Question

If I am right, on Linux (in C/C++, gcc/g++), one can read data from a regular file using read(2) or mmap(2) syscalls.

Two questions. Do read syscall use mmap internally? When is first faster than the second and vice versa?

Answer 1

If you’re reading the file sequentially, my default choice would be to repeatedly read into a largish buffer.

If you’re accessing small bits of data scattered around a large file, the choice is less clear, but mmap could lead to more readable code (since you could code things up as if the file were already in memory). Which would give better performance in this case is hard to tell a priori.

If you’re writing performance-critical code, then the only way to ascertain performance is by benchmarking/profiling actual code.