Perf records in C++

The library enables a typical workflow that I follow in investions of software performance. It is a tool for dedicated studies of performance in specific sections of software, possibly in connection with some external hardware or software configuration within a larger system.

• It provides a simple interface to the Linux perf_event subsytem that manages access to the Performance Monitor Unit of modern CPUs. And it also manages custom software counters.

  • The goal is to access all PMU events of a given CPU model. I do not really use the general events that have standard nicknames in Linux, like PERF_TYPE_HARDWARE type. Those events are more relevant for system monitoring in administration and operations. This library uses the Intel’s perfmon descriptions to access all PMU events as PERF_TYPE_RAW type.

• It is easy to integrate as one or two drop-in header files.

• It dumps the counters in HTML format that is easy to copy-paste into a Markdown document, which I typically use while logging a study.

• The Markdown is converted into an HTML, which is easy to parse and generate CSV files with the data in the long format. Then I use standard tools to analyse the data, like R and ggplot2.

  • The C++ library does not do any statistical operations, like aggregations or percentile calculations.
  • The goal is to get the data from the CPU and the software, print it out in a logable format, which is then easy to parse and dump into a long-format CSV. All the post-processing and analytics is done with standard tools for data analysis.

There are similar packages:

• libperf - the last commit is from 12 years ago.
• PAPI - a large framework with many large industrial partnerships that is “[Exa-PAPI team is] channeling the monitoring capabilities of hardware counters, power usage, software-defined events into a robust PAPI software package for exascale-level systems.”
• perf-cpp by Jan Mühlig - it is very similar to my tool, although it is somewhat larger (the built library is a couple 10s MB) and it does perform some statistical operations in Cpp. Also, my guess is that there is no intention to implement a workflow with dumping the data into an HTML for a log in Markdown. But, when I get a clearer idea for where my tool is headed, I want to look more into it and check whether there is a possibility to merge some features.