Date: Tue, 5 Mar 1996 16:19:08 -0800 (PST)
From: Mark Crispin <MRC@CAC.Washington.EDU>
Sender: Mark Crispin <mrc@Tomobiki-Cho.CAC.Washington.EDU>
Subject: Re: re PIPELINING
To: drums@cs.utk.edu
In-Reply-To: <19960305234819.10678.qmail@koobera.math.uic.edu>
Message-Id: <MailManager.826071548.17319.mrc@Tomobiki-Cho.CAC.Washington.EDU>
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; CHARSET=US-ASCII

On 5 Mar 1996 23:48:19 GMT, D. J. Bernstein wrote:
> It is technically inaccurate to call this an ``efficiency feature.''
> It is poor engineering to say that it SHOULD be implemented. Are you
> suggesting that I not point out these problems?

The only thing that D. J. Bernstein has done is observe that in certain cases,
the cumulative real-time cost of many RTTs, even of relatively small
transactions, can be greater than the corresponding bandwidth cost.

This is not a new observation.  Nor are Bernstein's florid hyperbole,
overstated (and incorrect) conclusions, and important omissions.  This last is
rather important; there are some non-obvious scaling and denial of service
issues associated with not aggregating multiple recipients on the same system
of the same message in a single SMTP session.

It is important that "efficiency" not be confused with "real time
performance."  These are two very different things.  In the SMTP delivery
content, parallelism (in the sense of multiple SMTP connections to the same
server to deliver the same message) is not efficient, even if under certain
conditions it can result in faster real-time delivery.