Ian Lance Tayler quotes Perry Metzer as writing:

> One wonders if, given that more apps seem to have performance issues
> with Nagle's algorithm than need it, if we shouldn't invert the sense
> and explicitly set Nagling for rlogin and telnet and disable it for
> other applications.

We should not.  RFC-1123 clearly sets the onus the other way (that is,
as we have it now). A TCP SHOULD implement Nagle, and MUST provide a
way for an application o disable Nagle on a per-connection basis.
(Not implementing Nagle covers the "MUST").