feat: add GPU Collector based on DCGM (alibaba#2360)

* feat: add GPU Collector based on DCGM

* fix: correct spelling error and add Linux platform protection

* fix: update copyright year to 2025 and modify comments to English

Author: shaochen0

.gitmodules

@@ -1,3 +1,6 @@
 [submodule "core/_thirdparty/coolbpf"]
 	path = core/_thirdparty/coolbpf
 	url = https://gitee.com/anolis/coolbpf.git
+[submodule "core/_thirdparty/DCGM"]
+	path = core/_thirdparty/DCGM
+	url = https://github.com/NVIDIA/DCGM.git

core/_thirdparty/DCGM

@@ -45,5 +45,8 @@ const int EXECUTE_FAIL = -1;
 const std::filesystem::path PROCESS_NET_IF_INET6 = "net/if_inet6";
 const int64_t SYSTEM_HERTZ = sysconf(_SC_CLK_TCK);
 const long PAGE_SIZE = sysconf(_SC_PAGESIZE);
+const std::string NVSMI = "nvidia-smi";
+const std::string LIB_DCGM = "dcgm";
+const std::filesystem::path NVIDIACTL = "/dev/nvidiactl";
 
 } // namespace logtail

core/host_monitor/Constants.cpp

@@ -17,6 +17,7 @@
 #pragma once
 
 #include <filesystem>
+#include <string>
 
 #include "common/StringView.h"
 
@@ -56,5 +57,8 @@ inline constexpr StringView DEFAULT_USER_ID_LABEL = "user_id";
 #else
 inline constexpr StringView DEFAULT_HOST_IP_LABEL = "host_ip";
 #endif
+const extern std::string NVSMI;
+const extern std::string LIB_DCGM;
+const extern std::filesystem::path NVIDIACTL;
 
 } // namespace logtail

core/host_monitor/Constants.h

@@ -36,6 +36,7 @@
 #include "host_monitor/HostMonitorTimerEvent.h"
 #include "host_monitor/collector/CPUCollector.h"
 #include "host_monitor/collector/DiskCollector.h"
+#include "host_monitor/collector/GPUCollector.h"
 #include "host_monitor/collector/MemCollector.h"
 #include "host_monitor/collector/NetCollector.h"
 #include "host_monitor/collector/ProcessCollector.h"
@@ -63,6 +64,7 @@ HostMonitorInputRunner::HostMonitorInputRunner() {
     RegisterCollector<DiskCollector>();
     RegisterCollector<ProcessCollector>();
     RegisterCollector<NetCollector>();
+    RegisterCollector<GPUCollector>();
 
     size_t threadPoolSize = 1;
     // threadPoolSize should be greater than 0

core/host_monitor/HostMonitorInputRunner.cpp

@@ -36,6 +36,7 @@ using namespace std::chrono;
 #include "common/FileSystemUtil.h"
 #include "common/StringTools.h"
 #include "host_monitor/Constants.h"
+#include "host_monitor/SystemInformationTools.h"
 #include "host_monitor/common/FastFieldParser.h"
 #include "logger/Logger.h"
 
@@ -1148,4 +1149,48 @@ bool LinuxSystemInterface::GetProcessOpenFilesOnce(pid_t pid, ProcessFd& process
 
     return true;
 }
+
+bool LinuxSystemInterface::InitGPUCollectorOnce(const FieldMap& fieldMap) {
+    if (!CheckGPUDevice()) {
+        return false;
+    }
+
+    if (!mDcgmCollector.IsLibraryLoaded()) {
+        LOG_ERROR(sLogger, ("GPU collector initialization failed", "DCGM library not loaded"));
+        return false;
+    }
+
+    if (!mDcgmCollector.CanInitialize()) {
+        if (mDcgmCollector.IsFullyInitialized()) {
+            LOG_INFO(sLogger, ("GPU collector already initialized", "skipping initialization"));
+            return true;
+        } else {
+            LOG_ERROR(sLogger, ("GPU collector initialization failed", "DCGM collector in invalid state"));
+            return false;
+        }
+    }
+
+    if (!mDcgmCollector.Initialize(fieldMap)) {
+        LOG_ERROR(sLogger, ("GPU collector initialization failed", "DCGM initialization error"));
+        return false;
+    }
+
+    return true;
+}
+
+bool LinuxSystemInterface::GetGPUInformationOnce(GPUInformation& gpuInfo) {
+    if (!mDcgmCollector.IsFullyInitialized()) {
+        LOG_ERROR(sLogger, ("GPU data retrieval failed", "DCGM collector not ready"));
+        return false;
+    }
+
+    bool success = mDcgmCollector.Collect(gpuInfo);
+    if (!success) {
+        LOG_ERROR(sLogger, ("GPU data retrieval failed", "collection operation failed"));
+    } else {
+        LOG_DEBUG(sLogger, ("GPU data retrieval successful", "metrics collected")("gpu_count", gpuInfo.stats.size()));
+    }
+
+    return success;
+}
 } // namespace logtail

core/host_monitor/LinuxSystemInterface.cpp

@@ -17,6 +17,8 @@
 #pragma once
 
 #include "common/ProcParser.h"
+#include "host_monitor/Constants.h"
+#include "host_monitor/SystemInformationTools.h"
 #include "host_monitor/SystemInterface.h"
 
 namespace logtail {
@@ -32,7 +34,7 @@ class LinuxSystemInterface : public SystemInterface {
     }
 
 private:
-    explicit LinuxSystemInterface() : mProcParser("") {}
+    explicit LinuxSystemInterface() : mProcParser(""), mDcgmCollector(LIB_DCGM) {}
     ~LinuxSystemInterface() = default;
 
     bool GetSystemInformationOnce(SystemInformation& systemInfo) override;
@@ -52,6 +54,9 @@ class LinuxSystemInterface : public SystemInterface {
     bool GetDiskStateInformationOnce(DiskStateInformation& diskStateInfo) override;
     bool GetFileSystemInformationOnce(std::string dirName, FileSystemInformation& fileSystemInfo) override;
 
+    bool InitGPUCollectorOnce(const FieldMap& fieldMap) override;
+    bool GetGPUInformationOnce(GPUInformation& gpuInfo) override;
+
     uint64_t GetMemoryValue(char unit, uint64_t value);
     bool GetProcessCmdlineStringOnce(pid_t pid, ProcessCmdlineString& cmdline) override;
     bool GetProcessStatmOnce(pid_t pid, ProcessMemoryInformation& processMemory) override;
@@ -67,5 +72,6 @@ class LinuxSystemInterface : public SystemInterface {
     bool GetInterfaceConfig(InterfaceConfig& interfaceConfig, const std::string& name);
 
     ProcParser mProcParser;
+    DCGMCollector mDcgmCollector;
 };
 } // namespace logtail