Path: utzoo!mnetor!uunet!husc6!mit-eddie!ll-xn!ames!pasteur!ucbvax!MITVMA.MIT.EDU!KASTEN
From: KASTEN@MITVMA.MIT.EDU (Frank Kastenholz)
Newsgroups: comp.protocols.tcp-ip
Subject: Re: Rumors about the death of the ARPANET
Message-ID: <8803230209.AA02570@ucbvax.Berkeley.EDU>
Date: 22 Mar 88 02:06:57 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 20


Jack, Vint, and friends,

By trusting the local I/O channel I had in mind just the transfer of
data from the hosts main storage to the protocol processor. For example,
the Block Multiplexer channel on an IBM mainframe provides a guaranteed
transfer - if the application issues a write to the channel, the final status
of the operation indicates success or failure. Also, the channel will transfer
data intact, or reports an error to the application. There is little need
(other than paranoia?) to provide an application level checksum on the
transferred data, or a high level ack mechanism across the channel.

End to End checksums, acks, etc are needed. Nolo Contendre. But they
can be between intelligent control units.

(This whole argument also assumes that the protocol processor has a
reasonably powerful CPU and amount of memory - imagine a IBM 3090
dumping data into a 8 Mhz 68000 with 512 Kb of memory!!! :-)

Frank