[e2e] on local ethernet throughput?

[Cannara]

This is a good msg in that it exemplifies misconceptions that we all should
have mastered:

1) A "hub" is not an arbitrator, it is simply a MAC extender (originally, more
correctly called a "multiport repeater").  So, a hub has no effect but to
repeat physical signalling (to be within standards) and to lengthen the
physical LAN segment (up to the MAC standard's limits) in terms of stations
and worst-case node-node time delay.  Not understanding that the backplane in
a hub, plus all full-duplex links to similar hubs and their backplanes is
directly equivalent to a coax segment is an unfortunate mistake to make.

2) Some 'hubs' today are mislabelled (e.g., 10/100) for the same reason cars
have had idiot lights.  A 10/100 'hub' should be relabelled a "switch" or
"bridge", because it in fact separates 10 and 100Mb/s ports into distinct
collision domains (segments).  Confusingly, it may not do this for ports
running at the same speed and behave properly as a true hub should -- creating
one segment on which all pkts can be seen from all connected, same-speed
nodes.

3) The old or new measures of Ethernet performance show the same thing (having
been paid to do this over more than 10 years) -- >90% throughput (good data)
easily possible among several contending nodes on the same segment.

4) Finally, the current marketing of "Ethernet" in multi-Gb/s forms, misleads
the reader into believing CSMA/CD has somehow magically evolved, when in fact
these forms are simply Ethernet frames on full-duplex, managed links for which
there are only a sender and receiver -- no contention, just end-system flow
control that may even be proprietary.  Again, myths continue to arise and we
must take the time to ask questions of ourselves and others before believing
manufacturers' glossies.  Even vendors of chips in this arena can mislead in
their product descriptions.

Alex

"David P. Reed" wrote:
> 
> The arguments about local ethernet throughput are interesting, but it's
> worth noting two things:
> 
> 1) Today's Ethernet is not your grandparents' Ethernet.  The analyses of
> the old days are pessimistic compared to the tiny& controlled arbitration
> environments that are presented by hubs, full-duplex, and switches.
> 
> 2) Ethernet was and still is intended to be operated at relatively light
> duty cycles (compared to the max capacity of the shared medium/switch) -
> the behavior as you approach saturation is important because loads are
> peaky, but those peaks are expected to be transient.  So the most important
> thing is that the system not have any hysteresis (history-dependent
> behavior) effect as loads peak and decay back.
> 
> So the original question about "local Ethernet throughput" is actually not
> fully answered by the old studies - in some ways hubs behave a lot more
> like Token Ring than the old coax nets do.