Why there is a fairly large variance in retired instructions between subsequent runs of 'perf stat'?

Issue

• Why there is a fairly large variance in retired instructions between subsequent runs of perf stat –e instructions <my-executable>. Disabling Address Space Layout Randomization with setarch $(uname –m) –LR does not seem to make any difference.

For example :

# cat testcase.c

#include <stdio.h>
int main(int argc, char* argv[])
{
    int i;
    for (i = 0; i < 100000000; ++i);
    return 0;
}

• If above program is run 5 times via setarch $(uname -m) -RL perf stat -e instructions ./testcase, it shows significantly different results:

# for i in {1..5}; do     setarch $(uname -m) perf stat -e instructions ./testcase;   done |& grep instructions
       300,930,235 instructions              #    0.00  insns per cycle        
       300,939,461 instructions              #    0.00  insns per cycle        
       300,925,166 instructions              #    0.00  insns per cycle        
       300,935,742 instructions              #    0.00  insns per cycle        
       300,930,628 instructions              #    0.00  insns per cycle

• Why there is such variance and if there is something we can do to reduce it?