Quick Start | Contribute
This project holds a library for the Linux perf event APIs. At its core lies an implementation of the perf_event_open (2) syscall, usable as is.
The library also contains several utilities necessary for stable usage of the APIs. The utilities provide error handling, support validation and more.
The focus of this library is highly accurate results without any overhead during measuring. That is, setting up measurements may be costly, but performing measurements requires zero overhead added by this library. This comes with a great benefit of achieving as accurate results as possible and being as flexible as possible, at the cost of UX.
To build the project, first install libcap:
sudo apt install libcap-devsudo yum install libcap-develsudo zapper install libcap-develThen, simply build the project.
make buildExamples are available in the examples directory, output to the build/examples directory.
./build/examples/full
Quickstart
Features
Documention
Contribute
CAP_PERFMON added in 5.9)There are many profiling tools available for Linux. Tools such as Valgrind (and Callgrind) are great and accurate ways of performing performance analysis of memory and the callstack. It does, however, produce simplified measurements for values such as CPU cycles and instructions since it doesn't measure native hardware. Other tools such as gprof, prof and gperftools are much closes to the hardware, but lack support for measuring kernel and idle events. Other ways of measuring instructions and cycles used in SUPERCOP, Papi etc. can be inaccurate, simply counting the overall cycles performed by a CPU core. This is where the Linux perf APIs shine. The APIs provide a comprehensive set of tools for measuring user and kernel events - be it via software defined breakpoints or hardware counters.
The perf CLI tool is great at providing a way of using these APIs, but it measures the execution of an entire program. The underlying system call, perf_event_open (2), has no glibc wrapper and is therefore quite verbose - especially when performing a lot of measurements.
This library aims to fill this small niche gap where highly accurate measurements of specific code in a binary.
To be continued as the API stabilizes.
Any contribution is welcome. If you're not able to code it yourself, perhaps someone else is - so post an issue if there's anything on your mind.
Make sure you meet the following prerequisites:
$CC refers to gcc 8 or newerlibcap is installedclang-format is installed (if you want to format code)compiledb is installed (if you want format code or analyze it with other llvm tools)# Clone the repository
git clone https://github.com/AlexGustafsson/perf && cd perf
# Format the code
make format
# Build the library
make library
# Build examples
make examplesThe examples can be tested using Docker.
docker build -t perf .
docker run -it --rm --privileged perf ./pimmaped events
