[optimizer] Add template prefix tracker and enhance offline analysis capabilities

kv_cache_manager/optimizer/analysis/script/README.md

@@ -35,9 +35,15 @@ bazel run //kv_cache_manager/optimizer/analysis/script:optimizer_run -- -c confi
 
 ### 输出
 
-- `<output_result_path>/*_hit_rates.csv` — 每个 instance 的命中率时序数据
-- `<output_result_path>/multi_instance_cache_analysis.png` — 命中率时序图（需 `--draw-chart`）
-- `<output_result_path>/*_lifecycle.csv` — block 生命周期数据（需 `--export-lifecycle`）
+```
+<output_result_path>/
+├── *_hit_rates.csv                       # 每个 instance 的命中率时序数据
+├── *_template_prefix_traces.csv          # per-trace 模板归属明细
+├── *_template_prefix_summary.csv         # 模板级汇总
+├── *_lifecycle.csv                       # block 生命周期数据（需 --export-lifecycle）
+└── timeseries/
+    └── multi_instance_cache_analysis.png # 命中率时序图（需 --draw-chart）
+```
 
 ---
 
@@ -106,9 +112,19 @@ bazel run //kv_cache_manager/optimizer/analysis/script:tradeoff -- \
 
 ### 输出
 
-- `pareto_curve_<type>.png` — 单策略 Pareto 散点图
-- `multi_policy_<type>.png` — 多策略对比子图
-- `csv_results/cap_<capacity>_<policy>/` — 每次运行的 CSV（需 `--save-csv`）
+```
+<output_result_path>/
+├── pareto/
+│   ├── pareto_curve_<type>.png           # 单策略 Pareto 散点图
+│   └── multi_policy_<type>.png           # 多策略对比子图
+├── timeseries/
+│   └── multi_instance_cache_analysis.png # 时序图（需 --plot-timeseries）
+└── csv_results/                          # 需 --save-csv
+    └── cap_<capacity>_<policy>/
+        ├── *_hit_rates.csv
+        ├── *_template_prefix_traces.csv
+        └── *_template_prefix_summary.csv
+```
 
 ---
 
@@ -180,9 +196,13 @@ python kv_cache_manager/optimizer/analysis/script/plot/radix_tree_plot.py \
 
 ### 输出
 
-- `<instance>_radix_tree.json` — 前缀树结构数据
-- `<instance>_radix_tree.png` — 完整树可视化
-- `<instance>_hot_paths.png` — 热点路径可视化
+```
+<output_result_path>/
+└── radix_tree/
+    ├── <instance>_radix_tree.json        # 前缀树结构数据
+    ├── <instance>_radix_tree.png         # 完整树可视化
+    └── <instance>_hot_paths.png          # 热点路径可视化
+```
 
 ---
 
@@ -227,10 +247,15 @@ bazel run //kv_cache_manager/optimizer/analysis/script:analyze_lifecycle -- \
 
 ### 输出
 
-- 控制台统计报告
-- `<instance>_physical_lifespan_cdf.png` — Physical Lifespan CDF（全量 + Evicted）
-- `<instance>_active_lifespan_cdf.png` — Active Lifespan CDF
-- `<instance>_access_count.png` — Access Count 直方图（全量 + 去零两张子图）
+```
+<output_result_path>/
+└── lifecycle/
+    ├── <instance>_physical_lifespan_cdf.png  # Physical Lifespan CDF（全量 + Evicted）
+    ├── <instance>_active_lifespan_cdf.png    # Active Lifespan CDF
+    └── <instance>_access_count.png           # Access Count 直方图（全量 + 去零两张子图）
+```
+
+控制台同步输出统计报告。
 
 ---
 
@@ -285,4 +310,45 @@ script/
     ├── optimizer_runner.py       # optimizer 运行封装
     ├── csv_loader.py             # CSV 加载 + 容量列表
     └── plot_utils.py             # 绘图风格 + Pareto 绘图
+```
+
+---
+
+## 输出目录总览
+
+所有脚本共享同一个根目录 `<output_result_path>`（来自 config.json `output_result_path` 字段）。
+
+```
+<output_result_path>/
+│
+│  # ── C++ optimizer 原始数据输出 ──────────────────────────────
+├── *_hit_rates.csv                           # 命中率时序（每条 trace 上报）
+├── *_template_prefix_traces.csv              # per-trace 模板归属明细
+├── *_template_prefix_summary.csv             # 模板级汇总
+├── *_lifecycle.csv                           # block 生命周期（需 --export-lifecycle）
+│
+│  # ── Python 图表输出 ────────────────────────────────────────
+├── pareto/                                   # tradeoff
+│   ├── pareto_curve_<type>.png
+│   └── multi_policy_<type>.png
+│
+├── timeseries/                               # optimizer_run --draw-chart
+│   └── multi_instance_cache_analysis.png     # tradeoff --plot-timeseries
+│
+├── lifecycle/                                # analyze_lifecycle
+│   ├── *_physical_lifespan_cdf.png
+│   ├── *_active_lifespan_cdf.png
+│   └── *_access_count.png
+│
+├── radix_tree/                               # export_tree
+│   ├── *_radix_tree.json
+│   ├── *_radix_tree.png
+│   └── *_hot_paths.png
+│
+│  # ── tradeoff --save-csv 实验中间数据 ─────────────────────────
+└── csv_results/
+    └── cap_<N>_<policy>/
+        ├── *_hit_rates.csv
+        ├── *_template_prefix_traces.csv
+        └── *_template_prefix_summary.csv
 ```

kv_cache_manager/optimizer/analysis/script/plot/hit_rate_plot.py

@@ -32,7 +32,49 @@ def read_csv_file(csv_file_path):
         print(f"Error reading {csv_file_path}: {str(e)}")
         return None
 
-def plot_multi_instance_analysis(csv_dir):
+def _load_sp_cumulative(csv_dir, instance_name):
+    """
+    从 template_prefix_traces.csv 计算 system prompt 累积命中率时序。
+
+    返回 DataFrame: [TimestampUs, AccSpHitRate]
+    AccSpHitRate = cumsum(min(hit, template_depth)) / cumsum(total_blocks)
+    """
+    basename = instance_name.replace("_hit_rates", "")
+    sp_path = os.path.join(csv_dir, f"{basename}_template_prefix_traces.csv")
+    if not os.path.exists(sp_path):
+        return None
+
+    df = pd.read_csv(sp_path)
+    # trace_id format: trace_<instance>_<timestamp_us>
+    df['TimestampUs'] = df['TraceId'].str.rsplit('_', n=1).str[-1].astype(np.int64)
+    df = df.sort_values('TimestampUs')
+
+    sp_hits = np.where(
+        (df['TemplateId'] != 'NONE') & (df['TemplateDepth'] > 0),
+        np.minimum(df['HitBlocks'].values, df['TemplateDepth'].values),
+        0
+    )
+
+    cum_sp_hits = np.cumsum(sp_hits)
+    cum_total = np.cumsum(df['TotalBlocks'].values)
+
+    acc_sp_rate = np.where(cum_total > 0, cum_sp_hits / cum_total, 0.0)
+
+    return pd.DataFrame({
+        'TimestampUs': df['TimestampUs'].values,
+        'AccSpHitRate': acc_sp_rate,
+    })
+
+
+def plot_multi_instance_analysis(csv_dir, output_dir: str = None):
+    """
+    读取 csv_dir 下的命中率 CSV，生成时序分析图。
+
+    Args:
+        csv_dir:    CSV 数据目录
+        output_dir: 图表根输出目录，图表保存至 output_dir/timeseries/
+                    默认为 csv_dir（向后兼容）
+    """
     csv_files = sorted(glob.glob(os.path.join(csv_dir, "*_hit_rates.csv")))
     if not csv_files:
         print(f"Error: No CSV files found in directory: {csv_dir}")
@@ -77,8 +119,8 @@ def plot_multi_instance_analysis(csv_dir):
     base = pd.DataFrame({'t': base_timestamps})  # 用于merge_asof
 
     all_acc_hit, all_acc_external_hit, all_time_ranges = [], [], []
-    # 用于瞬时命中率计算：累积读块数 / 累积命中块数（反推）
     all_acc_read_blocks, all_acc_hit_blocks, all_acc_ext_hit_blocks = [], [], []
+    all_acc_sp_hit = []
     global_updates_list = []
     for df in dataframes:
         d = df.copy()
@@ -114,6 +156,22 @@ def plot_multi_instance_analysis(csv_dir):
         all_acc_hit_blocks.append(aligned['AccHitBlocks'].to_numpy(float))
         all_acc_ext_hit_blocks.append(aligned['AccExtHitBlocks'].to_numpy(float))
 
+    # ---- SP 累积命中率对齐 ----
+    for idx, name in enumerate(instance_names):
+        sp_df = _load_sp_cumulative(csv_dir, name)
+        if sp_df is None:
+            all_acc_sp_hit.append(None)
+            continue
+        sp_df['t'] = (sp_df['TimestampUs'] - min_timestamp) / 1e6
+        sp_df = sp_df.sort_values('t')
+        sp_aligned = pd.merge_asof(
+            base, sp_df[['t', 'AccSpHitRate']], on='t',
+            direction='backward', allow_exact_matches=True
+        )
+        t0, _ = all_time_ranges[idx]
+        sp_aligned.loc[sp_aligned['t'] < t0, 'AccSpHitRate'] = np.nan
+        all_acc_sp_hit.append(sp_aligned['AccSpHitRate'].to_numpy(float))
+
     global_updates = pd.concat(global_updates_list, ignore_index=True)
     global_updates = global_updates.dropna(subset=['t', 'CachedBlocksAllInstance']).sort_values('t')
 
@@ -233,7 +291,7 @@ def window_hit_rate(timestamps, acc_hit_blocks, acc_read_blocks, window_seconds=
     top_lines = [ax_top.lines[0]]
     bot_lines = [ax_bot.lines[0]]
 
-    # 上图：累计命中率
+    # 上图：累计命中率 + system prompt 累积命中率
     for i, name in enumerate(instance_names):
         t0, t1 = all_time_ranges[i]
         valid = (base_timestamps >= t0) & (base_timestamps <= t1)
@@ -246,6 +304,14 @@ def window_hit_rate(timestamps, acc_hit_blocks, acc_read_blocks, window_seconds=
                     linewidth=1.5, drawstyle='steps-post')
         top_lines += l1
 
+        if all_acc_sp_hit[i] is not None:
+            sp_line = ax_top_r.plot(
+                base_timestamps[valid], np.array(all_acc_sp_hit[i])[valid],
+                color=colors[i], linestyle=':', linewidth=2.5, alpha=0.9,
+                label=f'{name} - SP AccHitRate',
+                drawstyle='steps-post')
+            top_lines += sp_line
+
     # 下图：时间窗口内真实命中率（累积量差值）+ 按时间降采样
     downsample_interval_s = 10   # 每隔 10 秒取一个代表点
     window_seconds         = 10  # 窗口内累积命中率的统计时间跨度
@@ -300,7 +366,9 @@ def window_hit_rate(timestamps, acc_hit_blocks, acc_read_blocks, window_seconds=
     ax_top.set_title(f'Cache Analysis - {len(instance_names)} Instances', fontsize=15, fontweight='bold', pad=12)
 
     fig.tight_layout()
-    output_file = os.path.join(csv_dir, "multi_instance_cache_analysis.png")
+    timeseries_dir = os.path.join(output_dir or csv_dir, "timeseries")
+    os.makedirs(timeseries_dir, exist_ok=True)
+    output_file = os.path.join(timeseries_dir, "multi_instance_cache_analysis.png")
     plt.savefig(output_file, dpi=300, bbox_inches='tight', facecolor='white')
     print(f"Chart saved to: {output_file}")
     plt.close()

kv_cache_manager/optimizer/analysis/script/plot/lifecycle_plot.py

@@ -33,15 +33,18 @@ def _annotate_percentiles(ax, sorted_data, percentiles, x_max):
 
 
 def plot_physical_lifespan_cdf(all_sorted, evicted_sorted, instance_name, output_path):
-    """Physical Lifespan CDF: 全量 + Evicted"""
+    """Physical Lifespan CDF: 全量 + Evicted，统计标注基于 Evicted Only"""
     n_all, n_ev = len(all_sorted), len(evicted_sorted)
     if n_all == 0:
         print("  跳过 Physical Lifespan CDF: 无数据")
         return
 
+    # 统计标注优先用 evicted_sorted（真实驱逐分布），无驱逐数据时 fallback 到 all
+    stat_data = evicted_sorted if n_ev > 0 else all_sorted
+
     cdf_all = np.arange(1, n_all + 1) / n_all * 100
-    p99 = float(np.percentile(all_sorted, 99))
-    x_max = min(p99 * 1.5, float(all_sorted[-1])) if all_sorted[-1] > 0 else 1.0
+    p99 = float(np.percentile(stat_data, 99))
+    x_max = min(p99 * 1.5, float(stat_data[-1])) if stat_data[-1] > 0 else 1.0
 
     fig, ax = plt.subplots(figsize=(14, 8))
 
@@ -55,14 +58,15 @@ def plot_physical_lifespan_cdf(all_sorted, evicted_sorted, instance_name, output
                 label=f"Evicted Only (n={n_ev:,})", alpha=0.8)
         ax.fill_between(evicted_sorted, cdf_ev, alpha=0.1, color="red")
 
-    _annotate_percentiles(ax, all_sorted, [50, 75, 90, 95, 99], x_max)
+    _annotate_percentiles(ax, stat_data, [50, 75, 90, 95, 99], x_max)
 
-    mean_val = float(np.mean(all_sorted))
-    median_val = float(np.median(all_sorted))
+    mean_val = float(np.mean(stat_data))
+    median_val = float(np.median(stat_data))
+    stat_label = "Evicted" if n_ev > 0 else "All"
     ax.axvline(mean_val, color="blue", linestyle="--", linewidth=2,
-               label=f"Mean: {mean_val:.1f}s", alpha=0.7)
+               label=f"Mean ({stat_label}): {mean_val:.1f}s", alpha=0.7)
     ax.axvline(median_val, color="orange", linestyle="--", linewidth=2,
-               label=f"Median: {median_val:.1f}s", alpha=0.7)
+               label=f"Median ({stat_label}): {median_val:.1f}s", alpha=0.7)
 
     ax.set_xlim([0, x_max])
     ax.set_ylim([0, 105])

kv_cache_manager/optimizer/analysis/script/run/analyze_lifecycle.py

@@ -31,7 +31,7 @@
 
 
 def analyze_single(csv_path: str, output_dir: str, stats_only: bool = False):
-    """分析单个 lifecycle CSV"""
+    """分析单个 lifecycle CSV，图表保存至 output_dir/lifecycle/"""
     name = Path(csv_path).stem.replace("_lifecycle", "")
     print(f"\n{'='*60}")
     print(f"分析: {name}")
@@ -48,21 +48,22 @@ def analyze_single(csv_path: str, output_dir: str, stats_only: bool = False):
     if stats_only:
         return
 
-    os.makedirs(output_dir, exist_ok=True)
+    lifecycle_dir = os.path.join(output_dir, "lifecycle")
+    os.makedirs(lifecycle_dir, exist_ok=True)
     plot_data = extract_plot_data(df)
 
     print(f"\n生成图表:")
     plot_physical_lifespan_cdf(
         plot_data["physical_all"], plot_data["physical_evicted"],
-        name, os.path.join(output_dir, f"{name}_physical_lifespan_cdf.png"))
+        name, os.path.join(lifecycle_dir, f"{name}_physical_lifespan_cdf.png"))
 
     plot_active_lifespan_cdf(
         plot_data["active_all"],
-        name, os.path.join(output_dir, f"{name}_active_lifespan_cdf.png"))
+        name, os.path.join(lifecycle_dir, f"{name}_active_lifespan_cdf.png"))
 
     plot_access_count_histogram(
         plot_data["access_counts"],
-        name, os.path.join(output_dir, f"{name}_access_count.png"))
+        name, os.path.join(lifecycle_dir, f"{name}_access_count.png"))
 
 
 def main():

kv_cache_manager/optimizer/analysis/script/run/export_tree.py

@@ -84,14 +84,15 @@ def main():
         sys.exit(1)
     config = config_loader.config()
 
-    output_dir = Path(args.output_dir) if args.output_dir else Path(config.output_result_path())
+    root_dir = Path(args.output_dir) if args.output_dir else Path(config.output_result_path())
+    output_dir = root_dir / "radix_tree"
     output_dir.mkdir(parents=True, exist_ok=True)
 
     print("=" * 80)
     print("Radix Tree Export and Visualization")
     print("=" * 80)
     print("Config: {}".format(args.config))
-    print("Output: {}".format(output_dir))
+    print("Output: {}".format(root_dir))
     print()
 
     optimizer = kvcm_py_optimizer.OptimizerManager(config)
@@ -160,7 +161,7 @@ def main():
             )
 
     print("\n" + "=" * 80)
-    print("Done! Output: {}".format(output_dir))
+    print("Done! Output: {}".format(root_dir))
     print("=" * 80)
 
     kvcm_py_optimizer.LoggerBroker.DestroyLogger()

kv_cache_manager/optimizer/analysis/script/run/optimizer_run.py

@@ -80,7 +80,7 @@ def main():
     if args.draw_chart:
         t5 = time.time()
         print("\n[5/5] Generating charts...")
-        plot_multi_instance_analysis(output_path)
+        plot_multi_instance_analysis(output_path, output_path)
         print("      Charts done: {:.2f}s".format(time.time() - t5))
     else:
         print("\n[5/5] Skipping chart generation.")

kv_cache_manager/optimizer/analysis/script/run/tradeoff.py

@@ -105,8 +105,8 @@ def _print_multi_policy_table(results_by_policy, policies):
 # 时序图
 # ============================================================================
 
-def _plot_timeseries(csv_save_dir, results_by_policy, target_caps=None):
-    """为指定容量点生成命中率时序图"""
+def _plot_timeseries(csv_save_dir, results_by_policy, output_dir, target_caps=None):
+    """为指定容量点生成命中率时序图，图表保存至 output_dir/timeseries/"""
     print("\n" + "=" * 60)
     print("Generating Timeseries Plots")
     print("=" * 60)
@@ -118,7 +118,7 @@ def _plot_timeseries(csv_save_dir, results_by_policy, target_caps=None):
             if os.path.exists(cap_dir):
                 print("Plotting {} capacity={}...".format(pol, cap))
                 try:
-                    plot_multi_instance_analysis(cap_dir)
+                    plot_multi_instance_analysis(cap_dir, output_dir)
                     count += 1
                 except Exception as e:
                     print("  Failed: {}".format(e))
@@ -264,7 +264,7 @@ def main():
     # ----------------------------------------------------------------
     has_csv = args.save_csv or args.skip_run
     if args.plot_timeseries and has_csv:
-        _plot_timeseries(csv_save_dir, results_by_policy, args.plot_capacity)
+        _plot_timeseries(csv_save_dir, results_by_policy, output_dir, args.plot_capacity)
     elif args.plot_timeseries and not has_csv:
         print("\nWarning: --plot-timeseries requires --save-csv or --skip-run")