zsb/fidc-backtest-engine
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

docs(perf): record small selection CPU savings and full shared-input parity

Browse Source
This commit is contained in:
boris
2026-09-14 03:05:06 +08:00
committed by boris
parent d2aa16a2f0
commit 3a3091a2cf
2 changed files with 714 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/selection-risk-plan-performance-20260914.md
+126
View File
@@ -0,0 +1,126 @@
# Selection Risk Plan Performance
## Status
Candidate tested, not deployed. The change removes selection calls that have
no possible effect under the current frozen policy. It does not disable any
configured rule, execution-day check or strategy expression. Engine time falls
slightly in the measured cases; this is not the solution to the main remaining
data construction cost and is not a general whole-backtest speedup claim.
The independent [intraday-clock counterexample](intraday-clock-causality-20260914.md)
remains open. This work does not remove that test or its evidence, change the
execution clock, or turn day-level parity into full framework acceptance.
The published service stays at e81bf47/c98bcc3. Source d5b682c6 remains frozen;
research and signal work stay paused. No trading operation was submitted.
## Evidence Leading to the Change
The official HTTP diagnostic replay btr_1789322878865_2871869_0 preserved the
original canonical and result-store SHA. Ten bounded Boris-only GDB snapshots
showed source inventory, PreparedDayBuilder, factor normalization and price
series construction, followed by repeated selection risk calls. GDB pauses are
not normal performance measurements and snapshot counts are not flamegraph
percentages. Source/target PID, binary SHA and CPU/thread resources stayed fixed.
The diagnostic helper now shares the existing canonical executable policy with
the saved-run profiler: it accepts both audited build roots and immutable API
release directories, but not arbitrary paths. Seven related tests passed.
## Implementation and Correctness
StaticRiskRuleConfig reports whether selection has an enabled state rule or an
enabled nonempty blacklist. The strategy computes this once before iterating
candidate symbols. If no such rule exists, the old selection function would
always return None, so that no-op call is omitted. Explicit universe conditions,
market/factor checks and all buy/sell execution paths are unchanged.
The ten state flags are also shared with the existing missing-risk-state checks
to avoid maintaining three separate flag lists. Blacklist presence is kept
separate: a blacklist is not missing market-risk data. No cross-strategy cache,
strategy identifier, fixed date, trading time or account state is introduced.
Tests enumerate all 4,096 combinations of ten selection flags and blacklist
enabled/populated states. The flag list is checked against the serialized
configuration, so adding a selection field requires updating the activation
test. Further tests retain missing-state behavior and show that paused buys
and sells remain rejected when selection checks are inactive.
On 177: 805 core unit/integration tests passed (9 ignored), 448 runner tests
passed (9 ignored), 119 API tests passed (5 ignored), and 28 benchmark/profiler
tests passed. These counts do not resolve the independently recorded clock
failure, which is not part of this frozen committed test tree.
## Reproducible Shared-Input Method
Each of the twelve replays has a new process and a new private result root.
The official runner benchmark gained --shared-runtime-cache. It resolves the
explicit cache root from the declared Boris service, requires canonical private
storage, hashes existing inputs before and after, and refuses any changed or
removed original. This mode cannot invoke copied-input disposal.
All twelve runs used the same 9,257 files / 12,596,608,049 bytes. Their complete
input inventories, file identities and byte SHA values are equal. No new Arrow
or binary cache input appeared. No backtest result was reused. Hash preparation
and verification are outside the measured runner interval; this is a shared
warm-input test, not raw-disk cold IO. Unlike the earlier copied-cache method,
it does not allocate another approximately 2 GB per replay on the nearly full
SSD. Original inputs and every result remain intact.
The common execution interval is 2021-08-23 through 2025-11-17 with 10,000,000
initial cash and each case's unchanged frozen strategy/bundle. This is not five
complete execution years. CPU affinity and 8 Rayon / 16 Tokio threads match the
declared reference service; no global resource limit was increased.
## Measurements
| Case | Wall seconds | Source validation | Data preparation | Engine |
|---|---:|---:|---:|---:|
| Rotation control 1 | 30.986 | 12.244 | 8.445 | 8.790 |
| Rotation candidate 1 | 12.977 | 0.004 | 5.132 | 6.682 |
| Rotation control 2 | 13.275 | 0.209 | 5.190 | 6.739 |
| Rotation candidate 2 | 12.927 | 0.005 | 5.128 | 6.640 |
| Rotation candidate 3 | 12.926 | 0.006 | 5.113 | 6.664 |
| Rotation control 3 | 13.044 | 0.005 | 5.159 | 6.732 |
| Trend 40 control | 15.628 | 0.694 | 5.234 | 8.199 |
| Trend 40 candidate | 14.826 | 0.004 | 5.276 | 8.087 |
| Pullback 40 control | 13.776 | 0.005 | 4.893 | 7.374 |
| Pullback 40 candidate | 13.928 | 0.003 | 5.137 | 7.239 |
| Volume momentum 80 control | 18.176 | 0.004 | 4.585 | 11.154 |
| Volume momentum 80 candidate | 18.627 | 0.005 | 5.200 | 11.013 |
Rotation engine medians are 6.739 versus 6.664 seconds, approximately 1.1%.
The other paired engine reductions are approximately 1.4%, 1.8% and 1.3%.
These are small CPU-path improvements. Pullback and volume total latency did
not improve because their preparation times were higher. The first control's
Source wait and unexplained slower construction are recorded, not attributed
to this code or discarded to manufacture a large speedup. Peak RSS stays about
6.76-6.83 GiB; there is no significant memory reduction claim.
Each case matches its independent prior baseline for all six canonical
sections and store bytes: 21,393 / 29,776 / 31,862 / 51,300 fills. Result receipts,
runtime/strategy identities, physical manifests and 3,506 fact blocks were
verified. The shared input inventory SHA is in the acceptance record. Full
unaltered receipts remain on 177; the repository stores the compact verified
summary rather than repeating the 9,257-file inventory in every document.
## Remaining Work
Prioritize direct typed-column reuse during daily snapshot and DataSet
construction; approximately five seconds of preparation remain in these warm
cases. Do not skip normalization, NULL, adjustment or date/uniqueness checks.
Source cold-query and contract-validation latency remain separate tasks under
the Source freeze. The earlier cache-boundary candidate still needs its missing
cold/same-window acceptance, and this combined candidate has no HTTP publication
gate yet. Financial PIT, minute-clock behavior, signal lifecycle and UI factor
condition acceptance are not claimed complete.
- Candidate engine: d2aa16a2f0064297d0d8c931060646d66422e9d4.
- Candidate service source: 4e23c7558d8301ba697543c39d5604289bb82c53.
- Control runner SHA: b90886b80634c7565ca215fbe1f9ed0cbb5a6bd967373a9b1f6753be5164737d.
- Candidate runner SHA: 1bda2d3acc016ca5addbb12e33cfcc31a23ece562f1d7d1ff8a825fbc83873fb.
- Candidate API SHA: 30ac3b50996e1769c1d93bd5d302a23c4af7ebe773d3e8110ee278c44aeb9501.
- Evidence root: /srv/fidc/canonical/run/research/selection-risk-plan-20260914.
- HTTP diagnostic: /srv/fidc/canonical/run/research/http-phase-profile-20260914.
[Verified acceptance summary](evidence/selection-risk-plan-20260914/acceptance.json).