perf(p6): add a reproducible baseline profile harness

- tools/perf/bench.py builds the release CLI and measures 7 .mq.md workloads
  three times with wall time and peak RSS, validating --json success.
- Check in the schema-tagged JSON report and interpretation; no bytecode/JIT/
  native backend was added, so P6 remains baseline-only, not complete.
- Update STATUS/changelog/tests README.

Gate: fmt, clippy -D warnings, 430 tests passed.
This commit is contained in:
cflmy committed 2026-11-20 23:16:02 +08:00
1 parent b020904199
commit 4c8e7fd8ab
6 files changed
+219 -1

No files matched your search

+2
View File
@@ -127,6 +127,8 @@ currently `0.1.0`; no release tag has been cut. Dates use ISO 8601.
`std:` resolution, tamper rejection and frozen-binary hash equality.
- P2 closure gate: add a pure no-Foreign/no-Plugin lane for C01–C06/C11/C12
(20 fixtures + stdlib table), source shortcut scan and golden comparison.
- P6 baseline profile: add `tools/perf/bench.py` and a checked-in 7-workload
JSON report; no bytecode/JIT/native backend is introduced.
- P4 cross-storage table equivalence: fixture 72 compares row-store Table
sort/filter output with a `.mq.md` columnar store cell-by-cell.
- P5 HTTP framing/client: `stdlib/net/http.mq.md` parses status/headers/chunked
+4 -1
View File
@@ -147,6 +147,9 @@ This is still a **local VM lane**, not a self-hosted GitHub Actions security URL
providers register additional resource kinds without parser/runtime/HostOp
changes; project-mode + name grant required; pure/strict refuse provider
kinds; C/Worker transport remains open.
- P6 baseline profile: `tools/perf/bench.py` measures 7 `.mq.md` workloads with
wall time and peak RSS; no optimizer/backend was added and P6 is not claimed
complete.
- Cross-storage table equivalence (P4, ADR-0032): fixture 72 compares T0 row
Table results with a `.mq.md` columnar List/Map store for sort/filter.
- Table stdlib algorithms (CM-A08, ADR-0032): `stdlib/data/table.mq.md`
@@ -207,7 +210,7 @@ behavior. This table records only what is in this tree now.
| P3 Resource/Stream ABI | **PARTIAL** | ADR-0025/0026 CM-A06 slices: opaque `Resource`, `resource.open/read/write/close`, stdout write in pure/project, granted project-file read and buffered project-file write in project, EOF/closed/handle errors, wire/host rejection; strict permits only stdio/stdout and refuses file/network resources. Fixture 66 + CLI/host tests. CM-A07 **Compat-C** is implemented: when a program calls `打印/print`, the loader injects the bundled `stdlib/io.mq.md` and the runtime routes to `io.写值` (all 15 value types, UTF-8 via `Memory`, `resource.write(stdout)`); `HostOp::Print` and the legacy `Value::display()` print path are removed, and a missing bundle fails with `kernel.unavailable`. Strict permits only the existing stdio console channel. Still missing: directories, network, generic provider registry, Table storage, strict file/network resources. |
| P4 Table logical/physical migration | **PARTIAL** | ADR-0032/CM-A08: `Value::Table` logical semantics unchanged (T0 `Vec<Vec<Value>>`); `stdlib/data/table.mq.md` provides select/filter/sort/append/set/join/group/`转GFM` as value-returning `.mq.md` functions with input immutability. GFM rendering and cross-storage equivalence (fixture 72: row Table vs `.mq.md` columnar store, same headers/rows/cells) are covered. No T1/T2/T3 physical storage, native column store, zero-copy views or Memory-backed table yet. |
| P5 third-party Provider / TCP / HTTP | **PARTIAL** | `stdlib/net/http.mq.md` + semantic fixture 71 implement HTTP/1.1 status/header/chunked parsing and request construction in `.mq.md`; `http_over_tcp.rs` runs a local HTTP exchange over `resource.open("tcp",...)`; no host HTTP/TLS and no external network test. ADR-0034 blocking TCP resource endpoint (`resource.open("tcp", "<ip>:<port>", ["read","write"])`, exact IP grant, project-only, 5s timeouts; loopback host+CLI tests). ADR-0033 in-process `ResourceProvider` v1: explicitly trusted providers register by kind, project-mode only, gated by `Capabilities.allow_resource_provider`, using the existing `resource.*`/`Result` ABI; a mock read/write provider is tested. No C/Worker provider transport, host DNS/TLS/HTTP implementation, UDP, clock/random/process/device providers, or provider package installation. |
| P6 performance / native backends | **NOT IMPLEMENTED** | no bytecode/JIT/GPU/QPU work |
| P6 performance / native backends | **BASELINE ONLY** | reproducible profile harness `tools/perf/bench.py` and [`p6-profile`](evidence/2026-11-20-p6-profile.md) for 7 workloads; no bytecode/JIT/columnar numeric backend or equivalence/optimization gate yet |
The P1 memory domain is pure computation: it needs no grant and is available in
`pure`, `project` and `strict` (bounded by the memory budgets). It adds no
+57
View File
@@ -0,0 +1,57 @@
{
"schema": "marqdo.perf-baseline",
"schema_version": 1,
"binary": ".cache/target/release/marqdo",
"iterations": 3,
"workloads": [
{
"name": "allocator-free-list",
"source": "tests/conformance/semantics/55-first-fit-free-list-allocator.mq.md",
"seconds_median": 0.0173,
"seconds_min": 0.017026568995788693,
"max_rss_kb_median": 8428
},
{
"name": "json-parser",
"source": "tests/conformance/semantics/62-json-array-parser.mq.md",
"seconds_median": 0.0188,
"seconds_min": 0.01880078506655991,
"max_rss_kb_median": 8960
},
{
"name": "table-sort",
"source": "tests/conformance/semantics/65-table-sort-stdlib.mq.md",
"seconds_median": 0.0152,
"seconds_min": 0.014747997978702188,
"max_rss_kb_median": 7476
},
{
"name": "bytecode-vm",
"source": "tests/conformance/semantics/63-memory-bytecode-vm.mq.md",
"seconds_median": 0.0155,
"seconds_min": 0.015298803802579641,
"max_rss_kb_median": 7692
},
{
"name": "matrix-multiply",
"source": "tests/conformance/semantics/70-matrix-multiply.mq.md",
"seconds_median": 0.0151,
"seconds_min": 0.014752828981727362,
"max_rss_kb_median": 7368
},
{
"name": "http-parser",
"source": "tests/conformance/semantics/71-http-parser.mq.md",
"seconds_median": 0.018,
"seconds_min": 0.017968795029446483,
"max_rss_kb_median": 8196
},
{
"name": "stdlib-print-demo",
"source": "stdlib/demo.mq.md",
"seconds_median": 0.0205,
"seconds_min": 0.02037098305299878,
"max_rss_kb_median": 9636
}
]
}
+53
View File
@@ -0,0 +1,53 @@
# RFC-CM P6 baseline profile
Date: 2026-11-20 (Asia/Shanghai)
Repository: `marqdo-cli`
Baseline: `b020904` plus this step's changes.
Scope: **baseline measurement only**; no bytecode/JIT/native backend was added.
## Harness
`tools/perf/bench.py` builds `marqdo-cli --release`, runs each workload three
times through `/usr/bin/time -f "%e %M"`, measures wall time with
`time.perf_counter`, verifies `--json` success and writes a schema-tagged JSON
report. The checked-in report is
[`2026-11-20-p6-profile.json`](2026-11-20-p6-profile.json).
Environment: Ubuntu 26.04 container, Linux x86_64, Rust 1.100.0, release
profile, 7 selected `.mq.md` workloads. These are small conformance workloads,
so most times are dominated by process/startup and Markdown parsing; they are
not throughput benchmarks of large scientific problems.
## Results (median of 3)
| Workload | Wall s | Peak RSS KiB |
|---|---:|---:|
| allocator-free-list | 0.0173 | 8428 |
| json-parser | 0.0188 | 8960 |
| table-sort | 0.0152 | 7476 |
| bytecode-vm | 0.0155 | 7692 |
| matrix-multiply | 0.0151 | 7368 |
| http-parser | 0.0180 | 8196 |
| stdlib-print-demo | 0.0205 | 9636 |
## Interpretation
* Startup + parse + load dominate these small workloads; the interpreter's
per-byte/per-call cost is visible mainly in the VM/matrix/table loops.
* `stdlib-print-demo` is highest RSS/wall because it loads the bundled stdlib
modules and formats all value types through `.mq.md`.
* The current release binary is a tree-walking Rust interpreter; no bytecode
compiler, JIT, columnar numeric backend or copy-elision pass exists.
* No claim of P6 pass: the RFC requires an optimized implementation that is
equivalent to the reference interpreter and reproducible performance/budget
evidence. This report only establishes the baseline and a reproducible
harness for future work.
## Reproduce
```bash
export PATH=/root/.cargo/bin:$PATH
export CARGO_HOME=/root/work/marqdo-cli/.cache/cargo
export CARGO_TARGET_DIR=/root/work/marqdo-cli/.cache/target
python3 tools/perf/bench.py > /tmp/marqdo-perf.json
```
+3
View File
@@ -185,6 +185,9 @@ grant, pure/strict, bad protocol version, reserved kinds and duplicate kinds
are rejected.
RFC-CM P2 closure (C01–C06, C11, C12) is pinned by
`tools/perf/bench.py` is the P6 baseline harness (not a test; it is not part of
the cargo gate).
`crates/cli/tests/p2_closure.rs`: 20 semantic fixtures run through
`run --mode pure --json`, sources are scanned for foreign/plugin/import
shortcuts, and value/stdout are compared with the goldens; the stdlib table
+100
View File
@@ -0,0 +1,100 @@
#!/usr/bin/env python3
"""Reproducible P6 baseline profile for selected .mq.md workloads.
Builds the release CLI, runs each workload three times through /usr/bin/time,
and writes a JSON report with wall seconds and peak RSS. This is a baseline
profile, not a claim that P6 optimization is complete.
"""
import json
import os
import statistics
import subprocess
import sys
import tempfile
import time
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
target_dir = Path(os.environ.get("CARGO_TARGET_DIR", ROOT / "target"))
if not target_dir.is_absolute():
target_dir = ROOT / target_dir
BIN = target_dir / "release" / "marqdo"
WORKLOADS = [
("allocator-free-list", "tests/conformance/semantics/55-first-fit-free-list-allocator.mq.md"),
("json-parser", "tests/conformance/semantics/62-json-array-parser.mq.md"),
("table-sort", "tests/conformance/semantics/65-table-sort-stdlib.mq.md"),
("bytecode-vm", "tests/conformance/semantics/63-memory-bytecode-vm.mq.md"),
("matrix-multiply", "tests/conformance/semantics/70-matrix-multiply.mq.md"),
("http-parser", "tests/conformance/semantics/71-http-parser.mq.md"),
("stdlib-print-demo", "stdlib/demo.mq.md"),
]
def run_once(relative: str) -> dict:
with tempfile.NamedTemporaryFile("r", delete=False) as handle:
time_path = handle.name
command = [
"/usr/bin/time",
"-f",
"%e %M",
"-o",
time_path,
str(BIN),
"run",
relative,
"--json",
]
env = dict(os.environ)
env["MARQDO_STDLIB_DIR"] = str(ROOT / "stdlib")
started = time.perf_counter()
output = subprocess.run(
command, cwd=ROOT, env=env, capture_output=True, text=True, timeout=120
)
elapsed = time.perf_counter() - started
timing = Path(time_path).read_text().split()
Path(time_path).unlink(missing_ok=True)
if output.returncode != 0:
raise RuntimeError(f"{relative} failed: {output.stderr}")
report = json.loads(output.stdout)
if not report.get("success"):
raise RuntimeError(f"{relative} reported failure")
return {
"seconds": elapsed,
"max_rss_kb": int(timing[1]),
}
def main() -> int:
subprocess.run(
["cargo", "build", "--release", "-p", "marqdo-cli"],
cwd=ROOT,
check=True,
)
report = {
"schema": "marqdo.perf-baseline",
"schema_version": 1,
"binary": str(BIN.relative_to(ROOT)),
"iterations": 3,
"workloads": [],
}
for name, relative in WORKLOADS:
runs = [run_once(relative) for _ in range(3)]
report["workloads"].append(
{
"name": name,
"source": relative,
"seconds_median": round(statistics.median(r["seconds"] for r in runs), 4),
"seconds_min": min(r["seconds"] for r in runs),
"max_rss_kb_median": int(
statistics.median(r["max_rss_kb"] for r in runs)
),
}
)
json.dump(report, sys.stdout, ensure_ascii=False, indent=2)
sys.stdout.write("\n")
return 0
if __name__ == "__main__":
raise SystemExit(main())