pe - perf event wrapper

linux perf event c++ wrapper

Measure how much your code costs in terms of hardware instructions, cachemisses, branchmisses and memory allocations.

1. Compile pe

You will need sashamakarenko/makefile project cloned next to the pe working copy.

$> git clone https://github.com/sashamakarenko/makefile.git makefile
$> git clone https://github.com/sashamakarenko/pe.git pe
$> cd pe
$> make

2. Create code snippet

Let us invoke gettimeofday 20 times on the CPU core 3 and get statistics from this. Instrument your code or create pe/src/tests/TestGettimeofday.cpp like this:

#include <pe/Measurement.h>

int main( int argc, char** argv )
{
    pe::Measurement m;
    m.pinToCpuCore( 3 );
    m.addEvent( pe::EventType::cpuCycles );
    m.addEvent( pe::EventType::hwInstructions );
    m.addEvent( pe::EventType::branchInstructions );
    m.addEvent( pe::EventType::llCacheReadMisses );
    m.addEvent( pe::EventType::branchMisses );
    m.addEvent( pe::EventType::memory ); // you will have to use LD_PRELOAD=build/lib/release/libPePreload-1.0.so to capture memory usage
    m.initialize( 20 );
    
    std::cout << "\ngettimeofday:" << std::endl;
    for( int i = 0; i < m.getMaxCaptures(); ++i )
    {
        m.startCapture();
        // this is what we measure
        gettimeofday( &tv, nullptr );
        m.stopCapture();
    }
    m.prepareResults();
    m.printCaptures();
    m.showAverageValues( std::cout );
    m.rewind();
    return 0;
}

3. Compile

$> make

4. Run measurement

Enable perf events:

$> sudo echo -1 > /proc/sys/kernel/perf_event_paranoid

Bring the frequency up

$> sudo cpufreq-set -c 3 -g performance

Run the code

$> make check

Relax the CPU

$> sudo cpufreq-set -g powersave

5. Results

The metrics correspond to 10 calls in a row. Hopefully the column names are self-explanatory. The column nanos/call correspond to a single call time on a 5GHz

Cost of time measurement functions

event	nanos/call	cpu.cycles	hw.instrs	br.instrs	cch.ll.rmiss	br.misses	bus.cycles	cch.l1d.rmiss	cch.l1i.rmiss
gettimeofday	13.2	661	821	140	0	0	3	2	0
clock_gettime	14.3	714	911	170	0	3	3	3	0
rdtsc	7.0	352	61	0	0	0	2	0	0

Functions and methods

See src/tests/TestLibCalls.cpp for details. In order to avoid compiler call evictions, for the functions returning int we actually measure:

externVolatileInt += function();
externVolatileInt += function();
... // repeated 10 times
externVolatileInt += function();

Without LTO

event	nanos/call	cpu.cycles	hw.instrs	br.instrs	cch.ll.rmiss	br.misses	bus.cycles	cch.l1d.rmiss	cch.l1i.rmiss
void()	1.4	70	20	20	0	0	0	0	0
void(int)	0.7	36	30	20	0	0	0	0	0
int(int)	0.9	43	81	20	0	0	0	0	0
inline int base.get()	0.7	37	43	0	0	0	0	0	0
int base.get()	0.9	45	83	20	0	0	0	0	0
int base.*getIntPtr()	1.2	61	123	30	0	0	0	0	0
virtual int base.get()	0.9	43	93	20	0	0	0	0	0
virtual int derived.get()	0.9	43	93	20	0	0	0	0	0
virtual int virtDerived.get()	1.3	67	133	20	0	0	0	0	0
inline base.getIndirect()	1.0	52	46	0	0	0	0	0	0
inline derived.getIndirect()	1.0	48	46	0	0	0	0	0	0

With LTO

event	nanos/call	cpu.cycles	hw.instrs	br.instrs	cch.ll.rmiss	br.misses	bus.cycles	cch.l1d.rmiss	cch.l1i.rmiss
void()	0.0	0	0	0	0	0	0	0	0
void(int)	0.0	0	0	0	0	0	0	0	0
int(int)	1.0	48	30	0	0	0	0	0	0
inline int base.get()	0.9	47	40	0	0	0	0	0	0
int base.get()	0.9	44	40	0	0	0	0	0	0
int base.*getIntPtr()	1.3	64	120	30	0	0	0	0	0
virtual int base.get()	1.0	48	71	20	0	0	0	0	0
virtual int derived.get()	0.9	45	71	20	0	0	0	0	0
virtual int virtDerived.get()	1.5	77	120	20	0	0	0	0	0
inline base.getIndirect()	1.0	48	40	0	0	0	0	0	0
inline derived.getIndirect()	0.9	43	40	0	0	0	0	0	0