[759] in Info-AFS_Redistribution
Re: mail & afs
daemon@ATHENA.MIT.EDU (Craig_Everhart@transarc.com)
Fri May 8 16:51:40 1992
Date: Fri, 8 May 1992 15:05:46 -0400 (EDT)
From: Craig_Everhart@transarc.com
To: Info-AFS@transarc.com
In-Reply-To: <0e2fuiq00WAqQc5ZVb@andrew.cmu.edu>
Transarc uses AMDS (and AMS) for its transarc.com AFS cell. AMDS is the
local mail delivery system, responsible for accepting messages from
anywhere and either delivering them locally into separate files in
~userid/Mailbox or queueing them for remote delivery by handing them to
a long-haul mailer like sendmail. As far as I know, AMDS is sensitive
to all AFS transient errors and postpones delivery until they pass; I
don't think that the same is true for the other solutions that piggyback
on /usr/spool/mail semantics.
While I can't contradict Wallace Colyer's points about AFS/AMDS
performance, I'll note that some of the limitations he spoke of reflect
the enormous andrew.cmu.edu installation, with about 2000 ``AMS
bboards''. Certainly you don't have to use that many bboards to use
AMDS, or even AMS (which is more the user-interface dual of AMDS).
AMS and AMDS are a bit more research-y than AFS. The fact that they're
all built on top of AFS is an advantage and a disadvantage, but it
seemed like an overriding advantage in 1985, when there was already a
collection of ``black-belt programmers'' (a Jim Morris term) building a
highly available, highly distributed system. The cross-cell stuff
(AMS/UI validation of non-local addresses, AMDS delivery via AFS to
non-local addresses) is a neat idea that runs aground in oddball cases,
but it hasn't caused us much trouble for a long time. It's been useful
even within cells at CMU: the AMDS daemons on andrew.cmu.edu and cmu.edu
refer to each others' databases to get mail through with minimal
hassles. cs.cmu.edu is set up to allow interactive address validation,
though not delivery.
As to DFS, I've worked pretty hard to allow an AMDS/AMS successor to be
able to work, but Wallace is right in that there's some work to be done.
Craig
