For most programs, small payloads are the critical issue.

This comparison was measured on a 120Mhz Pentium machine, in the last implementation that incorporated a fast IPC path.

The moral, such as it is, is that the cost of boundary crossing determines the scale of decomposition that is feasible. A crossing for every function call (which several early capability systems tried) won't work on current hardware (though it would be a lot closer on a RISC machine).