[e2e] on local ethernet throughput?

[Cannara]

Dan,

Do read the Dec WRL report from 1988, which answered this issue quiet clearly
and competently, by controlling one variable most folks just jabber (excuse
the pun) about:  Offered Load.  If you're interested, I can email a copy.  In
essence, depending on frame size, 20 or so nodes can maintain >90% throughput
at the DLC level on a classic CSMA/CD segment.  The paper was written to
dispel the then prevalent notion promulgated by IBM, for instance, that Token
passing was superior to CSMA/CD in guaranteeing LAN throughput.  It also shows
the fragility of the logic claiming a "capture effect" when all adpators are
playing by the same rules (which Seeq in the '80s tried, until punished, to
violate).  

The DEC team included one of the Xerox Parc contributors to Metcalfe's
invention and their paper makes clear why IBM's marketing eventually failed,
not without cost to many organizations who bought it and TR over more than a
decade.  When IBM started shipping 9xxx and AS4xx boxes with Ethernet
adaptors, it was a quiet admission of failure spawned by a '70s refusal to
license anything from Xerox.

Alex
Menlo Park, Calif.


Vernon Schryver wrote:
> 
> > From: Dan HE <he at enseirb.fr>
> 
> > I was confused by one on the questions of the available bandwidth in a
> > local segment.
> > This question could be very fundemantal but it would be confused.
> > Here is a graph that addresses a segment of ethernet topology.
> >
> >
> >            |A|     |B|
> >             |       |
> > --------------------------------- 10Mb
> >     |          |           |
> >    |A1|       |C|         |B1|
> >
> > There are five stations sharing a segment of ethernet. every stations
> > have
> > 10Mb/s interface card. A and A1 are communicating, B and B1 are
> > communicating.
> > C is an observer.
> > (1) how much maximum throughput can be reached between the communicating
> > pairs(A-A1, B-B1) at the same time? is it about 10Mb/s for both (A-A1)
> > and (B-B1)? or when communicating, both pairs only can share average
> > half of 10Mb/s bandwidth?
> > is it possible to say that adding both the average throughput of (A-A1)
> > and the average throughput of (B-B1)  is lager than 10Mb/s!?
> 
> If we are talking about an 802.3 10MHz CSMA/CD network instead of
> something involving multi-port bridges ("switches" in marketoon talk),
> then the total available TCP/IP payload bandwidth is less than 10
> Mbit/sec.  If all of the stations use classic BEB with fully compliant
> 802.3 state machines and no BLAM and no "polite" MACs like the venerable
> AMD 7990 LANCE and all use subtantial TCP/IP windows, then the total
> throughput will be about 6.5 Mbit/sec because of the Ethernet Capture
> Effect.  If all of the MACs use BLAM instead of BEB, all are "polite"
> (delay when they ought to defer), or if all of the TCP windows are
> tiny, then the total throughput will approach 9.3 Mbit/sec.
> 
> > (2) for the observer, how much bandwidth has been used in this segment
> > when the others
> > are communicating? because in some adaptive applications based on the
> > local segment adaptation, they usually addressed that they are measuring
> > the local available bandwidth.
> > but I don't know how a C can infer the available bandwidth on the link?
> 
> I don't understand that question.  If it is intended to ask about the
> fairness of CSMA/CD with or without BLAM, then the answer is that
> CSMA/CD is very fair.  Except for the Capture Effect, which is in some
> sense fair, the available bandwidth is fairl distributed among stations
> that want to transmit.  If there are 2 stations transmitting, each
> will have access to about 50% of the available 10 MHz (which is less
> than 10 Mbit/sec because of TCP/IP headers, IFG's, and collisions).
> If there are 5 stations, as would be the case with a heavy TCP stream
> from A to A1 (don't forget TCP ACKs), a heavy TCP stream from B to
> B1, and some from C, then each will have access to about 20%.
> 
> The most extensive collection of measurements and simulations on
> that issue is probably Mart Molle's.  The last I heard it was at
> ftp://ftp.cs.ucr.edu/pub/blam
> 
> That collection also includes discussions of BLAM, the capture
> effect, and related issues.
> 
> Finally, before giving credence to the old 802.5 and FDDI token ring
> salescritter and trade rag exspurt claim that CSMA/CD collisions
> consume lots of bandwidth, use a 'scope or a calculator.  (This is
> only a pre-emptive note.  It's nice that so far in this thread that
> old silly lie has not been revived.)
> 
> Vernon Schryver    vjs at rhyolite.com