Network problem

[Chris Brenton]

Tom Fogal wrote:
> once again ive forgotten to change the 'To:' to the list... *sigh*


and my response was....

-------- Original Message --------
Subject: Re: Network problem
Date: Fri, 29 Aug 2003 07:00:14 -0400
From: Chris Brenton <cbrenton at chrisbrenton.org>
To: Tom Fogal <tfogal at io.iol.unh.edu>
References: <200308290133.h7T1XRhm004270 at io.iol.unh.edu>

Tom Fogal wrote:
 >>Stephen Ingham wrote:
 >>
 >>>A good repeater hub will automatically turn off a port when 32 
consecutive
 >>>collisions are detected.
 >>
 >>Humm. A collision is defined as a system following the Ethernet CSMA/CD
 >>rules that detects a different bit pattern on the receive pair Vs. what
 >>it is currently sending out on the transmit pair. Since a hub is little
 >>more than a line amplifier, it does not "transmit" per CSMA/CD, and thus
 >>has no way of detecting collisions. If a hub had this type capability,
 >>we really would not need switching as it would be trivial to retime the
 >>circuit as well. My guess is you are thinking of a switch or perhaps a
 >>dual speed hub (which is still a form of a bridge).
 >
 >
 > A collision is actually the detection of an incoming frame while a 
device is
 > transmitting. It is not necessarily a different pattern;

Actually, it is. The system expects to see exactly the same pattern its
transmitting on the receive pair (remember the hub transmits the bits
out every port, including the one transmitting). If the pattern on the
receive pair does not match, its assumed there was some kind of error
(bad cable, collision, or what ever) has occurred.

I *think* what you may be thinking of is full duplex mode, as the mode
effectively tells a system to only follow CSMA. This effectively tells a
system "it's OK to see something different on the receive pair while you
are transmitting". Of course you can't support full duplex with a hub as
its a simple amplifier (or you can try to do full duplex with a hub but
it hoses the network ;-).

At least that's how we used to program NIC's when I was at 3COM. ;-)

 > 100mbs devices
 > constantly send out all '1's, and this could of course easily happen 
within
 > a frame.

OK, but I don't remember this either. If this did occur I'm not sure how
you would ever get a clean line to start the preamble, especially in a
large environment. :( The "MA" portion of CSMA states that if you don't
have anything useful to transmit you should remain quiet so others can
use the wire.

 > A 'repeater' is the term used in the standard, and to be compliant 
they must
 > implement CSMA/CD.

Again, if a hub did follow CSMA/CD then topology rules would apply to
every cable hooked into the hub, not to the entire collision domain. For
example if hubs/repeaters really did do CSMA/CD, the maximum overall
collision domain spec (3K ft for 10 Mb/s, 650 Ft for 100 Mb/s, etc.)
would apply to each cable off of the hub, rather than the entire
topology (up to a layer 2 or 3 device like a switch, bridge, or router).

 > The timing of transmissions is supposed to be set when a device starts
 > receiving the preamble of an incoming frame (this is why the preamble 
exists).
 > Not to say it must sync clocks via the preamble, its just what the 
preamble was
 > intended to do.

Agreed, which also implies that the device has some form of memory
storage to queue the packet before transmitting the preamble. Hubs do
not contain any such storage as they are simple amplifiers. Switches do,
which is why they retime the circuit, as well as extend the Ethernet
overall length rules.

 > 'Runts', or fragments of a packet should be discarded by a switch.
 > I forget about bad CRCs, but I believe a switch should drop the 
packet without
 > forwarding.

Again, it depends on the switch. If it's store and forward, absolutely.
If its running in cut through mode, it may not. I've actually tested
this (again 3 COM gear) and seen runts and bad CRC's come out the other
side when a switch runs in cut through mode. This is one of the reasons
why store and forward is so popular (that and it allows you to run a
mixed speed network).

 > apparently people have seen this, but it seems very awkward to me. 
only 4 of
 > the 8 wires are used in networks

Ya, you have to try really hard to shoot yourself in the foot this way. ;-)

 > <=100mpbs, it would seem to me a device
 > manufacturer would hook up the two 'idle' pairs to particularly strong
 > resistors, if anything at all - not the main circuit of their networking
 > device or anywhere it could do damage. of course, i dont design 
circuits,
 > either...

The idle pairs are usually a dead circuit. The problem occurs when you
do something like hook up + on receive to the ground reference on
transmit. Again, this goes back to what I was saying about hubs being
dumb amplifiers. When one system changes the ground reference that's
propagated through the collision domain because the hub changes the
ground reference on all connected systems. If it actually followed
CSMA/CD, it would shunt the signal at that point and only the one system
would be effected.

 > the worst ive seen with a bad cable is devices not establishing a link.

That's the usual. As I said, you have to try *really* hard to shoot
yourself in the foot this way. ;-)

Thanks for the note!
C