/* Measure Linux (or whatever) system call overhead: minimally about
 * 300 ns on this laptop [clarification needed].  Copying 8192 zero
 * bytes into userspace seems to take about 1400 ns more, or 171 ps
 * per byte.

On this Ryzen 5 3500U, mostly running at 3.667GHz, I get:

                          1234567
    $ \time ./syscallovh 10000000 1
    0.81user 5.02system 0:05.84elapsed 99%CPU (0avgtext+0avgdata 1152maxresident)k
    0inputs+0outputs (0major+72minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 1
    0.86user 4.97system 0:05.83elapsed 99%CPU (0avgtext+0avgdata 1316maxresident)k
    0inputs+0outputs (0major+73minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 1
    0.85user 5.08system 0:05.94elapsed 99%CPU (0avgtext+0avgdata 1240maxresident)k
    0inputs+0outputs (0major+74minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 1024
    0.79user 5.50system 0:06.29elapsed 99%CPU (0avgtext+0avgdata 1248maxresident)k
    0inputs+0outputs (0major+75minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 1024
    0.84user 5.56system 0:06.40elapsed 99%CPU (0avgtext+0avgdata 1328maxresident)k
    0inputs+0outputs (0major+75minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 1024
    0.83user 5.43system 0:06.26elapsed 100%CPU (0avgtext+0avgdata 1268maxresident)k
    0inputs+0outputs (0major+74minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 8192
    0.92user 7.50system 0:08.43elapsed 99%CPU (0avgtext+0avgdata 1256maxresident)k
    0inputs+0outputs (0major+76minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 8192
    0.85user 7.46system 0:08.31elapsed 99%CPU (0avgtext+0avgdata 1152maxresident)k
    0inputs+0outputs (0major+72minor)pagefaults 0swaps
    $ \time ./syscallovh 10000000 8192
    0.84user 7.52system 0:08.37elapsed 99%CPU (0avgtext+0avgdata 1148maxresident)k
    0inputs+0outputs (0major+75minor)pagefaults 0swaps
    $ \time ./syscallovh 10000 8192000
    0.00user 9.56system 0:09.57elapsed 99%CPU (0avgtext+0avgdata 9144maxresident)k
    0inputs+0outputs (0major+539minor)pagefaults 0swaps
    $ \time ./syscallovh 10000 8192000
    0.00user 9.79system 0:09.79elapsed 99%CPU (0avgtext+0avgdata 9328maxresident)k
    0inputs+0outputs (0major+542minor)pagefaults 0swaps
    $ \time ./syscallovh 10000 8192000
    0.01user 9.65system 0:09.66elapsed 99%CPU (0avgtext+0avgdata 9264maxresident)k
    0inputs+0outputs (0major+540minor)pagefaults 0swaps

So:

- the loop reading 1 byte took 583–594 nanoseconds per read;
- 1024 bytes took 626–640ns per read, an additional 32–57ns, or
  31–56ps per byte; and
- 8192 bytes took 831–843ns per read, an additional 191–217ns over
  the 1024-byte number, or 26–30ps per byte.
- 8192000 bytes took 957–979μs per read, which is 116–120ps per
  byte.
 
 */
#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>

char *devzero = "/dev/zero";

int main(int argc, char **argv) {
  if (argc != 3) {
    fprintf(stderr, "%s: usage: %s 102 4096\n"
            "Reads 4096 bytes from /dev/zero 102 times.\n",
            argv[0], argv[0]);
    return 1;
  }

  int n = atoi(argv[1]);
  size_t s = atoi(argv[2]);
  char c[s];
  int fd = open(devzero, O_RDONLY);
  if (fd < 0) {
    perror(devzero);
    return 1;
  }

  for (int i = 0; i < n; i++) {
    read(fd, c, s);
  }
  return 0;
}