Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!rutgers!sri-spam!mordor!lll-tis!ptsfa!ihnp4!homxb!houxm!houxa!mel1
From: mel1@houxa.UUCP (M.HAAS)
Newsgroups: comp.dcom.lans,comp.protocols.tcp-ip
Subject: Re: Smart Ethernet boards
Message-ID: <585@houxa.UUCP>
Date: Fri, 3-Jul-87 15:15:27 EDT
Article-I.D.: houxa.585
Posted: Fri Jul  3 15:15:27 1987
Date-Received: Sat, 4-Jul-87 14:38:27 EDT
References: <283@sering.cwi.nl> <8212@utzoo.UUCP> <17346@amdcad.AMD.COM> <365@parcvax.Xerox.COM>
Organization: AT&T Bell Laboratories, Holmdel
Lines: 23
Xref: mnetor comp.dcom.lans:602 comp.protocols.tcp-ip:489

I think there are two performance factors to consider.  If one wants
to devote CPU cycles to Ethernet processing, then an off-board TCP/IP
makes sense and high network traffic rates can be achieved.  If the
CPU is already burdened by having to handle too many time-share users
or by just being too small, then an on-board TCP/IP makes sense (with
network traffic rates limited by the board/CPU combination's power).

If the on-board TCP/IP imposes more load on the CPU than an off-board
TCP/IP, then something is radically wrong with the design.  If anyone
has evidence of such a situation, please post it here so we can avoid
purchasing that board.

For small systems with 80386 and 68020 class CPUs, I would think that
an on-board TCP/IP with similar CPU and decent memory buffers would give
optimum performance.  Does anyone have figures or thoughts as to what
impedes performance?  One article here commented on the multiple
copies being made of the data being transferred.  Is that necessary
or desirable?  Couldn't the on-board TCP/IP processor just handle
the headers on the board and DMA the data to/from user's I/O space?
(i.e. no copying, just pointer shuffling)  Are there TCP/IP
implementations that do this?

   Mel Haas  ,  odyssey!mel