diff options
| author | Somhairle H. Marisol <[email protected]> | 2026-09-29 04:23:26 +0800 |
|---|---|---|
| committer | Somhairle H. Marisol <[email protected]> | 2026-09-29 04:23:26 +0800 |
| commit | f681867692f575b078f46afaf07060faed457461 (patch) | |
| tree | 4ab11638357891257703e55f5aa790868e902e3c | |
| parent | 2d9e4d845019b6c8793633258a45b91e0358e0e9 (diff) | |
| download | living-village-f681867692f575b078f46afaf07060faed457461.tar.gz | |
p59: Sim 多世界并存后台压测(内存/GC 增长曲线证据单,只读)
目标:量化「多世界并存 × 天数」的内存/GC 增长,收口 P49 范围外项。
口径:新增 Headless `--profile-multiworld <worlds> <days> <npcs> <sampleEveryDays> <seedBase>`,
N 个 Sim 世界同时存活,每轮 Parallel.For 各推进 sampleEveryDays 天并采样一行
(worlds/events/rumors/annals/memory_entries/gc_live_mb/working_set_rss_mb/elapsed_s,
行尾 flush);未改 Kernel/Sim 任何语义(Sim.initialWorldN / Sim.step 原样调用)。
标定:1 世界 × 2 天探针 = 34.2s(与 overnight-status 的 34.6s 一致),据此定 N=10 × 20 天。
运行:独立 systemd --user transient unit living-village-multiworld.service
(scripts/profile-multiworld.sh,StandardOutput=append: + exec 2>&1,state.json 记
pid/unit/log/commit/参数/起止/exit)。主运行 wall=1024.4s(17.1min < 40min),exit=0。
结论(artifacts/perf-multiworld/20260928T195839Z/README.md):
- 随世界数近似线性,未超线性:N=1 vs N=10(同日 D=20)annals/memory_entries 恰 10×、
rumors 9.67×、gc_live 8.81×(每世界 1.23MB,共享运行时开销摊薄)。
- RSS 有界无泄漏:N=10 39→峰值123(day8)→稳定104~112MB;每 2 天一轮墙钟恒 101~105s。
原因是历史集合有界(rumorRetentionDays=3 / rumorCapacity=16384 / maxAnnalEntries=128 /
Mind.Memory 上限),约 day4 后进入稳态。
- 两轮世界0 digest 一致(2364CB74…)→ 并存+并行不改变单世界语义。
- 数字未越界 → 明确「不修」;只读本 slice 不修 Kernel/Sim。
回归:Desktop 304/304 + Kernel 119/119;--performance-baseline ×3 digest
953775FAEB2F… 不变,performance_determinism=PASS。未 push。
| -rw-r--r-- | artifacts/perf-multiworld/20260928T195839Z/README.md | 113 | ||||
| -rw-r--r-- | artifacts/perf-multiworld/20260928T195839Z/acceptance.log | 15 | ||||
| -rw-r--r-- | artifacts/perf-multiworld/20260928T195839Z/control-n1.log | 13 | ||||
| -rw-r--r-- | artifacts/perf-multiworld/20260928T195839Z/state.json | 16 | ||||
| -rw-r--r-- | deploy/living-village-multiworld.service | 40 | ||||
| -rwxr-xr-x | scripts/profile-multiworld.sh | 80 | ||||
| -rw-r--r-- | src/LivingVillage.Desktop.Tests/M3BatchOutputTests.fs | 7 | ||||
| -rw-r--r-- | src/LivingVillage.Headless/BatchOutput.fs | 17 | ||||
| -rw-r--r-- | src/LivingVillage.Headless/Program.fs | 64 |
9 files changed, 364 insertions, 1 deletions
diff --git a/artifacts/perf-multiworld/20260928T195839Z/README.md b/artifacts/perf-multiworld/20260928T195839Z/README.md new file mode 100644 index 0000000..f5dc676 --- /dev/null +++ b/artifacts/perf-multiworld/20260928T195839Z/README.md @@ -0,0 +1,113 @@ +# P59 多世界并存压测结论(Sim 多世界 × 天数的内存/GC 增长曲线) + +- 日期:2026-09-28 +- 代码 commit:`2c559a7`(本 slice 起点;压测期间未改 Kernel/Sim) +- 运行方式:独立 `systemd --user` transient unit `living-village-multiworld.service` + (wrapper `scripts/profile-multiworld.sh`,`StandardOutput=append:` + `exec 2>&1` 汇一个文件) +- 证据:本目录 `acceptance.log`(主运行 N=10)、`control-n1.log`(控制组 N=1)、`state.json` +- 只读结论,**未修改任何 Kernel/Sim 语义**;`--performance-baseline` ×3 digest 仍为 + `953775FAEB2F…`,Kernel 119/119。 + +## 口径 + +`dotnet LivingVillage.Headless.dll --profile-multiworld <worlds> <days> <npcs> <sampleEveryDays> <seedBase>` + +- N 个 `Sim.initialWorldN(seedBase+i, 30)` 世界**同时存活**; +- 按采样轮推进:每轮用 `Parallel.For` 让全部世界各前进 `sampleEveryDays` 天,随后采样一次; +- 采样字段:`worlds`(并存世界数)、`events`、`rumors`、`annals`、`memory_entries` + (全部世界的聚合)、`gc_live_mb`(`GC.GetTotalMemory(true)`)、`working_set_rss_mb` + (`Process.WorkingSet64`)、`elapsed_s`;每行 flush; +- ground truth 为 **gc_live 与 RSS**;条目计数只作解释用。 + +### 标定探针(先跑) + +`--profile-multiworld 1 2`(1 世界 × 2 天)= **34.2s**,与 `docs/overnight-status.md` +中 Hermes 单 worker 亲测 34.6s 一致。据此定 **N=10 × 20 天**(10 世界并行一轮约 100s)。 + +## 主运行 N=10 × 20 天(12 天口径内的采样,`acceptance.log`) + +| day | rumors | annals | memory_entries | gc_live_mb | rss_mb | elapsed_s | +|----:|------:|------:|--------------:|----------:|------:|---------:| +| 0 | 0 | 0 | 0 | 0.12 | 39.05 | 0.0 | +| 2 | 38536 | 1280 | 19200 | 7.28 | 74.61 | 101.0 | +| 4 | 76353 | 1280 | 19200 | 12.19 | 91.55 | 206.1 | +| 6 | 75488 | 1280 | 19200 | 12.09 | 102.58 | 309.6 | +| 8 | 75737 | 1280 | 19200 | 12.13 | 123.10 | 412.4 | +| 10 | 75667 | 1280 | 19200 | 12.11 | 106.71 | 514.9 | +| 12 | 75908 | 1280 | 19200 | 12.15 | 112.00 | 616.9 | +| 14 | 76017 | 1280 | 19200 | 12.16 | 110.58 | 719.2 | +| 16 | 75334 | 1280 | 19200 | 12.06 | 111.41 | 820.4 | +| 18 | 75907 | 1280 | 19200 | 12.15 | 105.52 | 922.5 | +| 20 | 76687 | 1280 | 19200 | 12.25 | 103.96 | 1024.3 | + +wall = **1024.4s(≈17.1 min,< 40 min 预算)**,exit=0。 + +## 控制组 N=1 × 20 天(`control-n1.log`) + +| day | rumors | annals | memory_entries | gc_live_mb | rss_mb | elapsed_s | +|----:|------:|------:|--------------:|----------:|------:|---------:| +| 2 | 3921 | 128 | 1920 | 0.85 | 57.19 | 34.0 | +| 4 | 7843 | 128 | 1920 | 1.37 | 61.71 | 68.1 | +| 20 | 7932 | 128 | 1920 | 1.39 | 67.94 | 339.0 | + +wall = 339.1s。世界 0 的 `world_digests=0:2364CB74F5FCA345…` 与 N=10 主运行的 +世界 0 **逐字节一致** → 并存 + 并行推进不改变单世界语义。 + +## 结论 + +### 1) 增长是否随世界数超线性?——否,近似线性(内存实为略次线性) + +以 N=1 与 N=10 同日(D=20)对比: + +| 量 | N=1 | N=10 | 倍率 | +|---|---:|---:|---:| +| rumors | 7932 | 76687 | 9.67× | +| annals | 128 | 1280 | 10.00× | +| memory_entries | 1920 | 19200 | 10.00× | +| gc_live_mb | 1.39 | 12.25 | 8.81×(每世界 1.23MB) | +| wall_s | 339 | 1024 | 3.02×(10× 工作量,16 核并行) | + +条目计数严格 ~N 倍;GC live set **略低于** N 倍(运行时/JIT/GC 堆共享的固定开销被摊薄)。 +**未观测到任何超线性增长**。 + +### 2) RSS 是否有界?——有界,无单调泄漏 + +- N=10:39MB → 峰值 123MB(day 8)→ 回落并稳定在 **104–112MB**,day 10 后不再上升。 +- N=1:39MB → day 4 后稳定在 **64–68MB**。 +- 时间维度上,10 世界每 2 天一轮墙钟恒为 **101–105s**(波动 <4%),说明 2–20 天窗口内 + 逐日成本**不随天数上升**——与早期 M3 phase2 在 100 天视野看到的逐日成本上升不同。 + +原因(读码确认,非猜测,未改码):三类历史集合均有界—— +`Sim.rumorRetentionDays = 3`(按天淘汰,`rumorCapacity = 16384`/世界)、 +`maxAnnalEntries = 128`、`Mind.Memory` 每 NPC 上限。因此 day ~4 后每条曲线进入稳态。 + +`events` 全程为 0:自主模拟不产生 `InteractionEvent`(仅玩家交互写入),符合预期,非泄漏信号。 + +### 3) 是否有建议的最小修复?——**不修** + +所有指标未越界:live set 有界、RSS 有界且回落、随 N 线性、无 NaN(判据门在 M3R 另测)。 +本 slice 只读,不改任何 Kernel/Sim。 + +观察项(留档,不在本 slice 处理): +- M3 phase2 在 100 天视野的逐日成本上升未在 20 天窗口复现;若需定位需另设 >20 天采样。 +- 10 世界并行时单世界墙钟约为串行的 3.0×(GC/CPU 争用),16 核下仍远低于预算。 + +## 复现 + +```bash +# 控制组(前台) +dotnet src/LivingVillage.Headless/bin/Release/net8.0/LivingVillage.Headless.dll \ + --profile-multiworld 1 20 30 2 42 + +# 主运行(独立 systemd --user transient unit,见 deploy/living-village-multiworld.service) +TS=$(date -u +%Y%m%dT%H%M%SZ) +DIR=$PWD/artifacts/perf-multiworld/$TS +systemd-run --user --unit=living-village-multiworld \ + --property=Type=simple --property=WorkingDirectory=$PWD \ + --setenv=PMW_REPO=$PWD --setenv=PMW_DIR=$DIR --setenv=PMW_COMMIT=2c559a7 \ + --setenv=PMW_WORLDS=10 --setenv=PMW_DAYS=20 --setenv=PMW_NPCS=30 \ + --setenv=PMW_SAMPLE_EVERY=2 --setenv=PMW_SEED_BASE=42 --setenv=PMW_LOG=$DIR/acceptance.log \ + --property=StandardOutput=append:$DIR/acceptance.log \ + --property=StandardError=append:$DIR/acceptance.log --property=Restart=no \ + /usr/bin/bash $PWD/scripts/profile-multiworld.sh +``` diff --git a/artifacts/perf-multiworld/20260928T195839Z/acceptance.log b/artifacts/perf-multiworld/20260928T195839Z/acceptance.log new file mode 100644 index 0000000..8f6ac82 --- /dev/null +++ b/artifacts/perf-multiworld/20260928T195839Z/acceptance.log @@ -0,0 +1,15 @@ +[pmw] start unit=living-village-multiworld.service pid=165537 commit=2c559a7 worlds=10 days=20 npcs=30 sample_every=2 seed_base=42 started_at=2026-09-28T19:58:39Z +multiworld_start worlds=10 days=20 npcs=30 sample_every_days=2 seed_base=42 ticks_per_day=5184000 +multiworld_sample worlds=10 day=0/20 events=0 rumors=0 annals=0 memory_entries=0 gc_live_mb=0.12 working_set_rss_mb=39.05 elapsed_s=0.0 +multiworld_sample worlds=10 day=2/20 events=0 rumors=38536 annals=1280 memory_entries=19200 gc_live_mb=7.28 working_set_rss_mb=74.61 elapsed_s=101.0 +multiworld_sample worlds=10 day=4/20 events=0 rumors=76353 annals=1280 memory_entries=19200 gc_live_mb=12.19 working_set_rss_mb=91.55 elapsed_s=206.1 +multiworld_sample worlds=10 day=6/20 events=0 rumors=75488 annals=1280 memory_entries=19200 gc_live_mb=12.09 working_set_rss_mb=102.58 elapsed_s=309.6 +multiworld_sample worlds=10 day=8/20 events=0 rumors=75737 annals=1280 memory_entries=19200 gc_live_mb=12.13 working_set_rss_mb=123.10 elapsed_s=412.4 +multiworld_sample worlds=10 day=10/20 events=0 rumors=75667 annals=1280 memory_entries=19200 gc_live_mb=12.11 working_set_rss_mb=106.71 elapsed_s=514.9 +multiworld_sample worlds=10 day=12/20 events=0 rumors=75908 annals=1280 memory_entries=19200 gc_live_mb=12.15 working_set_rss_mb=112.00 elapsed_s=616.9 +multiworld_sample worlds=10 day=14/20 events=0 rumors=76017 annals=1280 memory_entries=19200 gc_live_mb=12.16 working_set_rss_mb=110.58 elapsed_s=719.2 +multiworld_sample worlds=10 day=16/20 events=0 rumors=75334 annals=1280 memory_entries=19200 gc_live_mb=12.06 working_set_rss_mb=111.41 elapsed_s=820.4 +multiworld_sample worlds=10 day=18/20 events=0 rumors=75907 annals=1280 memory_entries=19200 gc_live_mb=12.15 working_set_rss_mb=105.52 elapsed_s=922.5 +multiworld_sample worlds=10 day=20/20 events=0 rumors=76687 annals=1280 memory_entries=19200 gc_live_mb=12.25 working_set_rss_mb=103.96 elapsed_s=1024.3 +multiworld_done worlds=10 day=20 elapsed_s=1024.4 world_digests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pmw] end pid=165537 exit=0 ended_at=2026-09-28T20:15:43Z diff --git a/artifacts/perf-multiworld/20260928T195839Z/control-n1.log b/artifacts/perf-multiworld/20260928T195839Z/control-n1.log new file mode 100644 index 0000000..88fe297 --- /dev/null +++ b/artifacts/perf-multiworld/20260928T195839Z/control-n1.log @@ -0,0 +1,13 @@ +multiworld_start worlds=1 days=20 npcs=30 sample_every_days=2 seed_base=42 ticks_per_day=5184000 +multiworld_sample worlds=1 day=0/20 events=0 rumors=0 annals=0 memory_entries=0 gc_live_mb=0.09 working_set_rss_mb=38.83 elapsed_s=0.0 +multiworld_sample worlds=1 day=2/20 events=0 rumors=3921 annals=128 memory_entries=1920 gc_live_mb=0.85 working_set_rss_mb=57.19 elapsed_s=34.0 +multiworld_sample worlds=1 day=4/20 events=0 rumors=7843 annals=128 memory_entries=1920 gc_live_mb=1.37 working_set_rss_mb=61.71 elapsed_s=68.1 +multiworld_sample worlds=1 day=6/20 events=0 rumors=7849 annals=128 memory_entries=1920 gc_live_mb=1.38 working_set_rss_mb=63.63 elapsed_s=101.9 +multiworld_sample worlds=1 day=8/20 events=0 rumors=7925 annals=128 memory_entries=1920 gc_live_mb=1.39 working_set_rss_mb=64.29 elapsed_s=135.8 +multiworld_sample worlds=1 day=10/20 events=0 rumors=7723 annals=128 memory_entries=1920 gc_live_mb=1.37 working_set_rss_mb=64.57 elapsed_s=169.6 +multiworld_sample worlds=1 day=12/20 events=0 rumors=7727 annals=128 memory_entries=1920 gc_live_mb=1.37 working_set_rss_mb=64.92 elapsed_s=203.6 +multiworld_sample worlds=1 day=14/20 events=0 rumors=8038 annals=128 memory_entries=1920 gc_live_mb=1.41 working_set_rss_mb=67.88 elapsed_s=237.2 +multiworld_sample worlds=1 day=16/20 events=0 rumors=7953 annals=128 memory_entries=1920 gc_live_mb=1.40 working_set_rss_mb=67.88 elapsed_s=271.2 +multiworld_sample worlds=1 day=18/20 events=0 rumors=7901 annals=128 memory_entries=1920 gc_live_mb=1.39 working_set_rss_mb=67.89 elapsed_s=305.2 +multiworld_sample worlds=1 day=20/20 events=0 rumors=7932 annals=128 memory_entries=1920 gc_live_mb=1.39 working_set_rss_mb=67.94 elapsed_s=339.0 +multiworld_done worlds=1 day=20 elapsed_s=339.1 world_digests=0:2364CB74F5FCA34503269F413F7184F455006C9975EB01B11219D25A3A33FDD2 diff --git a/artifacts/perf-multiworld/20260928T195839Z/state.json b/artifacts/perf-multiworld/20260928T195839Z/state.json new file mode 100644 index 0000000..3ead223 --- /dev/null +++ b/artifacts/perf-multiworld/20260928T195839Z/state.json @@ -0,0 +1,16 @@ +{ + "commit": "2c559a7", + "ended_at": "2026-09-28T20:15:43Z", + "exit": 0, + "log": "/home/somhairle/projects/living-village/artifacts/perf-multiworld/20260928T195839Z/acceptance.log", + "params": { + "days": 20, + "npcs": 30, + "sample_every_days": 2, + "seed_base": 42, + "worlds": 10 + }, + "pid": 165537, + "started_at": "2026-09-28T19:58:39Z", + "unit": "living-village-multiworld.service" +} diff --git a/deploy/living-village-multiworld.service b/deploy/living-village-multiworld.service new file mode 100644 index 0000000..a16544d --- /dev/null +++ b/deploy/living-village-multiworld.service @@ -0,0 +1,40 @@ +# P59 多世界并存压测 systemd --user 单元模板(N 世界 × D 天,逐采样行流式)。 +# +# 独立于 writer cgroup 运行(transient unit),输出 append 到 acceptance.log、exec 2>&1。 +# 安装示例: +# TS=$(date -u +%Y%m%dT%H%M%SZ) +# systemd-run --user --unit=living-village-multiworld \ +# --property=Type=simple \ +# --property=WorkingDirectory=@REPO@ \ +# --setenv=PMW_REPO=@REPO@ \ +# --setenv=PMW_DIR=@REPO@/artifacts/perf-multiworld/$TS \ +# --setenv=PMW_COMMIT=@COMMIT@ \ +# --setenv=PMW_WORLDS=10 --setenv=PMW_DAYS=20 --setenv=PMW_NPCS=30 \ +# --setenv=PMW_SAMPLE_EVERY=2 --setenv=PMW_SEED_BASE=42 \ +# --setenv=PMW_LOG=@REPO@/artifacts/perf-multiworld/$TS/acceptance.log \ +# --property=StandardOutput=append:@REPO@/artifacts/perf-multiworld/$TS/acceptance.log \ +# --property=StandardError=append:@REPO@/artifacts/perf-multiworld/$TS/acceptance.log \ +# --property=Restart=no \ +# /usr/bin/bash @REPO@/scripts/profile-multiworld.sh +[Unit] +Description=Living Village multi-world coexistence memory/GC profile +After=default.target + +[Service] +Type=simple +WorkingDirectory=@REPO@ +Environment=PMW_REPO=@REPO@ +Environment=PMW_DIR=@REPO@/artifacts/perf-multiworld/@TS@ +Environment=PMW_COMMIT=@COMMIT@ +Environment=PMW_WORLDS=10 +Environment=PMW_DAYS=20 +Environment=PMW_NPCS=30 +Environment=PMW_SAMPLE_EVERY=2 +Environment=PMW_SEED_BASE=42 +Environment=PMW_LOG=@REPO@/artifacts/perf-multiworld/@TS@/acceptance.log +ExecStart=/usr/bin/bash @REPO@/scripts/profile-multiworld.sh +StandardOutput=append:@REPO@/artifacts/perf-multiworld/@TS@/acceptance.log +StandardError=append:@REPO@/artifacts/perf-multiworld/@TS@/acceptance.log +Restart=no +TimeoutStartSec=infinity +TimeoutStopSec=30s diff --git a/scripts/profile-multiworld.sh b/scripts/profile-multiworld.sh new file mode 100755 index 0000000..3ab099d --- /dev/null +++ b/scripts/profile-multiworld.sh @@ -0,0 +1,80 @@ +#!/usr/bin/env bash +# P59 多世界并存后台压测 MasterProcess(systemd --user transient unit 的 ExecStart)。 +# +# 职责: +# * 在 <PMW_DIR> 落 state.json(pid/unit/log/commit/参数/起止/exit); +# * 以 2>&1 运行 Headless `--profile-multiworld N D npcs sampleEvery seedBase`, +# 逐采样行流式打印并 flush; +# * 退出时写回 ended_at/exit,并回传真实退出码。 +# +# 环境变量(由 unit/transient 属性注入): +# PMW_REPO / PMW_DIR / PMW_COMMIT / PMW_LOG +# PMW_WORLDS(默认 10)/ PMW_DAYS(默认 20)/ PMW_NPCS(默认 30) +# PMW_SAMPLE_EVERY(默认 2)/ PMW_SEED_BASE(默认 42) +set -u + +REPO="${PMW_REPO:?PMW_REPO required}" +DIR="${PMW_DIR:?PMW_DIR required}" +COMMIT="${PMW_COMMIT:?PMW_COMMIT required}" +WORLDS="${PMW_WORLDS:-10}" +DAYS="${PMW_DAYS:-20}" +NPCS="${PMW_NPCS:-30}" +SAMPLE_EVERY="${PMW_SAMPLE_EVERY:-2}" +SEED_BASE="${PMW_SEED_BASE:-42}" +LOG="${PMW_LOG:-$REPO/artifacts/perf-multiworld/latest/acceptance.log}" +UNIT="living-village-multiworld.service" +DLL="$REPO/src/LivingVillage.Headless/bin/Release/net8.0/LivingVillage.Headless.dll" + +mkdir -p "$DIR" +if [ ! -f "$DLL" ]; then + echo "[pmw] FATAL missing build: $DLL" >&2 + exit 90 +fi + +STARTED="$(date -u +%Y-%m-%dT%H:%M:%SZ)" +echo "$$" > "$DIR/pid" +echo "[pmw] start unit=$UNIT pid=$$ commit=$COMMIT worlds=$WORLDS days=$DAYS npcs=$NPCS sample_every=$SAMPLE_EVERY seed_base=$SEED_BASE started_at=$STARTED" + +python3 - "$DIR/state.json" "$UNIT" "$COMMIT" "$WORLDS" "$DAYS" "$NPCS" "$SAMPLE_EVERY" "$SEED_BASE" "$LOG" "$STARTED" "$$" <<'PY' +import json, sys +out, unit, commit, worlds, days, npcs, sample, seed, log, started, pid = sys.argv[1:12] +state = { + "pid": int(pid), + "unit": unit, + "log": log, + "commit": commit, + "params": { + "worlds": int(worlds), + "days": int(days), + "npcs": int(npcs), + "sample_every_days": int(sample), + "seed_base": int(seed), + }, + "started_at": started, + "ended_at": None, + "exit": None, +} +with open(out, "w", encoding="utf-8") as fh: + json.dump(state, fh, ensure_ascii=False, indent=2, sort_keys=True) + fh.write("\n") +PY + +( exec dotnet "$DLL" --profile-multiworld "$WORLDS" "$DAYS" "$NPCS" "$SAMPLE_EVERY" "$SEED_BASE" 2>&1 ) +CODE=$? + +ENDED="$(date -u +%Y-%m-%dT%H:%M:%SZ)" +echo "[pmw] end pid=$$ exit=$CODE ended_at=$ENDED" + +python3 - "$DIR/state.json" "$CODE" "$ENDED" <<'PY' +import json, sys +path, code, ended = sys.argv[1:4] +with open(path, encoding="utf-8") as fh: + state = json.load(fh) +state["ended_at"] = ended +state["exit"] = int(code) +with open(path, "w", encoding="utf-8") as fh: + json.dump(state, fh, ensure_ascii=False, indent=2, sort_keys=True) + fh.write("\n") +PY + +exit "$CODE" diff --git a/src/LivingVillage.Desktop.Tests/M3BatchOutputTests.fs b/src/LivingVillage.Desktop.Tests/M3BatchOutputTests.fs index b1822c7..963d571 100644 --- a/src/LivingVillage.Desktop.Tests/M3BatchOutputTests.fs +++ b/src/LivingVillage.Desktop.Tests/M3BatchOutputTests.fs @@ -32,6 +32,13 @@ type M3BatchOutputTests () = Assert.AreEqual<string>("[world 0] progress day=15/15 elapsed=260.0s", BatchOutput.progressLine 0 15L 15L 260.0) [<TestMethod>] + member _.MultiworldLineReportsCountsGcRssAndElapsed () = + let line = BatchOutput.multiworldLine 8 4L 20L 1600L 32000L 64L 5120L (1048576L * 96L) (1048576L * 240L) 31.5 + Assert.AreEqual<string>( + "multiworld_sample worlds=8 day=4/20 events=1600 rumors=32000 annals=64 memory_entries=5120 gc_live_mb=96.00 working_set_rss_mb=240.00 elapsed_s=31.5", + line) + + [<TestMethod>] member _.SummaryLineKeepsOldFieldsAndAppendsWallClock () = let line = BatchOutput.summaryLine 4 3 1 4 1.100 9.900 0.010 0.400 4 610.5 // 监督器正则只读前三个字段,前缀必须逐字保持。 diff --git a/src/LivingVillage.Headless/BatchOutput.fs b/src/LivingVillage.Headless/BatchOutput.fs index 971da97..aa79391 100644 --- a/src/LivingVillage.Headless/BatchOutput.fs +++ b/src/LivingVillage.Headless/BatchOutput.fs @@ -42,6 +42,23 @@ module BatchOutput = "batch_summary worlds=%d passed=%d failed=%d old_passed=%d/%d ratio_min=%.3f ratio_max=%.3f gini_min=%.3f gini_max=%.3f elapsed_s=%.1f workers=%d wall_s=%.1f" worlds passed failed oldPassed worlds ratioMin ratioMax giniMin giniMax elapsedSeconds workers elapsedSeconds + /// 多世界并存压测采样行:worlds/day 与各历史集合计数 + GC live set + OS RSS + 墙钟。 + let multiworldLine + (worlds: int) + (day: int64) + (totalDays: int64) + (events: int64) + (rumors: int64) + (annals: int64) + (memoryEntries: int64) + (gcLiveBytes: int64) + (rssBytes: int64) + (elapsedSeconds: float) : string = + let mb (bytes: int64) = float bytes / (1024.0 * 1024.0) + sprintf + "multiworld_sample worlds=%d day=%d/%d events=%d rumors=%d annals=%d memory_entries=%d gc_live_mb=%.2f working_set_rss_mb=%.2f elapsed_s=%.1f" + worlds day totalDays events rumors annals memoryEntries (mb gcLiveBytes) (mb rssBytes) elapsedSeconds + /// --cost-probe 单段输出:ticks/s、墙钟、分配与 GC 全部来自真实测量。 let costLine (days: int64) diff --git a/src/LivingVillage.Headless/Program.fs b/src/LivingVillage.Headless/Program.fs index 4bd6424..c9dd442 100644 --- a/src/LivingVillage.Headless/Program.fs +++ b/src/LivingVillage.Headless/Program.fs @@ -612,6 +612,66 @@ let runProfileLongRun (argv: string[]) : int = Console.Out.Flush() if world.Tick = days * ticksPerDay then 0 else 1 +/// `--profile-multiworld [worlds] [days] [npcs] [sampleEveryDays] [seedBase]`: +/// N 个 Sim 世界并存、按采样轮并行推进 D 天(只读 profile,不改 Kernel/Sim 语义)。 +/// 每 `sampleEveryDays` 天打印一行流式采样(worlds/count、events、rumors、annals、 +/// memory_entries、gc_live_mb、working_set_rss_mb、elapsed_s)并 flush;行的聚合口径为 +/// 全部并存世界之和,ground truth 是 GC live set 与 OS WorkingSet。 +let runProfileMultiworld (argv: string[]) : int = + let argInt64 index (dflt: int64) = + if argv.Length > index then + match Int64.TryParse argv.[index] with + | true, v when v > 0L -> v + | _ -> dflt + else dflt + let worlds = int (argInt64 1 10L) + let days = argInt64 2 20L + let npcs = int (argInt64 3 (int64 Sim.npcCount)) + let sampleEvery = argInt64 4 2L + let seedBase = if argv.Length > 5 then (match UInt64.TryParse argv.[5] with | true, v -> v | _ -> 42UL) else 42UL + let proc = System.Diagnostics.Process.GetCurrentProcess() + printfn "multiworld_start worlds=%d days=%d npcs=%d sample_every_days=%d seed_base=%d ticks_per_day=%d" + worlds days npcs sampleEvery seedBase ticksPerDay + Console.Out.Flush() + let ws = Array.init worlds (fun i -> Sim.initialWorldN (seedBase + uint64 i) npcs) + let drivers = Array.init worlds (fun i -> { Rng = Rng.ofSeed (seedBase + uint64 i); Until = segmentTicks; Input = zeroInput }) + let sw = System.Diagnostics.Stopwatch.StartNew() + let advanceOne (i: int) (targetDay: int64) = + let mutable world = ws.[i] + let mutable driver = drivers.[i] + let dayEnd = targetDay * ticksPerDay + while world.Tick < dayEnd do + let d, input = + if world.Tick < driver.Until then driver, driver.Input + else + let d = nextSegment driver.Rng driver.Until + d, d.Input + driver <- d + world <- Sim.step { Input = input } world + ws.[i] <- world + drivers.[i] <- driver + let sample (day: int64) = + let gcLive = System.GC.GetTotalMemory true + proc.Refresh() + let rss = proc.WorkingSet64 + let events = ws |> Array.sumBy (fun w -> int64 w.Events.Length) + let rumors = ws |> Array.sumBy (fun w -> int64 w.Rumors.Length) + let annals = ws |> Array.sumBy (fun w -> int64 w.Annals.Length) + let memEntries = ws |> Array.sumBy (fun w -> int64 (w.Npcs |> Array.sumBy (fun n -> n.Mind.Memory.Length))) + printfn "%s" (LivingVillage.Headless.BatchOutput.multiworldLine worlds day days events rumors annals memEntries gcLive rss sw.Elapsed.TotalSeconds) + Console.Out.Flush() + sample 0L + let mutable dayNo = 0L + while dayNo < days do + let target = min days (dayNo + sampleEvery) + System.Threading.Tasks.Parallel.For(0, worlds, fun i -> advanceOne i target) |> ignore + dayNo <- target + sample dayNo + let digests = ws |> Array.mapi (fun i w -> sprintf "%d:%s" i (LivingVillage.Headless.PerformanceProbe.worldDigest w)) |> String.concat "," + printfn "multiworld_done worlds=%d day=%d elapsed_s=%.1f world_digests=%s" worlds dayNo sw.Elapsed.TotalSeconds digests + Console.Out.Flush() + if ws |> Array.forall (fun w -> w.Tick = days * ticksPerDay) then 0 else 1 + let runExisting (argv: string[]) = let rec parse (i: int) @@ -675,7 +735,7 @@ let runExisting (argv: string[]) = if stats.NonFinite > 0L || stats.OutOfBounds > 0L || not replayOk then 1 else 0 | Error msg -> eprintfn $"headless: {msg}" - eprintfn "usage: dotnet run -c Release --project src/LivingVillage.Headless -- --days N --seed S [--npc N] [--dump-relations] [--dump-rumors] [--replay-rumors] | --m5-smoke | --m6a-smoke | --batch K D | --performance-baseline | --cost-probe [D1,D2,...] | --rumor-bench [days npcs chats_per_day]" + eprintfn "usage: dotnet run -c Release --project src/LivingVillage.Headless -- --days N --seed S [--npc N] [--dump-relations] [--dump-rumors] [--replay-rumors] | --m5-smoke | --m6a-smoke | --batch K D | --performance-baseline | --cost-probe [D1,D2,...] | --rumor-bench [days npcs chats_per_day] | --profile-long-run [days sampleEveryDays npcs seed] | --profile-multiworld [worlds days npcs sampleEveryDays seedBase]" 2 let private parseCostLadder (argv: string[]) : int64 list = @@ -706,5 +766,7 @@ let main argv = runRumorBench argv elif argv.Length >= 1 && argv.[0] = "--profile-long-run" then runProfileLongRun argv + elif argv.Length >= 1 && argv.[0] = "--profile-multiworld" then + runProfileMultiworld argv else runExisting argv |
