,** so `MU-WAIT-ALL` waits forever.
The error prints, but the trial produces no marker.
- **Worker execution errors are no longer counted.** `VM-ERROR?` can't see them, because
`sk_hermes_drain_checkpoint()` restores the VM's error flag after each payload (by
design: a bad message must not abort the enclosing context). `fail` now counts only
refused `SEND`s.
- **Hera's wait loop is a busy loop.** She takes a fair share of the CPU
(1/(N+1)), as the row tally shows.
- Per-VM, per-beat `HB-ON` rows make trials slow. One row in that capture shows a negative
(wrapped) `tick_interval_ns` for a VM a beat behind the others. That's a DoE-row
observation, not investigated.
- Stage F's remaining evidence (work proceeds with the console idle) is shown by Stage G's
interleave and by these trials.