Removing T/TCP and replacing it with something simpler

[Sean Chittenden]

> I intend to remove T/TCP (transactional TCP) support from our TCP
> implementation for the following reasons:

The special cases are evil.

> However something like T/TCP is certainly useful and I know of one 
> special
> purpose application using it (Web Proxy Server/Client for high-delay 
> Satellite
> connections).

Actually, there are two/three programs that I know of that use it.  
memcached(1), which found a fantastic decrease in its benchmarks.  
Here's an excerpt from the following link:

http://lists.danga.com/pipermail/memcached/2003-August/000111.html

> I benchmarked 3 different methods of doing network I/O:
>
>    1) the default way, with the Nagel algorithm.
>    2) using TCP_CORK (Linux, same as TCP_PUSH on BSD)
>    3) using TCP_NODELAY
>
> I measured both real time and number of packets on the wire.
> The test was doing 2,500 deletes, sets, and gets, then a 2,500 
> get_multi.
>
>            Seconds     Packets
> DEFAULT     102.48      22638
> TCP_CORK      3.88      15105
> TCP_CORK      3.87      15108
> TCP_CORK      3.86      15105
> TCP_NODELAY   3.99      20169
> TCP_NODELAY   4.04      20170
> TCP_NODELAY   4.00      20170

and an internal reverse proxy server/modified apache that I've hacked 
together (reduces latency in a tiered request hierarchy a great deal, 
on order of the benchmarks from above).

That said, I can't stress enough the usefulness of TCP_NOPUSH/ttcp(4).  
However it gets implemented, I don't really care.  If ttcp(4) is ready 
to bite the dust, so be it, but extensions/options like this are 
fantastically useful to those who know about its existence/usefulness.  
Good luck with the replacement.  :)

In 2001, there was a push to make Linux's TCP_CORK option behave the 
same as FreeBSD's TCP_NOPUSH.  Is maintaining that compatibility still 
a goal, or are we going to kick this change off to the Linux folk to 
have them play catchup (to what sounds like a more secure option than 
Linux's TCP_CORK)?

http://seclists.org/linux-kernel/2001/Feb/0993.html

-sc

-- 
Sean Chittenden