add the support for ub transport

docs/source/zh_archive/kunpeng_ub_transport.md

@@ -0,0 +1,90 @@
+# Kunpeng UB Transport 
+Kunpeng UbTransport源代码路径为Mooncake/mooncake-transfer-engine/src/transport/kunpneg_transport，该路径下有UB协议的Transport对接代码和实现逻辑。
+
+## 概述
+UB（Unified Bus，统一总线） 是与RDMA、CXL、NVLink 和TCP处于同一抽象层的传输协议，属于可在应用层灵活选择的传输方案。目前 UB 协议有两个开源实现：URMA（远程内存访问语义）和 OBMM（Load/Store 语义）。
+
+URMA（Unified Remote Memory Access，统一远程内存访问）是UB协议为上层应用提供的统一编程抽象与核心语义层。它基于 UB 协议低延迟、高带宽的底层特性，为远程共享内存的访问与操作提供统一的 API 和语义接口。
+
+URMA 开源代码仓库：https://atomgit.com/openeuler/umdk
+
+OBMM (Ownership Based Memory Management) 是面向超节点环境的内核内存管理系统，支持跨节点的物理内存共享。该系统通过内核模块 (obmm.ko) 和用户态库 (libobmm.so) 提供高效的远程内存访问能力。
+
+OBMM 开源代码仓库：https://atomgit.com/openeuler/obmm
+
+## 新增依赖
+Kunpeng UbTransport在Mooncake本身依赖的基础上，新增了一部分URMA和OBMM的依赖：
+
+- **硬件平台**: 支持原生UB互联架构的鲲鹏950 CPU
+- **OS版本**: openEuler 24.03 (LTS-SP3) [下载链接](https://www.openeuler.openatom.cn/zh/download/#openEuler%2024.03%20LTS%20SP3)
+- **URMA依赖**: UMDK: `yum install umdk-urma-devel` 或从[源码](https://atomgit.com/openeuler/umdk)构建。
+- **协议优势**: URMA 提供类似 RDMA 的内存语义，针对鲲鹏芯片片上互联进行了优化
+
+---
+
+## 构建与编译
+
+**前置条件**
+
+- openEuler 24.03 (LTS-SP3) [下载链接](https://www.openeuler.openatom.cn/zh/download/#openEuler%2024.03%20LTS%20SP3)
+- 已安装 UMDK: `yum install umdk-urma-devel` 或从[源码](https://atomgit.com/openeuler/umdk)构建
+
+**CMake 配置**
+
+```bash
+# 克隆 Mooncake 仓库
+git clone https://github.com/kvcache-ai/Mooncake.git 
+cd Mooncake
+
+# 启用 UB 传输层进行配置
+mkdir build && cd build
+cmake .. -DUSE_UB=ON \
+         -DURMA_INCLUDE_DIR=/usr/include \
+         -DURMA_LIBRARY=/usr/lib64/liburma.so
+
+# 编译
+make -j$(nproc)
+```
+
+**验证**
+
+```bash
+# 检查 UB 传输层是否已注册
+./mooncake_server --list-transports
+# 预期输出: rdma, tcp, nvlink, ub
+```
+
+---
+
+## 运行与测试
+
+**单节点基准测试**
+
+```bash
+# 终端 1: 目标端（Target）
+./transfer_engine_bench \
+    --mode=target \
+    --protocol=ub \
+    --device_name=urma0 \
+    --local_server_name=127.0.0.1 \
+    --metadata_server=P2PHANDSHAKE
+
+# 终端 2: 发起端（Initiator）
+./transfer_engine_bench \
+    --mode=initiator \
+    --protocol=ub \
+    --device_name=urma0 \
+    --metadata_server=P2PHANDSHAKE \
+    --segment_size=8388608 \
+    --batch_size=1\
+    --segment_id=127.0.0.1:$PORT
+```
+
+**多设备基准测试**
+
+```bash
+# 自动发现多个 URMA 设备
+./transfer_engine_bench \
+    --protocol=ub \
+    --device_name=urma0,urma1,urma2,urma3
+```

mooncake-common/common.cmake

@@ -73,7 +73,12 @@ option(USE_ASCEND_HETEROGENEOUS "option for transferring between ascend npu and
 option(USE_MNNVL "option for using Multi-Node NVLink transport" OFF)
 option(USE_CXL "option for using CXL protocol" OFF)
 option(USE_EFA "option for using AWS EFA transport" OFF)
+option(USE_UB "option for using UB protocol transport" OFF)
 
+if (USE_UB)
+  add_compile_definitions(USE_UB)
+  message(STATUS "ub transport is enabled")
+endif()
 if (USE_EFA)
   # Find libfabric headers and library; default to AWS EFA installer path
   find_path(LIBFABRIC_INCLUDE_DIR rdma/fabric.h

mooncake-transfer-engine/example/transfer_engine_bench.cpp

@@ -231,6 +231,9 @@ static void freeMemoryPool(void *addr, size_t size) {
 #endif
     }
 #else
+    if (FLAGS_protocol == "ub") {
+        munmap(addr, size);  // for urma
+    }
     numa_free(addr, size);
 #endif
 }
@@ -459,6 +462,9 @@ static Transport *installTransportFromFlags(TransferEngine *engine) {
         args.get()[0] = const_cast<char *>(nic_priority_matrix.c_str());
         args.get()[1] = nullptr;
         xport = engine->installTransport(FLAGS_protocol.c_str(), args.get());
+    } else if (FLAGS_protocol == "ub") {
+        engine->getLocalTopology()->discover({FLAGS_device_name});
+        xport = engine->installTransport(FLAGS_protocol, nullptr);
     } else if (FLAGS_protocol == "efa") {
         // EFA needs topology discovery to find devices, but auto_discovery
         // would auto-install RDMA transport. Manually discover instead.

mooncake-transfer-engine/example/transfer_engine_bench_with_notify.cpp

@@ -313,6 +313,9 @@ int initiator() {
             args[0] = (void *)nic_priority_matrix.c_str();
             args[1] = nullptr;
             xport = engine->installTransport("rdma", args);
+        } else if (FLAGS_protocol == "ub") {
+            engine->getLocalTopology()->discover({FLAGS_device_name});
+            xport = engine->installTransport(FLAGS_protocol, nullptr);
         } else if (FLAGS_protocol == "tcp") {
             xport = engine->installTransport("tcp", nullptr);
         } else if (FLAGS_protocol == "nvlink") {
@@ -412,6 +415,9 @@ int target() {
             args[0] = (void *)nic_priority_matrix.c_str();
             args[1] = nullptr;
             engine->installTransport("rdma", args);
+        } else if (FLAGS_protocol == "ub") {
+            engine->getLocalTopology()->discover({FLAGS_device_name});
+            engine->installTransport(FLAGS_protocol, nullptr);
         } else if (FLAGS_protocol == "tcp") {
             engine->installTransport("tcp", nullptr);
         } else if (FLAGS_protocol == "nvlink") {

mooncake-transfer-engine/example/transfer_engine_bench_with_retry.cpp

@@ -386,6 +386,9 @@ int target() {
             args[0] = (void *)nic_priority_matrix.c_str();
             args[1] = nullptr;
             engine->installTransport("rdma", args);
+        } else if (FLAGS_protocol == "ub") {
+            engine->getLocalTopology()->discover({FLAGS_device_name});
+            xport = engine->installTransport(FLAGS_protocol, nullptr);
         } else if (FLAGS_protocol == "tcp") {
             engine->installTransport("tcp", nullptr);
         } else {

mooncake-transfer-engine/include/config.h

@@ -65,6 +65,13 @@ struct GlobalConfig {
     int ib_pci_relaxed_ordering_mode = 0;
     bool ascend_use_fabric_mem = false;
     bool ascend_agent_mode = false;
+    // ub config parameters
+    size_t num_jfc_per_ctx = 2;
+    size_t num_jfce_per_ctx = 2;
+    int eid_index = 0;
+    uint64_t max_seg_size = 0x10000000000;
+    size_t max_jfc_e = 4096;  // urma is temporarily using this default value.
+    size_t num_jetty_per_ep = 1;
 };
 
 struct RpcCommunicatorConfig {

mooncake-transfer-engine/include/transfer_metadata.h

@@ -46,16 +46,19 @@ class TransferMetadata {
         std::string name;
         uint16_t lid;
         std::string gid;
+        std::string eid;  // for ub
     };
 
     struct BufferDesc {
         std::string name;
         uint64_t addr;
         uint64_t length;
-        std::vector<uint32_t> lkey;  // for rdma
-        std::vector<uint32_t> rkey;  // for rdma
-        std::string shm_name;        // for nvlink and hip
-        uint64_t offset;             // for cxl
+        std::vector<uint32_t> lkey;         // for rdma
+        std::vector<uint32_t> rkey;         // for rdma
+        std::string shm_name;               // for nvlink and hip
+        uint64_t offset;                    // for cxl
+        std::vector<std::string> tseg;      // for ub/urma
+        std::vector<uint32_t> l_seg_index;  // for ub/urma
     };
 
     struct NVMeoFBufferDesc {
@@ -82,7 +85,7 @@ class TransferMetadata {
     struct SegmentDesc {
         std::string name;
         std::string protocol;
-        // this is for rdma/shm
+        // this is for rdma/shm/urma
         std::vector<DeviceDesc> devices;
         Topology topology;
         std::vector<BufferDesc> buffers;
@@ -113,6 +116,9 @@ class TransferMetadata {
     struct HandShakeDesc {
         std::string local_nic_path;
         std::string peer_nic_path;
+#ifdef USE_UB
+        std::vector<uint32_t> jetty_num;  // for ub/urma
+#endif
 #ifdef USE_BAREX
         uint16_t barex_port;
 #endif