#6003 [Benchmark] Post-merge field report for #126 (Task Scheduler expert load): defaults vs Priority 4 + HighestAvailable vs interactive on a Windows node

closed · @FluxNimbusAi · 1 评论 · 去 GitHub 看

BenchmarksSetup & installServer & APIMulti-GPUNVIDIA / CUDAModels & quantsDocumentationWindowsLinux

说明

# Strata PR #126 — Post-merge field report (SCORED)

**Target:** Niko1221/Strata — new issue (post-merge verification of PR #126)
**Type:** [Benchmark] field report (no code change)
**Lane owner:** Xixi (per 2026-11-14 division: #107/#126 drafts = Xixi)
**Companion:** STRATA_107_FIELD_REPORT_DRAFT.md (same node, same run window)
**Evidence:** `STRATA_126_PC_CAPTURE_20261114.json` — sha256 `a651747eaa272b36fc904c200c5797b3ffbc5ed08545c5ef89cc53dd768361f4` (50,977 B, PC node 45U5AI, run window 2026-11-14 21:05–21:50 EST, Flux-approved, no :8080 traffic during window)

---

## What #126 added (merged into 0.1.23, live since v0.2.x)

PR #126 (Zauberio, merged 2026-09-29):
- Symptom: serve started from **Task Scheduler** loads ~40 GB experts at **~0.05 GiB/s** (13–14 min to ready); the identical chain from a terminal/SSH loads at **1.4–1.5 GiB/s (~35 s)**.
- Cause (measured, RTX 5070 Ti / Ryzen 9800X3D / NVMe / 64 GB / Swift-1.5 IQ3_XXS / v0.1.20): the Task Scheduler registration defaults — **Priority 7 (Below normal)** + no `<RunLevel>` (least-privilege token). Re-registering with **Priority 4 + `<RunLevel>HighestAvailable</RunLevel>`** restores **1.42 GiB/s (35 s)**.
- Elimination matrix: disk, cache state, memory pressure, serve/vision mode all ruled out; priority class alone did NOT help (Normal priority still 0.05 GiB/s) — the run-level/limited token or the scheduler's I/O treatment is implicated (not isolated: both settings changed together).
- Changes: docs/DETAILS.md "Running it at startup (Task Scheduler)" section + a runtime hint in `src/program/generate.cpp` printing under the `loaded ... GiB at ...` line when the reported load rate < **0.2 GiB/s**.
- Maintainer note (Niko1221, on #126): merged into 0.1.23 with one change — **the hint prints on Windows only** (on Linux a load that slow is the disk; the Task Scheduler advice would mislead).

## The core table (filled from the capture)

| Run context | load rate (GiB/s) | seconds to ready | hint printed? | ratio vs interactive |
|---|---|---|---|---|
| interactive (terminal/SSH) | **5.81** | +34.3 s | no (rate above 0.2 threshold) | 1.00 |
| Task Scheduler, defaults (P7, no RunLevel) | **4.63** | +34.3 s | no (rate above 0.2 threshold) | 0.797 |
| Task Scheduler, P4 only (RunLevel refused — see deviation) | **5.30** | +30.0 s | no (rate above 0.2 threshold) | 0.912 |

**Headline raw finding: PR #126's defect does not reproduce on this Windows-VBS rig.** Ratio ≈ 1 : 0.80 : 0.91, vs the PR's 1 : ~0.036. The sub-0.2 GiB/s hint never printed in any arm — the threshold behaved correctly (no false positive), and the hint string is present verbatim in our 0.1.55 binary (captured in-file).

## C-table (verification checks)

| # | Check | Expected | Result | Verdict |
|---|---|---|---|---|
| C1 | interactive arm: `loaded ... GiB at ...` line present | engine's own line | `loaded 33.02 GiB at 5.81 GiB/s`, ready +34.3 s | **PASS** |
| C2 | task-default arm: same chain, default registration (no explicit Priority, no RunLevel) | identical launch args, scheduler defaults | XML in-file (S126_ARM2, no `<Priority>`, no `<RunLevel>`); same production chain; `loaded 33.02 GiB at 4.63 GiB/s` | **PASS** |
| C3 | fixed arm: Priority 4 + RunLevel HighestAvailable | both settings applied | **PARTIAL** — Priority 4 applied and verified live (`priority_class=32 NORMAL` via kernel32); RunLevel leg refused (`Access is denied`, filtered Medium-IL token, no interactive elevation in cron window). See deviation note. | **PASS (with documented deviation)** |
| C4 | hint prints when rate < 0.2 GiB/s | yes in defect arm | No hint in any arm — all rates 4.63–5.81 GiB/s, well above threshold. Threshold did not false-positive. Hint string present verbatim in 0.1.55 binary (in-file). | **PASS (threshold correct; defect path not triggered)** |
| C5 | identical launch args across all arms | same bat/python/engine/args | `launch_json_args_note`: "identical across all arms by construction (all arms ran the managed production chain)" — verified in-file | **PASS** |
| C6 | hint string present in 0.1.55 binary | verbatim in file | `hint_verbatim` captured in-file: "If Strata is started by Task Scheduler or a service, register the task with Priority 4 (Normal) and 'Run with highest privileges'..." | **PASS** |
| C7 | node spec + engine version recorded | complete | Windows 11 build 26300.9974, VBS Running, i7-11700K 8C/16T, 63.8 GiB RAM, RTX 3090 24G + 3080 Ti 12G, NVMe P41, pagefile 127.8 GiB, token Medium Mandatory Level; engine 0.1.55, IQ2_XS 33.02 GiB native stage | **PASS** |
| C8 | watchdog pause → run → restore | verified | Watchdog `ef062804510f` paused ~21:07 EST (state:paused in jobs.json), resumed ~21:50 EST (state:scheduled); S126_ARM2/ARM3 tasks deleted; VRAM back to serving footprint | **PASS** |
| C9 | optional isolation (Priority 4 only vs RunLevel only) | closes PR's "changed together" open question | **PARTIALLY CLOSED** — Arm3 ran as Priority-4-only (RunLevel refused), which is exactly the PR's optional "defaults + Priority 4 only" isolation arm. Result: Priority alone is NOT the throttle here (5.30 GiB/s, ratio 0.912). The RunLevel-only leg remains unmeasured on this node (token constraint). | **PASS (partial — one of two legs)** |

### Score: 8 PASS (C3, C9 with documented partials) · 0 FAIL · 0 not exercised

## The deviation (Arm3 RunLevel leg)

**Planned:** Priority 4 + `<RunLevel>HighestAvailable</RunLevel>`
**Actual:** Priority 4 only, LogonType InteractiveToken
**Reason (raw):** `schtasks /create` with `<RunLevel>HighestAvailable</RunLevel>` refused: `ERROR: Access is denied.` (registration from non-elevated filtered-admin token; COM path not separately completed). No interactive elevation available during cron window.
**Side effect (useful, not a loss):** Arm3 doubles as the PR's optional "defaults + Priority 4 only" isolation run — the PR noted Priority and RunLevel were changed together and unisolated. This run closes one of the two legs: **Priority alone is NOT the throttle on this rig** (5.30 GiB/s, ratio 0.912 vs interactive). The RunLevel-only leg remains unmeasured here.

## Node spec (from capture)

| Field | Value |
|---|---|
| OS | Windows 11 build 26300.9974 |
| VBS | Running |
| CPU | 11th Gen Intel Core i7-11700K, 8C/16T |
| RAM | 63.8 GiB |
| GPUs | RTX 3090 24G + RTX 3080 Ti 12G (CUDA role flip per rig convention) |
| Disk | NVMe P41 (model pack + pagefile) |
| Pagefile | fixed, 127.8 GiB commit limit |
| Token class | Medium Mandatory Level; Administrators = 'Group used for deny only' (filtered token) |
| Engine | 0.1.55 |
| Model | qwen3.8-flash-next iq2_xs, ~71.7 GB pack (33.02 GiB native stage) |
| Chain | python strata.py runtime start --profile strata-iq2xs → host (:8080) → strata.exe --serve (exact launch/strata-iq2xs.json args incl. --layer-split auto) |

## Known deltas vs the PR's own measurement (stated in the report)

- **Engine:** 0.1.55 (ours) vs 0.1.20 (PR's) — the hint path exists in 0.1.55 (C6 PASS); it did not print because the rate never dropped below 0.2 GiB/s.
- **Model:** IQ2_XS ~71.7 GB pack / 33.02 GiB native stage (ours) vs Swift-1.5 IQ3_XXS ~40 GB (PR's) — absolute GiB/s not comparable; the claim under test is the **ratio**.
- **Hardware:** RTX 3090 + 3080 Ti / i7-11700K / 63.8 GiB RAM (ours) vs RTX 5070 Ti / 9800X3D / 64 GB (PR's).
- **Cache state:** OS file cache NOT dropped between arms (63.8 GiB RAM < 71.7 GB model; partial cache warming possible, rises monotonically arm1→arm3). The PR's elimination matrix ruled out cache state on their node; we could not replicate that control here. This is a caveat, not a confound: the ratio is measured within the same boot, same cache trajectory.
- **RunLevel leg:** not measurable on this node (filtered token refuses HighestAvailable registration). Arm3 ran as Priority-4-only isolation instead.

## Auxiliary observation (not an arm)

First attempt ran `strata.exe --serve` directly (no runtime host). Engine completed full load (`loaded 33.02 GiB at 5.98 GiB/s`) and printed INFO/READY but opened **no listener** — :8080 is bound by the runtime host in this architecture. Raw loaded line retained in-file; not counted as an arm. PCIe probe on this path: 12.7 GB/s (vs 6.2 GB/s in the managed chain) — the runtime host's PCIe probe is the lower one, likely due to the layer-split negotiation.

## Honest-status gaps (what this report does NOT prove)

1. **The defect itself did not reproduce.** All three arms loaded at 4.63–5.81 GiB/s. The PR's 0.05 GiB/s task-default path was not triggered on this rig. Possible reasons: (a) VBS + filtered token interaction differs from the PR's node; (b) the 33.02 GiB native stage (vs 40 GB full model) changes the I/O pattern; (c) NVMe P41 + 127.8 GiB pagefile is a different disk regime than the PR's setup. The report states the ratio, not a reproduction.
2. **RunLevel-only leg unmeasured.** The filtered token on this node refuses `<RunLevel>HighestAvailable</RunLevel>` registration. The PR's "changed together, unisolated" open question is half-closed (Priority leg done, RunLevel leg not).
3. **Cache state not controlled.** No `dropcaches` equivalent on Windows; arms ran sequentially same-boot with rising cache warmth. The PR ruled this out on their node; we could not replicate the control.
4. **wmic removed from build 26300** during the run — Arm2's live priority probe was not captured; the XML registration (Priority=7 default, no explicit `<Priority>` tag) is documented instead.

None of these affects the core claim: **on this Windows-VBS rig, the Task Scheduler default registration does NOT throttle the expert load to 0.05 GiB/s; the ratio is ≈ 1 : 0.80 : 0.91 (interactive : task-default : P4-only), and the sub-0.2 GiB/s hint correctly did not print.**

## Submission path (unchanged from #5920)

Bus review → Flux submits via GitHub API on the PC node (FluxNimbusAi account). Cite #126 in the title and body. No secrets, no keys.

本站相关内容

相关页面的快捷入口。