Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10 5/3/83; site cmu-cs-k.ARPA
Path: utzoo!watmath!clyde!burl!ulysses!allegra!mit-eddie!godot!harvard!seismo!rochester!cmu-cs-pt!cmu-cs-k!tim
From: tim@cmu-cs-k.ARPA (Tim Maroney)
Newsgroups: net.dcom,net.lan,net.unix-wizards
Subject: Re: Wanted: Ethernet/IPC Info
Message-ID: <299@cmu-cs-k.ARPA>
Date: Mon, 11-Mar-85 18:59:04 EST
Article-I.D.: cmu-cs-k.299
Posted: Mon Mar 11 18:59:04 1985
Date-Received: Wed, 13-Mar-85 01:01:33 EST
References: <3912@umcp-cs.UUCP>
Organization: Carnegie-Mellon University, Networking
Lines: 18
Xref: watmath net.dcom:894 net.lan:717 net.unix-wizards:12408

If throughput is really vital, your best bet may be to forget the Internet
protocols entirely.  The overhead is fairly high.  It is easy to modify the
4.2bsd Ethernet drivers to allow your own packet types.  The data from the
packets can be directed to raw or customized socket drivers.  By tailoring
the protocol to the application, I expect you could get significantly better
performance, albeit with some loss of portability.

Incidentally, I have actually done something like this, a raw socket
interface directly to the Ethernet, but the sources belong to IBM or C-MU or
someone else who is not me.  It was easy.  Any experienced 4.2bsd kernel
hacker should have little difficulty and finish in under two weeks
(depending on what fraction of his time is devoted to it, of course).  The
only hassle with raw sockets is that their performance tends to be inversely
proportional to the number of them in use.
-=-
Tim Maroney, Carnegie-Mellon University, Networking
ARPA:	Tim.Maroney@CMU-CS-K	uucp:	seismo!cmu-cs-k!tim
CompuServe:	74176,1360	audio:	shout "Hey, Tim!"