phase0: align benchmark v2 workload protocol

plan_exec.md

@@ -0,0 +1,20 @@
+# kv_bench 执行记录（benchmark_refactor）
+
+## Phase 0（已完成）
+- 日期：2026-03-03
+- 范围：
+  - 重构 `src/main.rs` 与 `rocksdb/main.cpp`，完成 v2 方法学最小清单：
+    - workload preset：`W1..W6`
+    - mixed/read/scan 的 prefill + shared keyspace
+    - 时长模式：`--warmup-secs` / `--measure-secs`
+    - 显式 read path parity：`--read-path snapshot|rw_txn`
+    - 统一 schema 结果落盘（CSV）并自动附带机器/环境元数据
+  - 更新脚本：`scripts/mace.sh`、`scripts/rocksdb.sh`、`scripts/plot.py`、`scripts/init.sh`
+  - 默认数据目录切换为 `/nvme` 体系（脚本强制 db_root 在 `/nvme` 下）
+- 编译验证：
+  - `cargo check -q` 通过
+  - `cargo build --release -q` 通过
+  - `cmake --build --preset release -j` 通过
+- 运行烟测：
+  - `mace` 与 `rocksdb` 均可按新参数运行并写入统一 schema 结果文件
+- 提交：待本阶段 commit

scripts/init.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
 
 python3 -m venv .
-./bin/pip3 install pandas matplotlib adjustText
+./bin/pip3 install pandas matplotlib
 rm -f .gitignore

scripts/mace.sh

@@ -2,55 +2,61 @@
 
 set -euo pipefail
 
-if [ "$#" -ne 1 ]
+if [ "$#" -lt 1 ] || [ "$#" -gt 2 ]; then
-then
+    printf "Usage: %s <db_root_under_/nvme> [result_csv]\n" "$0"
-    printf "\033[m$0 path\033[0m\n"
     exit 1
 fi
 
 script_dir="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
 root_dir="$(cd -- "${script_dir}/.." && pwd)"
 
-cargo build --release --manifest-path "${root_dir}/Cargo.toml" 1>/dev/null 2>/dev/null
+# The runner creates per-case unique paths under this root; each path must not exist.
+db_root="$1"
+result_file="${2:-${script_dir}/benchmark_results.csv}"
 
-function samples() {
+warmup_secs="${WARMUP_SECS:-10}"
-        export RUST_BACKTRACE=full
+measure_secs="${MEASURE_SECS:-20}"
-        kv_sz=(16 16 100 1024 1024 1024 16 10240)
+prefill_keys="${PREFILL_KEYS:-200000}"
-        mode=(insert get mixed scan)
+read_path="${READ_PATH:-snapshot}"
-        # set -x
-        db_root="$1"
 
-        cnt=100000
+if [[ "${db_root}" != /nvme* ]]; then
-        for ((i = 1; i <= $(nproc); i *= 2))
+    printf "db_root must be under /nvme, got: %s\n" "${db_root}" >&2
-        do
+    exit 1
-                for ((j = 0; j < ${#kv_sz[@]}; j += 2))
-                do
-                      for ((k = 0; k < ${#mode[@]}; k += 1))
-                      do 
-                            if [ "${mode[k]}" == "insert" ]
-                            then
-                                  "${root_dir}/target/release/kv_bench" --path "${db_root}" --threads "${i}" --iterations "${cnt}" --mode "${mode[k]}" --key-size "${kv_sz[j]}" --value-size "${kv_sz[j+1]}" --random
-                                  if test $? -ne 0
-                                  then
-                                          echo "${mode[k]} threads $i ksz ${kv_sz[j]} vsz ${kv_sz[j+1]} random fail"
-                                          exit 1
-                                  fi
-                            fi
-                            "${root_dir}/target/release/kv_bench" --path "${db_root}" --threads "${i}" --iterations "${cnt}" --mode "${mode[k]}" --key-size "${kv_sz[j]}" --value-size "${kv_sz[j+1]}"
-                            if test $? -ne 0
-                            then
-                                    echo "${mode[k]} threads $i ksz ${kv_sz[j]} vsz ${kv_sz[j+1]} fail"
-                                    exit 1
-                            fi
-                      done
-                done
-        done
-}
-
-echo mode,threads,key_size,value_size,insert_ratio,ops,elapsed_us > "${script_dir}/mace.csv"
-samples "$1" 1>> "${script_dir}/mace.csv"
-if [ -x "${script_dir}/bin/python" ]; then
-    (cd "${script_dir}" && "${script_dir}/bin/python" plot.py mace.csv)
-else
-    (cd "${script_dir}" && python3 plot.py mace.csv)
 fi
+
+mkdir -p "${db_root}"
+mkdir -p "$(dirname -- "${result_file}")"
+
+cargo build --release --manifest-path "${root_dir}/Cargo.toml"
+
+workloads=(W1 W2 W3 W4 W5 W6)
+threads=(1 6 12)
+profiles=(
+  "32 1024"
+  "32 16384"
+)
+
+for workload in "${workloads[@]}"; do
+    for t in "${threads[@]}"; do
+        for kv in "${profiles[@]}"; do
+            read -r key_size value_size <<< "${kv}"
+            run_path="$(mktemp -u -p "${db_root}" "mace_${workload}_${t}_${key_size}_${value_size}_XXXXXX")"
+            printf "[mace] workload=%s threads=%s key=%s value=%s path=%s\n" \
+              "${workload}" "${t}" "${key_size}" "${value_size}" "${run_path}"
+            "${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}"
+        done
+    done
+done
+
+printf "Mace runs finished. Results appended to: %s\n" "${result_file}"

scripts/plot.py

@@ -1,65 +1,81 @@
-import pandas as pd
+#!/usr/bin/env python3
-import matplotlib.pyplot as plt
+
-from adjustText import adjust_text
 import sys
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import pandas as pd
 
 
-def real_mode(m):
+def main() -> int:
-    if m == "mixed":
+    if len(sys.argv) not in (2, 3):
-        return "Mixed (70% Get, 30% Insert)"
+        print(f"Usage: {sys.argv[0]} <result_csv> [output_dir]")
-    elif m == "get":
+        return 1
-        return "Random Get"
+
-    elif m == "scan":
+    result_csv = Path(sys.argv[1])
-        return "Sequential Scan"
+    output_dir = Path(sys.argv[2]) if len(sys.argv) == 3 else result_csv.parent
-    return m.capitalize()
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    df = pd.read_csv(result_csv)
+
+    required = {
+        "engine",
+        "workload_id",
+        "threads",
+        "key_size",
+        "value_size",
+        "ops_per_sec",
+        "p99_us",
+    }
+    missing = required - set(df.columns)
+    if missing:
+        raise ValueError(f"Missing required columns: {sorted(missing)}")
+
+    for engine in sorted(df["engine"].unique()):
+        engine_df = df[df["engine"] == engine]
+        profiles = (
+            engine_df[["key_size", "value_size"]]
+            .drop_duplicates()
+            .sort_values(["key_size", "value_size"])
+            .itertuples(index=False)
+        )
+
+        for key_size, value_size in profiles:
+            sub = engine_df[
+                (engine_df["key_size"] == key_size)
+                & (engine_df["value_size"] == value_size)
+            ]
+            if sub.empty:
+                continue
+
+            for metric, ylabel in (("ops_per_sec", "OPS/s"), ("p99_us", "P99 Latency (us)")):
+                plt.figure(figsize=(12, 7))
+                for workload in sorted(sub["workload_id"].unique()):
+                    wdf = sub[sub["workload_id"] == workload].sort_values("threads")
+                    plt.plot(
+                        wdf["threads"],
+                        wdf[metric],
+                        marker="o",
+                        linewidth=2,
+                        label=workload,
+                    )
+
+                plt.title(
+                    f"{engine.upper()} {metric} (key={key_size}, value={value_size})",
+                    fontsize=14,
+                )
+                plt.xlabel("Threads")
+                plt.ylabel(ylabel)
+                plt.grid(True, linestyle="--", alpha=0.5)
+                plt.legend()
+                plt.tight_layout()
+                out = output_dir / f"{engine}_{metric}_k{key_size}_v{value_size}.png"
+                plt.savefig(out)
+                plt.close()
+
+    print(f"Charts written to: {output_dir}")
+    return 0
 
 
-name = sys.argv[1]
+if __name__ == "__main__":
-prefix = name.split(".")[0]
+    raise SystemExit(main())
-
-# read benchmark data
-# keep compatibility with older csv files that used elapsed/elasped
-# and normalize to elapsed_us
-
-df = pd.read_csv(f"./{name}")
-if "elapsed_us" not in df.columns:
-    if "elapsed" in df.columns:
-        df = df.rename(columns={"elapsed": "elapsed_us"})
-    elif "elasped" in df.columns:
-        df = df.rename(columns={"elasped": "elapsed_us"})
-
-# group by mode
-modes = df["mode"].unique()
-
-for mode in modes:
-    plt.figure(figsize=(16, 9))
-    subset = df[df["mode"] == mode]
-
-    # group by key/value size
-    key_value_combinations = subset.groupby(["key_size", "value_size"])
-
-    texts = []
-    for (key_size, value_size), group in key_value_combinations:
-        label = f"key={key_size}B, val={value_size}B"
-        x = group["threads"]
-        y = group["ops"]
-
-        # draw line
-        line, = plt.plot(x, y, marker="o", label=label)
-
-        # add labels
-        for xi, yi, ops in zip(x, y, group["ops"]):
-            texts.append(
-                plt.text(xi, yi, f"{int(ops)}", color=line.get_color(), fontsize=12)
-            )
-
-    adjust_text(texts, arrowprops=dict(arrowstyle="->", color="gray"))
-
-    plt.title(f"{prefix.upper()}: {real_mode(mode)}", fontsize=16)
-    plt.xlabel("Threads", fontsize=14)
-    plt.ylabel("OPS", fontsize=14)
-    plt.grid(True, linestyle="--", alpha=0.6)
-    plt.legend()
-    plt.tight_layout()
-    plt.savefig(f"{prefix}_{mode}.png")
-    plt.close()

scripts/rocksdb.sh

@@ -2,55 +2,61 @@
 
 set -euo pipefail
 
-if [ "$#" -ne 1 ]
+if [ "$#" -lt 1 ] || [ "$#" -gt 2 ]; then
-then
+    printf "Usage: %s <db_root_under_/nvme> [result_csv]\n" "$0"
-  printf "\033[m$0 path\033[0m\n"
+    exit 1
-  exit 1
 fi
 
 script_dir="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
 root_dir="$(cd -- "${script_dir}/.." && pwd)"
 rocksdb_dir="${root_dir}/rocksdb"
 
-(cd "${rocksdb_dir}" && cmake --preset release 1>/dev/null 2>/dev/null)
+db_root="$1"
-(cd "${rocksdb_dir}" && cmake --build --preset release 1>/dev/null 2>/dev/null)
+result_file="${2:-${script_dir}/benchmark_results.csv}"
 
-function samples() {
+warmup_secs="${WARMUP_SECS:-10}"
-        kv_sz=(16 16 100 1024 1024 1024 16 10240)
+measure_secs="${MEASURE_SECS:-20}"
-        mode=(insert get mixed scan)
+prefill_keys="${PREFILL_KEYS:-200000}"
-        # set -x
+read_path="${READ_PATH:-snapshot}"
-        db_root="$1"
-        cnt=100000
-        for ((i = 1; i <= $(nproc); i *= 2))
-        do
-                for ((j = 0; j < ${#kv_sz[@]}; j += 2))
-                do
-                        for ((k = 0; k < ${#mode[@]}; k += 1))
-                        do
-                            if [ "${mode[k]}" == "insert" ]
-                            then
-                                "${rocksdb_dir}/build/release/rocksdb_bench" --path "${db_root}" --threads "${i}" --iterations "${cnt}" --mode "${mode[k]}" --key-size "${kv_sz[j]}" --value-size "${kv_sz[j+1]}" --random
-                                if test $? -ne 0
-                                then
-                                        echo "${mode[k]} threads $i ksz ${kv_sz[j]} vsz ${kv_sz[j+1]} random fail"
-                                        exit 1
-                                fi
-                            fi
-                            "${rocksdb_dir}/build/release/rocksdb_bench" --path "${db_root}" --threads "${i}" --iterations "${cnt}" --mode "${mode[k]}" --key-size "${kv_sz[j]}" --value-size "${kv_sz[j+1]}"
-                            if test $? -ne 0
-                            then
-                                    echo "${mode[k]} threads $i ksz ${kv_sz[j]} vsz ${kv_sz[j+1]} fail"
-                                    exit 1
-                            fi
-                      done
-                done
-        done
-}
 
-echo mode,threads,key_size,value_size,insert_ratio,ops,elapsed_us > "${script_dir}/rocksdb.csv"
+if [[ "${db_root}" != /nvme* ]]; then
-samples "$1" 1>> "${script_dir}/rocksdb.csv"
+    printf "db_root must be under /nvme, got: %s\n" "${db_root}" >&2
-if [ -x "${script_dir}/bin/python" ]; then
+    exit 1
-    (cd "${script_dir}" && "${script_dir}/bin/python" plot.py rocksdb.csv)
-else
-    (cd "${script_dir}" && python3 plot.py rocksdb.csv)
 fi
+
+mkdir -p "${db_root}"
+mkdir -p "$(dirname -- "${result_file}")"
+
+(cd "${rocksdb_dir}" && cmake --preset release)
+(cd "${rocksdb_dir}" && cmake --build --preset release)
+
+workloads=(W1 W2 W3 W4 W5 W6)
+threads=(1 6 12)
+profiles=(
+  "32 1024"
+  "32 16384"
+)
+
+for workload in "${workloads[@]}"; do
+    for t in "${threads[@]}"; do
+        for kv in "${profiles[@]}"; do
+            read -r key_size value_size <<< "${kv}"
+            run_path="$(mktemp -u -p "${db_root}" "rocksdb_${workload}_${t}_${key_size}_${value_size}_XXXXXX")"
+            printf "[rocksdb] workload=%s threads=%s key=%s value=%s path=%s\n" \
+              "${workload}" "${t}" "${key_size}" "${value_size}" "${run_path}"
+            "${rocksdb_dir}/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}"
+        done
+    done
+done
+
+printf "RocksDB runs finished. Results appended to: %s\n" "${result_file}"