Path: utzoo!attcan!uunet!jarthur!usc!sdd.hp.com!wuarchive!mit-eddie!bloom-beacon!deccrl!news.crl.dec.com!jg
From: jg@crl.dec.com (Jim Gettys)
Newsgroups: comp.windows.x
Subject: Re:  Short circuiting network communication
Message-ID: <1990Dec10.220630.261@crl.dec.com>
Date: 10 Dec 90 22:06:30 GMT
References: <9012091641.AA14521@Larry.McRCIM.McGill.EDU> <1990Dec10.181808.1528@news.arc.nasa.gov>
Sender: news@crl.dec.com (USENET News System)
Reply-To: jg@crl.dec.com (Jim Gettys)
Organization: DEC Cambridge Research Lab
Lines: 20

In article <1990Dec10.181808.1528@news.arc.nasa.gov>, schoch@sheba.arc.nasa.gov (Steve Schoch) writes:
> Of course, this still means that requests will take about 2 memory copies
> to get from client to server (one from client memory to kernel memory and
> one from kernel memory to server memory).  However, using TCP also gives
> the overhead of window-size calculations, checksum verification, IP address
> checking, etc.

Not quite true (on many systems).  A long time ago, in a galaxy far, far,
away (MIT), a hack was put into the top tcp_output in the network code, allowing
local TCP connections to avoid all the low level TCP transport costs of
window size, checksumming, etc..  The data copies do of course happen.

I believe this hack got into more recent Berkeley TCP implementations (certainly
got into Ultrix, shortly after it went in at MIT's), and many manufacturers
may now have picked up the optimization.  

-- 
Jim Gettys
Digital Equipment Corporation
Cambridge Research Laboratory