The direct loop before 046ce0be was architecturally incorrect. The fabric
wiring removed it and exposed the slowdowns, and the instruments (ASan,
UBSan, perf) locate them.
FABRIC-3 XLVIII.24 and .25 and both 11-17 plans are corrected in place,
and every measurement is kept. Mechanism A is the first speed-up found,
not a fix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RzZgrf7V7Zk2iczBXDRbm
166 lines
8.3 KiB
Markdown
166 lines
8.3 KiB
Markdown
# The regression at 046ce0be -- write-up, for ruling
|
|
|
|
**Corrected 2026-11-17, Captain Bob:** this is not a regression. "It was architecturally incorrect. That's
|
|
why we went back and did this wiring; to eliminate the defective architecture and locate slowdowns" -- with
|
|
ASan, UBSan, perf et cetera. The direct loop before 046ce0be was the defective architecture. The fabric
|
|
wiring removed it and exposed the slowdowns, and the instruments locate them. The measurements below stand;
|
|
where this document says "regression", read "the slowdown the correct architecture exposed".
|
|
|
|
Written 2026-11-17, after reading the record again in full (notes: `.claude/memory/reading-2026-11-17.md`).
|
|
**Nothing here is built, reverted or reset.** Every decision below waits for Captain Bob.
|
|
|
|
## 1. What and when
|
|
|
|
v4's boot became 1.5 to 2.4 times slower at **046ce0be** ("the mesh boot to spec -- Hera born empty in
|
|
the kernel's fabric; ... every node stepping together", 2026-11-13). The last good commit is
|
|
**f10d2bc8**, the commit before it.
|
|
|
|
Both commits ran the same POST, 539 cases, and both protocols passed. Each stage's time is taken from
|
|
the start of its QEMU log to the start of the next stage's log. A stage is a `clean` build, the boot and
|
|
the session.
|
|
|
|
| Stage | f10d2bc8 (`logs/20261113-142432`) | 046ce0be (`logs/20261113-170011`) | Ratio |
|
|
|---|---|---|---|
|
|
| amd64 | 14:25:04 to 14:26:51: **107 s** | 17:00:39 to 17:03:43: **184 s** | 1.7 |
|
|
| aarch64 | 14:26:51 to 14:30:20: **209 s** | 17:03:43 to 17:12:02: **499 s** | 2.4 |
|
|
| riscv64 | 14:30:20 to 14:32:05: **105 s** | 17:12:02 to 17:14:45: **163 s** | 1.6 |
|
|
|
|
**What these numbers do not prove on their own:**
|
|
- There is one run of each commit.
|
|
- The time includes the build.
|
|
- This host varies: the same commit took 146 s, then 175 s an hour later (Phase 7 plan, Task 2b).
|
|
|
|
Section 5 proposes the proof.
|
|
|
|
Two more measurements were taken in the previous session but are **not committed evidence**:
|
|
- A gprof profile of the host mesh boot, with `v4_fabric_step` at about 40% of the time spent in its own code (`v4/gmon.out`, untracked).
|
|
- A loop of a million, on hosted v3 and on Hera in the mesh:
|
|
|
|
| Run | Time |
|
|
|---|---|
|
|
| hosted v3 | about 172 ms |
|
|
| Hera in the mesh | about 1362 ms |
|
|
|
|
## 2. The mechanism, in the code
|
|
|
|
- **f10d2bc8.** One node runs the boot. For each instruction word, `boot_exec`
|
|
(`git show f10d2bc8:v4/system/boot.c`, lines 200-209) does three things:
|
|
- sets the node's have-mask;
|
|
- calls `v4_exec_step_word` once;
|
|
- checks one flag.
|
|
- **046ce0be onward.** Every instruction word Hera executes is one `v4_fabric_step` (`v4/src/fabric.c`,
|
|
lines 423-490), and the step walks all 16 places of the kernel's fabric (`SK_V4_PLACES`,
|
|
`kernel/src/v4/sk_v4.c:63`) in ten passes:
|
|
1. the have-mask, with an 8-port loop for each awake node;
|
|
2. `look_about` for each node that is not blocked, with another 8-port loop;
|
|
3. one `v4_exec_step_word` for each awake node, blocked ones included;
|
|
4. clearing `have_fetched`;
|
|
5. and 6. the two operation passes;
|
|
7. handing over writes;
|
|
8. handing over reads, which asks the console device on every step that Hera reads;
|
|
9. the gone-error pass.
|
|
|
|
After the step, `takes()` polls the console's queue (`v4/system/mesh.c`, `v4_mesh_step`, `v4_mesh_tell`).
|
|
|
|
So the cost of a step grows with how many places there are, not with how many nodes run. In the boot one
|
|
node runs, so nearly all of each step is overhead.
|
|
|
|
## 3. What the record says. I was wrong about this before.
|
|
|
|
**The rotation is your ruled design. It is not a scheduler.**
|
|
- ENGINE.md §3d, ruled "A" on 2026-10-05: "Every node that is not blocked runs; on one processor that is
|
|
each in rotation, an instruction word at a time, which is what 'all at once' is there and is not a
|
|
policy."
|
|
- MESH.md §2, ruling 7, and §4.3: "There is no choice of whose turn it is."
|
|
- `fabric.h` says the same.
|
|
|
|
Calling the lock-step itself a scheduler was my mistake. The scheduler I actually offered was the
|
|
withdrawn ruling "Hera by compudynamics now" (Phase 1 plan, 2026-11-13). The rotation stays as ruled.
|
|
|
|
**What the change broke is the cost.**
|
|
- **"An idle node does not spin"** (ENGINE §3d; MESH ruling 7). The node does not, but the fabric now
|
|
spins over it at every instruction word. MESH 9a.6 tested only the node: "its anti-clock ... does not
|
|
move over thousands of steps". It never tested what the step costs.
|
|
- **FABRIC-3.5 §XXII and §XXXV.2** rule out per-tick walks over every VM as a scaling wall: "we really will
|
|
be giving birth to potentially hundreds of VMs".
|
|
- **ENGINE §6** still lists as open "the node's safe moment for message delivery and switching". 046ce0be
|
|
did not settle it. It made every instruction word that moment, for every place.
|
|
|
|
**The slowdown was seen at each step and explained away.** It was never measured against f10d2bc8.
|
|
- **Phase 1 plan:** "The amd64 boot: 160 s (about 90 s before -- POST and the lines now go as messages)".
|
|
- **Phase 2:** "142 s (Phase 1: 153 s)", compared with the boot that was already slow.
|
|
- **Phase 7, Task 2b:** "The host slowed."
|
|
- **POST-PARITY §6:** "about 3 m 15 s at 539" was taken as the baseline.
|
|
|
|
Your standing order, "Find the TCG cost first" (POST-PARITY §6), is what this answers.
|
|
|
|
## 4. What it costs to go back to f10d2bc8
|
|
|
|
There are 33 commits after f10d2bc8:
|
|
|
|
| Commits | Where they are |
|
|
|---|---|
|
|
| 046ce0be | **pushed**, in PR 164 |
|
|
| Phase 2, e8e7ec8b | **pushed**, in PR 164 |
|
|
| The remote's head, ee28530b | **pushed** |
|
|
| 28 more: Phase 7 (the loops, the opcode physics, the physics words, the whole POST of 990 cases) and my five | **local only** |
|
|
|
|
So starting over from f10d2bc8 changes pushed history. That needs your explicit word. The choices are:
|
|
- reset the branch, with a force push;
|
|
- add revert commits;
|
|
- keep the history and redesign the step forward from HEAD.
|
|
|
|
## 5. Proposed proof, before anything is changed
|
|
|
|
**Intent:** show that the gap is the commit, not the host.
|
|
|
|
**How:**
|
|
- Use `git archive` to build f10d2bc8 and 046ce0be into the scratchpad. Nothing on the branch changes.
|
|
- Boot each one, back to back, on amd64, then on aarch64. That is four boots.
|
|
- Each boot runs in the foreground, one QEMU at a time, under `timeout 570`.
|
|
|
|
**The observation that proves it**, named before the run: 046ce0be's boot time is above f10d2bc8's by more
|
|
than the host's variance on both ISAs (about 1.7 times on amd64 and 2.4 times on aarch64), and the POST
|
|
line is the same.
|
|
|
|
**The logs** are committed and labelled.
|
|
|
|
## 6. What escaped, and how (a CAPA candidate, to rule)
|
|
|
|
- **What got through:** a 1.6 to 2.4 times regression passed the protocol.
|
|
- **Why:** no stage compares the boot's time with the last good commit, and acceptance has no criterion for cost.
|
|
- **Proposed control (not built):** the protocol records each stage's time and compares it with the last
|
|
passing PROTOCOL.md. The threshold is yours to set.
|
|
|
|
## 7. The uncommitted work you rejected, still untouched
|
|
|
|
**Untracked or modified files:**
|
|
- `v4/system/post_all_cases.c`, `post_v3_cases.c`, `post_cases.c`, `POST_FLOOR`
|
|
- `v4/tests/test_mesh_boot.c`, `v4/Makefile`
|
|
- `v4/include/v4/post.h`, `mesh.h`, `v4/system/mesh.c`
|
|
- `v4/tools/mkpost.py`, `post_all_rules.py`
|
|
- `docs/v4.0.0/POST.md`, `docs/v4.0.0/POST-PARITY.md`
|
|
- `logs/20261117-135508/aarch64/README.md`
|
|
|
|
**Logs not yet committed:**
|
|
- `logs/20261114-064855`, `-065851`, `-081210`, `-090737`, and `logs/20261117-135508`
|
|
- two DoE CSVs
|
|
|
|
**One false statement:** the addition to POST-PARITY §6 says "Ruled ... Whole POST in tests; boot runs
|
|
less". You withdrew that. It must not be committed as a ruling.
|
|
|
|
## 8. For ruling
|
|
|
|
1. **The write-up:** commit it locally as it stands, or correct it first?
|
|
2. **The proof of §5:** run it, yes or no?
|
|
3. **How far back:** start over from f10d2bc8, and if so by reset or by revert? Or redesign the step from HEAD?
|
|
4. **The step's design, whichever way 3 goes.** Keep the rotation as ruled, but make the step's work grow
|
|
with what runs and what moves, not with how many places there are. v3's own answers are on record:
|
|
- FABRIC-3 §XXVIII: a flag set, not a walk;
|
|
- FABRIC-0 GAP-A1: a virtual tick;
|
|
- FABRIC-3.5 §XLIII: the target takes its own messages on its own turn.
|
|
|
|
The design is to be written first, in the record (ENGINE §6, "the node's safe moment"), for your ruling.
|
|
5. **The rejected work (§7):** discard it, or keep any part of it? And which logs get committed as labelled evidence?
|
|
6. **The CAPA (§6):** record it now, or later?
|