[tcpm] Nagle Algorithm and Delayed ACKs

[Thomas Narten <[e-mail]>]

It has been long known that the Nagle Algorithm and Delayed ACKs
combine in ways that can be problematic. Googling around, I found
Stewart Chesire's writeup
(http://www.stuartcheshire.org/papers/NagleDelayedAck/) as well as an
ancient (!) ID from Greg Minshall
(http://lists.w3.org/Archives/Public/ietf-discuss/1998Dec/0025.html)

The TCP_NODELAY socket option can be used to turn off the Nagle
Algorithm, but applications have to know to enable it -- which they
often don't know, until after they've run into problems and traced
them down to the specific problem.  This can be a long and expensive
process (e.g., involve help desk calls).

What is the current thinking on this problem? Am I right in assuming
that:

 - the Nagle Algorithm is still a SHOULD to implement
 - Delayed ACKs are also a SHOULD to implement?
 - Most everyone provides a TCP_NODELAY type socket option?

Is that all there is to it? Is this considered good enough?

Have there not been in any attempts over the last ten years to update
the above recommendations to make them work together better without an
application needing to set the TCP_NODELAY option, in those cases
where it matters?

Thomas