add benchmarks

Author: vickiegpt

.github/workflows/cmake.yml

@@ -22,12 +22,12 @@ jobs:
     - uses: actions/checkout@v3
 
     - name: Install dependencies
-      run: sudo apt install llvm-dev clang libbpf-dev libclang-dev python3-pip gcc-13 g++-13 libcxxopts-dev libboost-dev nvidia-cuda-dev libfmt-dev libspdlog-dev && git submodule update --init --recursive
+      run: sudo apt install llvm-dev clang libbpf-dev libclang-dev python3-pip gcc-14 g++-14 libcxxopts-dev libboost-dev nvidia-cuda-dev libfmt-dev libspdlog-dev && git submodule update --init --recursive
 
     - name: Configure CMake
       # Configure CMake in a 'build' subdirectory. `CMAKE_BUILD_TYPE` is only required if you are using a single-configuration generator such as make.
       # See https://cmake.org/cmake/help/latest/variable/CMAKE_BUILD_TYPE.html?highlight=cmake_build_type
-      run: CC=gcc-13 CXX=g++-13 cmake -B ${{github.workspace}}/build -DCMAKE_BUILD_TYPE=${{env.BUILD_TYPE}} -DCMAKE_POLICY_DEFAULT_CMP0091=NEW && sudo cp -r ${{github.workspace}}/include/vmlinux.h /usr/include
+      run: CC=gcc-14 CXX=g++-14 cmake -B ${{github.workspace}}/build -DCMAKE_BUILD_TYPE=${{env.BUILD_TYPE}} -DCMAKE_POLICY_DEFAULT_CMP0091=NEW -DSPDLOG_USE_STD_FORMAT=ON && sudo cp -r ${{github.workspace}}/include/vmlinux.h /usr/include
 
     - name: Build
       # Build your program with the given configuration

.gitmodules

@@ -19,3 +19,9 @@
 [submodule "workloads/llama.cpp"]
 	path = workloads/llama.cpp
 	url = https://github.com/ggml-org/llama.cpp
+[submodule "workloads/vsag"]
+	path = workloads/vsag
+	url = https://github.com/antgroup/vsag
+[submodule "workloads/ann-benchmarks"]
+	path = workloads/ann-benchmarks
+	url = https://github.com/erikbern/ann-benchmarks

include/cxlcontroller.h

@@ -42,7 +42,7 @@ class MigrationPolicy : public Policy {
     virtual ~MigrationPolicy() = default;
 
     // 基本的compute_once方法，决定是否需要执行迁移
-    int compute_once(CXLController* controller) override {
+    int compute_once(CXLController *controller) override {
         auto migration_list = get_migration_list(controller);
         return migration_list.empty() ? 0 : 1;
     }
@@ -54,20 +54,19 @@ class MigrationPolicy : public Policy {
         return migration_list;
     }
     // 判断特定地址是否应该迁移
-    virtual bool should_migrate(uint64_t addr, uint64_t timestamp, int current_device) {
-        return false;
-    }
+    virtual bool should_migrate(uint64_t addr, uint64_t timestamp, int current_device) { return false; }
 
     // 为给定地址选择最佳的目标设备
-    virtual int select_target_device(uint64_t addr, int current_device, CXLController* controller) {
-        return -1;  // -1表示不迁移
+    virtual int select_target_device(uint64_t addr, int current_device, CXLController *controller) {
+        return -1; // -1表示不迁移
     }
 };
 
 // need to give a timeout and will be added latency later,
 class PagingPolicy : public Policy {
 public:
     PagingPolicy();
+    int compute_once(CXLController *) override{};
     // paging related
 };
 
@@ -86,7 +85,7 @@ class CachingPolicy : public Policy {
     }
 
     // 获取需要失效的地址列表
-    virtual std::vector<uint64_t> get_invalidation_list(CXLController* controller) {
+    virtual std::vector<uint64_t> get_invalidation_list(CXLController *controller) {
         return {}; // 默认行为，可以被子类覆盖
     }
 };
@@ -99,7 +98,7 @@ struct LRUCacheEntry {
 
 // LRU缓存
 class LRUCache {
-    public:
+public:
     int capacity; // 缓存容量
     std::unordered_map<uint64_t, LRUCacheEntry> cache; // 缓存映射
     std::list<uint64_t> lru_list; // LRU列表，最近使用的在前面
@@ -155,9 +154,7 @@ class LRUCache {
         lru_map.clear();
     }
     // LRU缓存类中添加size()方法
-    size_t size() const {
-        return cache.size();
-    }
+    size_t size() const { return cache.size(); }
 
     // LRU缓存类中添加remove()方法
     bool remove(uint64_t key) {
@@ -198,7 +195,7 @@ class CXLController : public CXLSwitch {
     double latency_lat{};
     double bandwidth_lat{};
     double dramlatency;
-    std::unordered_map<int, CXLMemExpander*> device_map;
+    std::unordered_map<int, CXLMemExpander *> device_map;
     // ring buffer
     std::queue<lbr> ring_buffer;
     // rob info
@@ -232,7 +229,7 @@ class CXLController : public CXLSwitch {
     void update_cache(uint64_t addr, uint64_t value, uint64_t timestamp) { lru_cache.put(addr, value, timestamp); }
     void perform_back_invalidation();
     void invalidate_in_expanders(uint64_t addr);
-    void invalidate_in_switch(CXLSwitch* switch_, uint64_t addr);
+    void invalidate_in_switch(CXLSwitch *switch_, uint64_t addr);
 };
 
 template <> struct std::formatter<CXLController> {

src/cxlcontroller.cpp

@@ -337,7 +337,7 @@ int CXLController::insert(uint64_t timestamp, uint64_t tid, lbr lbrs[32], cntr c
     // 从当前controller开始DFS遍历
     dfs_calculate(this);
 
-    latency_lat += std::max(total_latency, 0.0);
+    latency_lat += std::max(total_latency + std::get<0>(calculate_congestion()), 0.0);
     bandwidth_lat += std::max(calculate_bandwidth(all_access), 0.0);
 
     return 0;
@@ -421,14 +421,15 @@ void CXLController::invalidate_in_expanders(uint64_t addr) {
 }
 
 // 在交换机及其子节点中执行失效
-void  CXLController::invalidate_in_switch(CXLSwitch* switch_, uint64_t addr) {
-    if (!switch_) return;
+void CXLController::invalidate_in_switch(CXLSwitch *switch_, uint64_t addr) {
+    if (!switch_)
+        return;
 
     // 处理此交换机连接的扩展器
     for (auto expander : switch_->expanders) {
         if (expander) {
             // 从expander的occupation中移除指定地址
-            for (auto it = expander->occupation.begin(); it != expander->occupation.end(); ) {
+            for (auto it = expander->occupation.begin(); it != expander->occupation.end();) {
                 if (it->address == addr) {
                     it = expander->occupation.erase(it);
                     counter.inc_backinv();

src/cxlendpoint.cpp

@@ -251,7 +251,7 @@ double CXLSwitch::get_endpoint_rob_latency(CXLMemExpander *endpoint,
         // 使用哈希表快速检查地址是否属于这个endpoint
         bool is_endpoint_access = endpoint->is_address_local(addr);
 
-        if (!is_endpoint_access)
+        if (is_endpoint_access)
             continue;
 
         double current_latency = base_latency;

src/main.cc

@@ -96,7 +96,7 @@ int main(int argc, char *argv[]) {
 
     auto *policy1 = new InterleavePolicy();
     auto *policy2 = new HeatAwareMigrationPolicy();
-    auto *policy3 = new HugePagePolicy();
+    auto *policy3 = new PagingPolicy();
     auto *policy4 = new FIFOPolicy();
 
     uint64_t use_cpus = 0;

src/monitor.cpp

@@ -207,13 +207,14 @@ int Monitors::terminate(const uint32_t tgid, const uint32_t tid, const int32_t t
         double emulated_time =
             (double)(mon[target].end_exec_ts.tv_sec - mon[target].start_exec_ts.tv_sec) +
             (double)(mon[target].end_exec_ts.tv_nsec - mon[target].start_exec_ts.tv_nsec) / 1000000000;
-        SPDLOG_INFO("emulated time ={}", emulated_time);
-        SPDLOG_INFO("total delay   ={}", mon[target].total_delay);
-
-        SPDLOG_INFO("PEBS sample total {} {}", mon[target].before->pebs.total, mon[target].after->pebs.llcmiss);
-        SPDLOG_INFO("LBR sample total {}", mon[target].before->lbr.total);
-        SPDLOG_INFO("bpftime sample total {}", mon[target].before->bpftime.total);
-        SPDLOG_INFO("{}", *controller);
+        std::cout << std::format("emulated time ={}", emulated_time) << std::endl;
+        std::cout << std::format("total delay   ={}", mon[target].total_delay) << std::endl;
+        std::cout << std::format("PEBS sample total {} {}", mon[target].before->pebs.total,
+                                 mon[target].after->pebs.llcmiss)
+                  << std::endl;
+        std::cout << std::format("LBR sample total {}", mon[target].before->lbr.total) << std::endl;
+        std::cout << std::format("bpftime sample total {}", mon[target].before->bpftime.total) << std::endl;
+        std::cout << std::format("{}", *controller) << std::endl;
         break;
     }
 
@@ -335,7 +336,6 @@ timespec operator*(const timespec &lhs, const timespec &rhs) {
 void Monitor::wait(std::vector<Monitor> *mons, int target) {
     auto &mon = (*mons)[target];
     uint64_t diff_nsec, target_nsec;
-    SPDLOG_ERROR("[{}:{}][OFF] total:", mon.tgid, mon.tid);
     timespec start_ts{}, end_ts{};
     timespec sleep_target{};
     // timespec wanted_delay;