feat(spdiag): Add PerfPoint instrumentation and MC_LOG logging to the vLLM invocation path.（为 vLLM 调用路径添加 PerfPoint 打点与 MC_LOG 日志

[ZhiningPan]

英文翻译结果
Change Overview

Add SpDiag performance instrumentation and MC_LOG log output to the path where vLLM calls Mooncake Store, covering the three-layer call chain of the Python binding layer, RealClient core layer, and Client service layer.

Change Content

1. PerfKey Definition (mooncake-integration/store/mooncake_perf_points.def)

25 new PerfKeys are added to cover the vLLM invocation path across three layers:

Python Binding Layer‌ (store_py.cpp): 8 STORE_PY keys
RealClient Core Layer‌ (real_client.cpp): 9 RC_* keys
Client Service Layer‌ (client_service.cpp): 1 PUT_BATCH_FULL key
2. PerfPoint Instrumentation Implementation
File Modification
mooncake-integration/store/store_py.cpp Add PerfPoint (KEY_MODULE) to 8 Python binding methods
mooncake-store/src/real_client.cpp Add PerfPoint + MC_LOG summary + per-key logs to 5 core methods
mooncake-store/src/client_service.cpp Add PUT_BATCH_FULL instrumentation to the Client::BatchPut method
3. Design Decisions (Q1b / Q2A / Q3b)
Q1b‌: Only output 1 summary log + per-key logs at the underlying layer to avoid duplicate logs between the entry layer and the underlying layer, while ensuring field integrity
Q2A‌: Use the KEY_MODULE level for main entry methods and the MODULE level for underlying _internal methods, which is consistent with the existing code style
Q3b‌: Add instrumentation only in the L1117 method with the actual implementation for the BatchQuery overload, to avoid repeated statistics on the same invocation chain
4. Benchmark Tool

Add mooncake-store/benchmarks/store_connector_bench.cpp to specifically test the instrumentation of the vLLM invocation path:

Cover the three phases of WRITE / IS_EXIST / GET
Allocate memory using numa_alloc_local (supports RDMA registration)
Correctly determine the success conditions where IS_EXIST returns 1 and GET returns the number of bytes
5. MC_LOG Log Format Example
Summary line + per-key line (missing fields will not be output)

改动概述

为 vLLM 调用 Mooncake Store 的路径添加 SpDiag 性能打点和 MC_LOG 日志输出，覆盖 Python 绑定层、RealClient 核心层、Client 服务层三层调用链。

改动内容

1. PerfKey 定义（mooncake-integration/store/mooncake_perf_points.def）

新增 25 个 PerfKey，分三层覆盖 vLLM 调用路径：

• Python 绑定层（store_py.cpp）：8 个 *_STORE_PY_* key
• RealClient 核心层（real_client.cpp）：9 个 RC_* key
• Client 服务层（client_service.cpp）：1 个 PUT_BATCH_FULL key

2. PerfPoint 打点实现

文件	改动
mooncake-integration/store/store_py.cpp	8 个 Python 绑定方法添加 PerfPoint（KEY_MODULE）
mooncake-store/src/real_client.cpp	5 个核心方法添加 PerfPoint + MC_LOG 汇总+per-key 日志
mooncake-store/src/client_service.cpp	Client::BatchPut 方法添加 PUT_BATCH_FULL 打点

3. 设计决策（Q1b / Q2A / Q3b）

• Q1b：只在下沉层输出 1 条汇总+per-key 日志，避免入口层与下沉层日志重复，同时确保字段完整性
• Q2A：主入口方法使用 KEY_MODULE 级别，下沉 _internal 方法使用 MODULE 级别，与现有代码风格保持一致
• Q3b：BatchQuery 重载只在实际实现的 L1117 方法中添加打点，避免同一调用链重复统计

4. Benchmark 工具

新增 mooncake-store/benchmarks/store_connector_bench.cpp，专测 vLLM 调用路径打点：

• 覆盖 WRITE / IS_EXIST / GET 三阶段
• 使用 numa_alloc_local 分配内存（支持 RDMA 注册）
• 正确判断 IS_EXIST 返回 1 和 GET 返回字节数的成功条件

5. MC_LOG 日志格式示例

汇总行 + per-key 行（缺失字段不输出）：

[ZhiningPan]

[batch_put_from_multi_buffers] elapsed_us=1849 num_keys=8 total_bytes=8388608
success=1 key=layer.0.req_0 size=1048576
success=1 key=layer.1.req_0 size=1048576
...

[batch_is_exist] elapsed_us=484 num_keys=8
success=1 key=layer.0.req_0
success=1 key=layer.1.req_0
...

[batch_get_into_multi_buffers] elapsed_us=158237 num_keys=8
success=1 key=layer.0.req_0 size=1048576 replica=memory endpoint=141.61.84.149:14024
success=0 key=layer.2.req_0 size=1048576
...