[e2e] TCP Local Area Normal behaviour? any references?

[Matt Mathis]

On Fri, 21 Jan 2005, Aaron Falk wrote:

> Jon Crowcroft wrote:
> >
> > how fair/efficient is TCP in normal operation when there's no router
> > or buffer in an intermediate node (yes i knoiw some switches have
> > more than 1 packet buffers but ignore those)
>
> I seem to recall Matt Mathis talking about this at IETF around 5 years
> ago in the context of MAC 'capture effects'.
>
> --aaron

Yea, I looked at it, but in the end I decided that the problem is a rathole,
and never really finished anything.   Here are some "hints".

There is a fundamental clash between self clocked (constant) window
protocols (e.g. TCP, especially in congestion avoidance) and channels that
give priority to the current sender.   These channels cause self interference
in the self clock because they couple the forward and reverse paths.

The most common example of a channel that gives priority to the sender is half
duplex Ethernet, with its so-called capture effect, but I believe that this
applies to most channels that are not true full duplex.  (It may even apply to
any channel where there is any interference at all between the forward and return
paths - Including wireless????)

TCP does especially poorly on a long path with a half duplex span that is not
adjacent to the sender.  TCP can do some things to help if the half duplex span
is the first (or only) span because it can inspect the queue in the NIC,
but I would characterize most of these solutions as "hacks".  (And some have
been deployed in some OS's).

BTW, my conjectured general "solution" for Ethernet is that a sending NIC MUST
give up (release) the channel periodically (e.g. every N packets), and must
have at least 2N queue space.  This modulates the entire flow into bursts of
size N, and assures that there is enough queue to maintain the proper average
data rate.

But the more pragmatic solution (adopted here at PSC and may other places) is
to declare half duplex Ethernet to be broken, and eradicate it wherever
possible.  Where not possible, tell people that the maximum theoretical
utilization is 1/e (35%), and they should be pleased if they get any better
than that, because they are operating beyond the designed operating point for
the media.

Good luck,
--MM--
-------------------------------------------
Matt Mathis      http://www.psc.edu/~mathis
Work:412.268.3319    Home/Cell:412.654.7529
-------------------------------------------
Evil is defined by people who think they know
"The Truth" and use force to apply it to others.