On the surprising behaviour of memory operations at high thread counts

Must read post about multicore I/O scalability using mmap() under Linux