phase2: add steady-state matrix runner and report

plan_exec.md

@@ -36,3 +36,15 @@
   - `bash -n scripts/phase1.sh` 通过
   - `python3 -m py_compile scripts/phase1_eval.py` 通过
 - 提交：待本阶段 commit
+
+## Phase 2（已完成）
+- 日期：2026-03-03
+- 范围：
+  - 新增 `scripts/phase2.sh`：稳态核心报告矩阵执行器
+    - `tier-m` 全量：`W1/W2/W3/W4/W6` × `P2/P3` × `threads(1/6/12)` × `repeats(默认5)`
+    - 可选 `tier-l` 代表集：`RUN_TIER_L_REPRESENTATIVE=1` 启用，默认 `TIER_L_REPEATS=1`
+  - 新增 `scripts/phase2_report.py`：输出按 case 的 `throughput/p95/p99 median`，并给出慢场景对比表
+- 验证：
+  - `bash -n scripts/phase2.sh` 通过
+  - `python3 -m py_compile scripts/phase2_report.py` 通过
+- 提交：待本阶段 commit

scripts/phase2.sh

@@ -0,0 +1,150 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+if [ "$#" -lt 1 ] || [ "$#" -gt 2 ]; then
+    printf "Usage: %s <db_root_under_/nvme> [result_csv]\n" "$0"
+    exit 1
+fi
+
+script_dir="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+root_dir="$(cd -- "${script_dir}/.." && pwd)"
+
+if [[ "$1" != /nvme* ]]; then
+    printf "db_root must be under /nvme, got: %s\n" "$1" >&2
+    exit 1
+fi
+
+db_root="$1"
+result_file="${2:-${script_dir}/phase2_results.csv}"
+
+warmup_secs="${WARMUP_SECS:-120}"
+measure_secs="${MEASURE_SECS:-300}"
+repeats="${REPEATS:-5}"
+read_path="${READ_PATH:-snapshot}"
+run_tier_l_rep="${RUN_TIER_L_REPRESENTATIVE:-0}"
+tier_l_repeats="${TIER_L_REPEATS:-1}"
+
+mkdir -p "${db_root}"
+mkdir -p "$(dirname -- "${result_file}")"
+
+cargo build --release --manifest-path "${root_dir}/Cargo.toml"
+(cd "${root_dir}/rocksdb" && cmake --preset release)
+(cd "${root_dir}/rocksdb" && cmake --build --preset release)
+
+workloads_tier_m=(W1 W2 W3 W4 W6)
+workloads_tier_l_rep=(W1 W3 W6)
+threads_tier_m=(1 6 12)
+threads_tier_l_rep=(1 12)
+profiles=(P2 P3)
+
+profile_key() {
+    case "$1" in
+        P2) echo 32 ;;
+        P3) echo 32 ;;
+        *) printf "unknown profile: %s\n" "$1" >&2; exit 1 ;;
+    esac
+}
+
+profile_val() {
+    case "$1" in
+        P2) echo 1024 ;;
+        P3) echo 16384 ;;
+        *) printf "unknown profile: %s\n" "$1" >&2; exit 1 ;;
+    esac
+}
+
+prefill_for() {
+    local tier="$1"
+    local profile="$2"
+    if [ "${tier}" = "tier-m" ]; then
+        case "${profile}" in
+            P2) echo 18302417 ;;
+            P3) echo 1177348 ;;
+            *) printf "unknown profile: %s\n" "${profile}" >&2; exit 1 ;;
+        esac
+    elif [ "${tier}" = "tier-l" ]; then
+        case "${profile}" in
+            P2) echo 28470427 ;;
+            P3) echo 1831430 ;;
+            *) printf "unknown profile: %s\n" "${profile}" >&2; exit 1 ;;
+        esac
+    else
+        printf "unknown tier: %s\n" "${tier}" >&2
+        exit 1
+    fi
+}
+
+run_case() {
+    local engine="$1"
+    local tier="$2"
+    local workload="$3"
+    local profile="$4"
+    local t="$5"
+    local repeat="$6"
+
+    local key_size value_size prefill_keys run_path
+    key_size="$(profile_key "${profile}")"
+    value_size="$(profile_val "${profile}")"
+    prefill_keys="$(prefill_for "${tier}" "${profile}")"
+    run_path="$(mktemp -u -p "${db_root}" "${engine}_phase2_${tier}_${workload}_${profile}_t${t}_r${repeat}_XXXXXX")"
+
+    printf "[phase2][%s] tier=%s repeat=%s workload=%s profile=%s threads=%s path=%s\n" \
+      "${engine}" "${tier}" "${repeat}" "${workload}" "${profile}" "${t}" "${run_path}"
+
+    if [ "${engine}" = "mace" ]; then
+        "${root_dir}/target/release/kv_bench" \
+          --path "${run_path}" \
+          --workload "${workload}" \
+          --threads "${t}" \
+          --key-size "${key_size}" \
+          --value-size "${value_size}" \
+          --prefill-keys "${prefill_keys}" \
+          --warmup-secs "${warmup_secs}" \
+          --measure-secs "${measure_secs}" \
+          --shared-keyspace true \
+          --read-path "${read_path}" \
+          --result-file "${result_file}"
+    else
+        "${root_dir}/rocksdb/build/release/rocksdb_bench" \
+          --path "${run_path}" \
+          --workload "${workload}" \
+          --threads "${t}" \
+          --key-size "${key_size}" \
+          --value-size "${value_size}" \
+          --prefill-keys "${prefill_keys}" \
+          --warmup-secs "${warmup_secs}" \
+          --measure-secs "${measure_secs}" \
+          --read-path "${read_path}" \
+          --result-file "${result_file}"
+    fi
+}
+
+# tier-m full matrix
+for repeat in $(seq 1 "${repeats}"); do
+    for workload in "${workloads_tier_m[@]}"; do
+        for profile in "${profiles[@]}"; do
+            for t in "${threads_tier_m[@]}"; do
+                run_case mace tier-m "${workload}" "${profile}" "${t}" "${repeat}"
+                run_case rocksdb tier-m "${workload}" "${profile}" "${t}" "${repeat}"
+            done
+        done
+    done
+done
+
+# tier-l representative subset (optional)
+if [ "${run_tier_l_rep}" = "1" ]; then
+    for repeat in $(seq 1 "${tier_l_repeats}"); do
+        for workload in "${workloads_tier_l_rep[@]}"; do
+            for profile in "${profiles[@]}"; do
+                for t in "${threads_tier_l_rep[@]}"; do
+                    run_case mace tier-l "${workload}" "${profile}" "${t}" "${repeat}"
+                    run_case rocksdb tier-l "${workload}" "${profile}" "${t}" "${repeat}"
+                done
+            done
+        done
+    done
+fi
+
+python3 "${script_dir}/phase2_report.py" "${result_file}"
+printf "Phase 2 finished. Results: %s\n" "${result_file}"

scripts/phase2_report.py

@@ -0,0 +1,105 @@
+#!/usr/bin/env python3
+
+import sys
+import pandas as pd
+
+
+def infer_tier(row: pd.Series) -> str:
+    key = int(row["key_size"])
+    val = int(row["value_size"])
+    prefill = int(row["prefill_keys"])
+
+    table = {
+        (32, 1024, 18302417): "tier-m",
+        (32, 16384, 1177348): "tier-m",
+        (32, 1024, 28470427): "tier-l",
+        (32, 16384, 1831430): "tier-l",
+    }
+    return table.get((key, val, prefill), "unknown")
+
+
+def main() -> int:
+    if len(sys.argv) != 2:
+        print(f"Usage: {sys.argv[0]} <result_csv>")
+        return 1
+
+    df = pd.read_csv(sys.argv[1])
+
+    needed = {
+        "engine",
+        "workload_id",
+        "key_size",
+        "value_size",
+        "prefill_keys",
+        "threads",
+        "ops_per_sec",
+        "p95_us",
+        "p99_us",
+    }
+    missing = needed - set(df.columns)
+    if missing:
+        raise ValueError(f"Missing columns: {sorted(missing)}")
+
+    sub = df[df["workload_id"].isin(["W1", "W2", "W3", "W4", "W6"])].copy()
+    if sub.empty:
+        print("No phase2 rows found in csv")
+        return 0
+
+    sub["tier"] = sub.apply(infer_tier, axis=1)
+
+    grp_cols = ["tier", "engine", "workload_id", "key_size", "value_size", "threads"]
+    summary = (
+        sub.groupby(grp_cols)
+        .agg(
+            repeats=("ops_per_sec", "count"),
+            throughput_median=("ops_per_sec", "median"),
+            p95_median=("p95_us", "median"),
+            p99_median=("p99_us", "median"),
+        )
+        .reset_index()
+    )
+
+    with pd.option_context("display.max_rows", None, "display.max_columns", None):
+        print(summary.to_string(index=False))
+
+    # Slow-scenario extraction by throughput median.
+    piv = summary.pivot_table(
+        index=["tier", "workload_id", "key_size", "value_size", "threads"],
+        columns="engine",
+        values="throughput_median",
+        aggfunc="first",
+    ).reset_index()
+
+    if {"mace", "rocksdb"}.issubset(set(piv.columns)):
+        piv["slower_engine"] = piv.apply(
+            lambda r: "mace" if r["mace"] < r["rocksdb"] else "rocksdb",
+            axis=1,
+        )
+        piv["slower_ratio"] = piv.apply(
+            lambda r: min(r["mace"], r["rocksdb"]) / max(r["mace"], r["rocksdb"])
+            if max(r["mace"], r["rocksdb"]) > 0
+            else 0.0,
+            axis=1,
+        )
+        print("\nSlow scenarios (by throughput median):")
+        print(
+            piv[
+                [
+                    "tier",
+                    "workload_id",
+                    "key_size",
+                    "value_size",
+                    "threads",
+                    "mace",
+                    "rocksdb",
+                    "slower_engine",
+                    "slower_ratio",
+                ]
+            ].to_string(index=False)
+        )
+
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())