PerfEvent

A header-only C++ wrapper for Linux' perf event API

Basic Usage:

#include "PerfEvent.hpp"
...
PerfEvent e;
e.startCounters();
for (int i=0; i<n; i++) // this code will be measured
  ...
e.stopCounters();
e.printReport(std::cout, n); // use n as scale factor
std::cout << std::endl;

This prints something like this:

cycles, instructions, L1-misses, LLC-misses, branch-misses, task-clock,    scale, IPC, CPUs,  GHz
 10.97,     28.01,      0.22,       0.00,          0.00,       3.89, 10000000, 2.55, 1.00, 2.82

Usage of PerfEventBlock (convenience wrapper):

#include "PerfEvent.hpp"

// Define some global params
BenchmarkParameters params;
params.setParam("name","Dummy Benchmark");
params.setParam("dataSize","100 GB");

for (int threads=1;threads<maxThreads;++threads) {

  // Change local parameters like num threads
  params.setParam("threads",numThreads);

  // Only print the header for the first iteration
  bool printHeader=numThreads==1;

  PerfEventBlock e(n,params,printHeader);
  // Counter are started in constructor

  yourBenchmark();

  // Benchmark counters are automatically stopped and printed on destruction of e
}

This prints something like this:

           name, dataSize, threads, time sec,      cycles, instructions, L1-misses, LLC-misses, branch-misses, task-clock,   scale,      IPC,     CPUs,      GHz
Dummy Benchmark,   100 GB,       1, 1.400645, 1075.520519,  1931.465504,  8.888315,   0.070063,      0.121389, 280.115649, 5000000, 1.795843, 0.999952, 3.839559
Dummy Benchmark,   100 GB,       2, 1.133364, 2386.772941,  2062.313141, 32.095011,   0.043248,      0.918986, 650.737357, 5000000, 0.864059, 1.870823, 3.667798
...

Sometimes the measured counters differ depending on when you construct PerfEvent, for example before vs. after starting threads. You can control this by passing an existing PerfEvent instance to PerfEventBlock:

PerfEvent perfevent;
// start threads etc.
{
  PerfEventBlock perf(perfevent);
  yourBenchmark();
}

Troubleshooting

You may need to run sudo sysctl -w kernel.perf_event_paranoid=-1 and/or add kernel.perf_event_paranoid = -1 to /etc/sysctl.conf