The man pages really do try to put you off using the syscalls. I only used getdents for the first time when I was writing a syscall interface for Lua https://github.com/justincormack/ljsyscall - not tried it on 8m files but you can set the buffer size so it should work.

The native aio interface has advantages over the posix wrapper too, and some of the other interfaces. On the other hand the futex interface is not nice to use, as it requires architecture specific assembly which seems undocumented...