Direct parsing for Nvidia devices from NVML (#297)

Add NVML-based GPU monitor with runtime dlopen loading (#242)

This implements unified NVIDIA monitoring preferring NVML, but supporting nvidia-smi as a fallback.

The *stable* headers needed for the NVML monitoring are copied from `nvml.h`, so as to avoid any build time dependency on CUDA. At runtime a `dlopen` call is made to see if the NVML methods can be used.

---------

Co-authored-by: Graeme Andrew Stewart <graeme.andrew.stewart@desy.de>

Author: EyupKeremBas7

CMakeLists.txt

@@ -6,6 +6,11 @@ cmake_minimum_required(VERSION 3.10)
 # From CMake 3, we can set the project version easily in one go
 project(prmon VERSION 3.2.0 LANGUAGES CXX)
 
+# For newer CMakes, allow target_link_libraries to link targets defined in other directories
+if(POLICY CMP0079)
+  cmake_policy(SET CMP0079 NEW)
+endif()
+
 # For newer CMakes use normalized install destination paths
 if(POLICY CMP0177)
   cmake_policy(SET CMP0177 NEW)
@@ -53,7 +58,7 @@ set(BUILD_GTESTS "${BUILD_GTESTS}"
 
 # Define a default build type when using a single-mode tool like make/ninja
 # We make this default to Release, unless profiling is enabled, in which
-# case do RelWithDebInfo (bcause you need the debug symbols)
+# case do RelWithDebInfo (because you need the debug symbols)
 if(NOT CMAKE_BUILD_TYPE)
   if(PROFILE_GPROF OR PROFILE_GPERFTOOLS)
     set(CMAKE_BUILD_TYPE RelWithDebInfo)
@@ -85,7 +90,6 @@ if("${ARM_PROCESSOR}" STREQUAL "arm")
   message(STATUS "Disabled ABI warnings on ${CMAKE_HOST_SYSTEM_PROCESSOR}")
 endif()
 
-
 # Add package utilities to CMake path
 set(CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/cmake;${CMAKE_MODULE_PATH}")
 
@@ -122,6 +126,7 @@ target_include_directories(prmon PRIVATE ${PROJECT_SOURCE_DIR}/package/include)
 if (BUILD_BENCHMARK_LOG)
   target_include_directories(benchmark-log PRIVATE ${PROJECT_SOURCE_DIR}/package/include)
 endif(BUILD_BENCHMARK_LOG)
+
 #--- create uninstall target ---------------------------------------------------
 include(cmake/prmonUninstall.cmake)
 

README.md

@@ -163,6 +163,17 @@ might not contain all metrics that the default approach supports, e.g.,
 If any of these issues are encountered, a relevant message is printed
 to notify the user.
 
+### NVIDIA GPU Monitoring
+
+When NVIDIA GPUs are present, `prmon` will preferentially use the
+NVIDIA Management Library (NVML) for direct GPU monitoring via
+`libnvidia-ml.so`. If NVML is not available or fails to initialise,
+`prmon` falls back to parsing the output of `nvidia-smi`.
+No extra build-time dependencies are needed: NVML symbols are loaded
+at runtime via `dlopen`, so `prmon` can be compiled on machines without
+NVIDIA drivers and will detect GPU support when run on a machine that
+has them.
+
 ### Environment Variables
 
 The `PRMON_DISABLE_MONITOR` environment variable can be used to specify a comma

package/CMakeLists.txt

@@ -4,17 +4,18 @@ add_executable(prmon src/prmon.cpp
            src/MessageBase.cpp
            src/parameter.cpp
            src/netmon.cpp
+           src/nvidiamon.cpp
            src/iomon.cpp
            src/cpumon.cpp
            src/countmon.cpp
            src/wallmon.cpp
            src/memmon.cpp
-           src/nvidiamon.cpp
            )
 
 target_link_libraries(prmon PRIVATE
            nlohmann_json::nlohmann_json
            spdlog::spdlog_header_only
+           dl
            )
 
 if (BUILD_BENCHMARK_LOG)