Test whether mmap or read is faster on your computer

mmap_vs_read.c

// As seen on:
// https://federated.saagarjha.com/notice/AaEMQpJBSbxhLyxYzg
// https://twitter.com/_saagarjha/status/1707423903969341949

// Compiling: gcc mmap_vs_read.c -O3 -o mmap_vs_read
// Usage: ./mmap_vs_read <bigfile> <mmap|read>

#include <fcntl.h>
#include <stddef.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/fcntl.h>
#include <sys/mman.h>
#include <unistd.h>

unsigned char hash_buffer(unsigned char *buffer, size_t size) {
	unsigned char hash = 0;
	while (size--) {
		hash ^= *buffer++;
	}
	return hash;
}

unsigned char test_read(int fd) {
	unsigned char hash = 0;
	size_t buffer_size = sysconf(_SC_PAGESIZE) * 16;
	unsigned char *buffer = malloc(buffer_size);
	ssize_t size = 0;
	while (size = read(fd, buffer, buffer_size), size > 0) {
		hash ^= hash_buffer(buffer, size);
	}
	return hash;
}

unsigned char test_mmap(int fd) {
	off_t size = lseek(fd, 0, SEEK_END);
	unsigned char *buffer = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
	return hash_buffer(buffer, size);
}

int main(int argc, char **argv) {
	int fd = open(*++argv, O_RDONLY);
	char *choice = *++argv;
	if (!choice) {
		return EXIT_FAILURE;
	} else if (!strcmp(choice, "read")) {
		return test_read(fd);
	} else if (!strcmp(choice, "mmap")) {
		return test_mmap(fd);
	} else {
		return EXIT_FAILURE;
	}
}

$ vers Mac
MacBook Pro (14-inch, 2021) running macOS Sonoma 14.1 Beta (23B5046f)
$ hyperfine --warmup 3 -N --prepare 'sudo purge' './mmap_vs_read mmap' --prepare 'sudo purge' './mmap_vs_read read' -i
Benchmark 1: ./mmap_vs_read mmap
  Time (mean ± σ):      2.391 s ±  0.256 s    [User: 0.091 s, System: 0.392 s]
  Range (min … max):    2.180 s …  2.957 s    10 runs

  Warning: Ignoring non-zero exit code.

Benchmark 2: ./mmap_vs_read read
  Time (mean ± σ):     932.4 ms ± 132.2 ms    [User: 18.4 ms, System: 343.9 ms]
  Range (min … max):   786.0 ms … 1120.7 ms    10 runs

  Warning: Ignoring non-zero exit code.

Summary
  ./mmap_vs_read read ran
    2.56 ± 0.46 times faster than ./mmap_vs_read mmap

With some slight changes, on a wimpy FreeBSD machine:

% repeat 3 time ./tperf ~/freenas-boot.zfs.gz read
4.195u 1.291s 0:05.48 100.0%5+172k 0+1io 6pf+0w
4.098u 1.387s 0:05.48 99.8%5+173k 0+1io 0pf+0w
4.054u 1.432s 0:05.48 100.0%5+172k 0+1io 0pf+0w
% repeat 3 time ./tperf ~/freenas-boot.zfs.gz mmap
4.628u 2.022s 0:06.66 99.6%5+173k 0+1io 18457pf+0w
4.450u 2.174s 0:06.62 100.0%5+172k 0+1io 18465pf+0w
4.457u 2.118s 0:06.58 99.6%5+173k 0+1io 18454pf+0w

However, in a beefy Mac Pro, with those same changes:

% repeat 3 time ./tperf freenas-boot.zfs.gz mmap
3.090u 1.235s 0:04.52 95.5%0+0k 0+0io 0pf+0w
3.087u 1.187s 0:04.28 99.5%0+0k 0+0io 0pf+0w
3.083u 1.234s 0:04.32 99.7%0+0k 0+0io 0pf+0w
% repeat 3 time ./tperf freenas-boot.zfs.gz read
1.898u 0.817s 0:02.71 99.6%0+0k 0+0io 0pf+0w
1.900u 0.817s 0:02.71 100.0%0+0k 0+0io 0pf+0w
1.901u 0.826s 0:02.72 100.0%0+0k 0+0io 0pf+0w

The changes were to use madvise and break it up into smaller chunks:

unsigned char test_mmap(int fd) {
        unsigned char hash = 0;
        off_t size = lseek(fd, 0, SEEK_END);
        size_t offset = 0;
        size_t buffer_size = sysconf(_SC_PAGESIZE) * 16;
        unsigned char *file = mmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
        posix_madvise(file, size, POSIX_MADV_SEQUENTIAL);

        while (offset < size) {
                size_t chunk_size = MIN(size - offset, buffer_size);
                unsigned char *buffer = file + offset;
                hash ^= hash_buffer(buffer, chunk_size);
                offset += chunk_size;
        }
        return hash;
}

My big FreeBSD system isn't booting, and I'm 5000 miles away, so I'll have to spend effort into figuring why.

And on the big FreeBSD system:

% repeat 5 time ./tperf ./freenas-boot.zfs.gz read
4.899u 1.149s 0:06.04 99.8%	5+168k 0+0io 0pf+0w
4.913u 1.275s 0:06.18 100.0%	5+167k 0+0io 0pf+0w
4.721u 1.416s 0:06.13 100.0%	5+167k 0+0io 0pf+0w
4.940u 1.189s 0:06.13 99.8%	5+167k 0+0io 0pf+0w
4.950u 1.188s 0:06.13 100.0%	5+167k 0+0io 0pf+0w
% repeat 5 time ./tperf ./freenas-boot.zfs.gz mmap
5.526u 1.637s 0:07.16 99.8%	5+168k 0+0io 0pf+0w
5.647u 1.535s 0:07.18 99.8%	5+167k 0+0io 0pf+0w
5.576u 1.575s 0:07.15 99.8%	5+167k 0+0io 0pf+0w
5.593u 1.551s 0:07.14 100.0%	5+167k 0+0io 0pf+0w
5.397u 1.762s 0:07.15 100.0%	5+168k 0+0io 0pf+0w

The big Mac and FreeBSD systems each have hundreds of gigabytes of RAM, and all three systems are SSD-only. (However: on the big FreeSBD system, I used /tmp which is tmpfs; on the smaller one, which doesn't have as much RAM, the test file was on a ZFS dataset.)