If you need a network server to respond in 40 ms p99, letting it start swapping is crazy. But it made sense for timesharing undersized computers running jobs that take minutes or hours. You overcommit because your institution couldn't afford two computers.

I/O also used to have a smaller speed penalty. There was just one core and it wasn't so dramatically faster than the rest of the machine. Hell, there used to be faster disks for swap than for the filesystem, and tuning was about scheduling jobs and placing inputs and outputs so as to fully utilize the fs disks.