This project provides a tiny C++ profiling library for monitoring:

• execution time
• CPU(s) usage
• memory usage
• GPU(s) usage and memory

This library aims at collecting metrics on embedded devices to monitor device performance while operating heavy tasks or booting for example. Those metrics can be useful to check that load is properly spread onto all CPU cores or that memory is not starved.

This library can also run on non-embedded devices like servers or desktop PCs. It is compatible with Linux and Windows.

Metrics are stored in a CSV file (the path is configurable).

The library is lightweight, simple and easy to use. It can be easily used from an existing application or integrated in a dedicated monitoring application.

Full API documentation is available here.

It will generate two rotating files with the most recent events. With the above example, both files will be uprofile_0.log and uprofile_1.log.

The library also supports GPU metrics monitoring like usage and memory. Since GPU monitoring is specific to each vendor, an interface IGPUMonitor is available to abstract each vendor monitor system.

To monitor a specific GPU, you must subclass IGPUMonitor:

As you can see from the interface methods, ccpuprofile supports multi-gpu monitoring.

And then inject it at runtime to the uprofile monitoring system:

Here is the list of GPUs supported by cppuprofile

• NVidia Graphics Cards (through nvidia-smi). Pass -DGPU_MONITOR_NVIDIA=ON as compile option and inject uprofile::NvidiaMonitor from monitors/nvidiamonitor.h as GPUMonitor. The nvidia-smi tool should be installed into /usr/bin directory.

The build process is based on CMake. Minimum version is 2.8.

By default, it generates a shared library on Linux and a dynamic library (DLL) on Windows. To link with this library on Windows, you must pass -DUPROFILE_DLL definition to CMake.

If you want to generate a static library, you must use -DBUILD_SHARED_LIBS=OFF CMake option.

If you want to disable profiling in Release mode or if you want to only enable profiling in particular cases, you can use the PROFILE_ENABLED option (set to ON by default).

To disable the profiling:

The project also brings a tool for displaying the different metrics in a single view:

This tool is written in Python3. It requires a set of dependency packages. To install them:

Then

Note that you can filter the metrics to display with --metric argument.

The project provides a C++ sample application called uprof-sample that shows how to use the cppuprofile library. You can build it with SAMPLE_ENABLED option:

A python binding pyuprofile is available for exposing cppuprofile library APIs to Python.

pyuprofile depends on PyBind11, so you need to install this library first:

For Ubuntu:

Then, to build and install the module in your system

The library compiles on Windows but only time execution is supported so far. Monitoring metrics like CPU Usage and system, process and nvidia GPU memory are not supported.

Contributions are welcomed.

This project is licensed under BSD-3-Clause license. See LICENSE file for any further information.