Network problem

[Stephen Ingham]

According to the Ethernet standards I found at
"http://www.erg.abdn.ac.uk/users/gorry/course/lan-pages/hub.html" the 
"Auto Partition" feature that Cabletron and many other quality vendors use
is optional. Therefore a miswired cable can bring down the entire network if
you use cheap hubs.


-----Original Message-----
From: gnhlug-discuss-admin at mail.gnhlug.org
[mailto:gnhlug-discuss-admin at mail.gnhlug.org] On Behalf Of Chris Brenton
Sent: Friday, August 29, 2003 8:10 AM
To: Stephen Ingham
Cc: 'Neal Richardson'; discuss at gnhlug.org
Subject: Re: Network problem

Stephen Ingham wrote:
> A repeater is more complicated than a "line amplifier" as Chris described
it
> below.

Actually, its not. Of course vendors are always free to add features. 
Routers by definition simply route traffic and don't filter it, and yet 
that's a feature most of us have learned to expect from a routing device.

> The following is a quote from "Cabletron Systems - Ethernet Technology
> Guide" Page 7-4.

LOL! Oh ya, let's quote Cabletron because they are *so* well known for 
their stability and market share in the industry. ;-)

> When the repeater detects 32 consecutive collisions on one port it will
> logically turn off or segment

Cabletron is not the only one doing this. I've seen this in other vendor 
products as well, however everyone does it a bit differently. This is an 
add on feature, not something that's part of the normal repeating process.

You have two problems coming into play here. The first is that the only 
100% accurate way to detect a collision is to know the exact bit pattern 
that is being put onto the wire. The only systems with that knowledge 
are the systems actually involved with the collision (because they know 
the pattern they are currently transmitting). This means you have to 
have a full layer 2 implementation to do this right. Hubs are only layer 
1. So what vendors are trying to do here is implement layer 2 
functionality without actually implementing a full layer 2. This can be 
a really bad idea.

This "feature" can actually miss a number of legitimate collisions 
depending on the bit pattern (again, we are talking hubs not switches 
that can implement this feature just fine). It can also guess wrong. For 
example I've had to troubleshoot networks where this feature has 
incorrectly triggered on frames with heavy padding. Remove the data 
stream from the wire, and things return to normal. I've tracked that you 
can isolate the problem to the payload level rather than to a specific 
NIC or system.

The second is that collision storms can actually be quite common on a 
large network just after certain broadcasts. If you plot collisions over 
time, you see they peak and valley rather than remain at a constant 
level (assuming the problem is a lot of systems rather than over 
extending the Ethernet topology rules).

So segmenting off a portion of the network because a burst of collisions 
has been detected can actually leave functional systems isolated from 
the rest of the net. I think the intent is to isolate systems that have 
gone into an overrun or jabber condition, but you need to measure this 
over time versus raw quantity. Some vendors do this over time, some raw 
quantity, some only on excessive collisions (my personal favorite). 
Again it varies because this is a _feature_, not a part of the actual 
repeating process, and thus not available with every hub on the market.

Of course all this is kind of mute anyway as most of us have switched 
networks which can deal with collisions just nicely.

HTH,
C

_______________________________________________
gnhlug-discuss mailing list
gnhlug-discuss at mail.gnhlug.org
http://mail.gnhlug.org/mailman/listinfo/gnhlug-discuss