feldman%mycroft@GSWD-VMS.ARPA.UUCP (04/22/87)
GVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGV From: feldman%mycroft@gswd-vms.ARPA (Mike Feldman) To: TCP-IP@sri-nic.arpa Subject: Re: Ken Olsen vs MAP Return-Path: <TCP-IP-RELAY@SRI-NIC.ARPA> Redistributed: tcp-ip^.x Received: from SRI-NIC.ARPA by Xerox.COM ; 10 APR 87 15:29:50 PDT Received: from gswd-vms.ARPA by SRI-NIC.ARPA with TCP; Fri 10 Apr 87 06:49:15-PDT Received: from mycroft.GSD (mycroft.Gould.COM) by gswd-vms.ARPA (5.52/) Message-Id: <8704101349.AA14705@gswd-vms.ARPA> Original-Date: Fri, 10 Apr 87 07:47:18 CST GVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGVGV > The answer to your question is that the MAP people do not fully understand > all the sources of probabilistic delay that you catalog. ... > > In their defense however, they have always been concerned about the situation > which might arise when a failure in a manufacturing process causes 500 > sensors to all start sending emergency warning messages at the same instant. > This can lead to lockup on an Ethernet in a way which should not occur on a > token bus. I was envolved in designing an early token bus process control system for a control system manufacturer in '69 - '70, when I first saw the Ethernet spec and articles describing it. I was a junior software engineer at the time, and I suggested to the powers-that-be that Ethernet was much simpler to implement, and that we were already doing it for the token recovery sequence (silence on the wire). It's my opinion that it was dogma among communications engineers even then that you needed deterministic bus access time and guaranteed throughput for control system muxes. There were two cases where the system had to perform under Ethernet-busting loads. One was during customer accecptance testing, and the other was when the reactor pressure relief valve failed open. I got angry reactions when I suggested Ethernet would be cheaper and easier to build and would have better performance almost all of the time. They were even more irritated when I suggested the if determinism and known throughput were so important that they use time division multiplexing. These days, there is a lot of support among control system suppliers for MAP. There hope is that the semiconductor industry will see a large market and implement all the hard, expensive lower layers in silicon, because so far it's been too expensive for each system house to develop one that's a comercial and technical success. It might happen, but they still won't get the deterministic round trip message delay needed to implement stable fast feedback loops. These opinions are my own. Mike Feldman Gould Computer Systems Division 1101 East University Avenue Urbana, IL 61801