What is this

This is a simple C library for collecting and reporting on TLB-related hardware performance counters. It’s built for Linux x86-64, using the Linux perf_event_open() system call.

Configuration and building

Configure using CMake and then build using your favoite build engine like GNU Make or Ninja. For example:

cd libperf-tlb-report
mkdir build
cd build
cmake -G Ninja -D STANDALONE=ON -D USE_GROUPS=OFF ..
ninja

The following cmake options are available:

STANDALONE

Whether to register a constructor and destructor to automatically call perfevents_init() and perfevents_finalize().

If OFF, you need to call these functions yourself in the application you’re linking it to. Include libperf-tlb-report.h for the function prototypes.

Defaults to ON.

USE_GROUPS

Whether to collect all performance events in a single group.

In theory, from what I understand, we would want to do this, so that the performance event counters represent the same period of code execution. In reality, this seems to cause the events to be almost never actually run. Couldn’t find anything online that’d explain this or propose a fix.

Because of this, this option defaults to OFF.

Usage

Configure and build as per the above to create a single static library libperf-tlb-report.a. Link your application to this statically, and you’re done.

Resources

• The only good resouce on the topic I found: the manpage for perf_event_open().
• StackOverflow discussion on using groups

License

The 3-Clause BSD License, see LICENSE.