Re: comments on RFC1323.bis

Vern Paxson <vern@ee.lbl.gov> Thu, 07 May 1998 22:52 UTC

Delivery-Date: Thu, 07 May 1998 18:52:40 -0400
Return-Path: tcplw-relay@services.BSDI.COM
Received: from cnri.reston.va.us (ns.cnri.reston.va.us [132.151.1.1]) by ietf.org (8.8.5/8.8.7a) with ESMTP id SAA29573 for <ietf-archive@ietf.org>; Thu, 7 May 1998 18:52:35 -0400 (EDT)
Received: from services.BSDI.COM (services.BSDI.COM [205.230.225.19]) by cnri.reston.va.us (8.8.5/8.8.7a) with ESMTP id SAA19614 for <IETF-archive@cnri.reston.va.us>; Thu, 7 May 1998 18:55:01 -0400 (EDT)
Received: (from daemon@localhost) by services.BSDI.COM (8.8.7/8.8.8) id QAA29655 for tcplw-list@bsdi.com; Thu, 7 May 1998 16:50:41 -0600 (MDT)
Received: from mailfilter.bsdi.com (mailfilter.BSDI.COM [205.230.225.21]) by services.BSDI.COM (8.8.7/8.8.8) with ESMTP id QAA29652 for <tcplw@bsdi.com>; Thu, 7 May 1998 16:50:40 -0600 (MDT)
Received: from daffy.ee.lbl.gov (daffy.ee.lbl.gov [131.243.1.31]) by mailfilter.bsdi.com (BSDI-MF 1.0) with ESMTP id QAA22231 for <tcplw@bsdi.com> env-from (vern@ee.lbl.gov); Thu, 7 May 1998 16:49:31 -0600 (MDT)
Received: by daffy.ee.lbl.gov (8.8.8/8.8.5) id PAA09087; Thu, 7 May 1998 15:50:36 -0700 (PDT)
Message-Id: <199805072250.PAA09087@daffy.ee.lbl.gov>
To: braden@isi.edu
Cc: tcplw@bsdi.com, floyd@ee.lbl.gov
Subject: Re: comments on RFC1323.bis
In-reply-to: Your message of Thu, 07 May 1998 08:56:47 PDT.
Date: Thu, 07 May 1998 15:50:36 -0700
From: Vern Paxson <vern@ee.lbl.gov>

> Did you notice that Van did not answer your question?  Does that mean
> there is some terrible flaw in the current scheme, so that NO values of
> the constants will actually improve the RTO?

Sorry, I didn't mean to imply that.  Presumably one can find coefficients
for the EWMA that will yield estimates similar to what you get with the
once-per-window estimator, which we know is basically sound, if quite
conservative.  However, the coefficients will be a function of the window
size, instead of fixed as they are today.

Sally believes that figuring out the form for the coefficients to give
equivalent behavior isn't all that hard, and will be posting something
to the list about this quite soon.

> If the answer is "yes",
> we should abandon RTTM now; otherwise, we should be able to specify
> recommended constants in the document, even if they may be improved
> with subsequent research.

I think the best we can do in the near-term is recommend coefficients that
give equivalent behavior (so this still leaves the question, Why bother
with RTTM?).  We might hope in the medium-term to have either better
coefficients, or a modified estimation algorithm per Van's suggestions.
This latter is something I'm going to try to pursue further in my CST
(copious spare time), using the NPD data.

		Vern