Re: Company network slowdown

In the Usenet newsgroup alt.internet.wireless, in article
<87aji15brj5flcejtplo4b6hibidfv41fd@xxxxxxx>, Jeff Liebermann wrote:

>My guess(tm) is that I was seeing the intentional trashing part of the
>ethernet collision detection mechanism going continuously. If the PAD
>(packet assembler/disassembler) detects a partial collision, it is
>suppose to intentionally trash the packet to prevent propogating
>garbage. The waveforms I saw looked like small pulses of this
>intentional trashing algorithm.

"The Specification" says this is supposed to be a 32 to 48 bit time
jam - content not specified except that it must not make a valid CRC
if by some bizarre reason the collision is detected that late. The
detection to jam delay is supposed to be no more than two bit periods.

>The packet disassembler is programmed to detect this trashing and
>temporarily increase the collision backoff timers resulting in fairly
>long delays where nobody is transmitting.

Collision backoff should only effect the two (or more) systems who
collided. Others on the wire shouldn't be effected, and they can
slip in 9.6 usec after the wires stop vibrating.

>Ah, nostalgia. Much my HF ham radio station is thrown together with
>bright yellow coax cable. I always knew it was good for something.

It's a good HF cable, but isn't rated for voltage (given that the
normal signals are about 2 Vpp), and of course has the different
jacket, solid center conductor, and double shielding.

>I also have a few 3C500 cards and transceivers which I'm sure will end
>up in a museum somewhere. About half the problems I found with the
>stuff were mechanical. Poor probe contact due to loose clamping and
>lousy terminations due to flakey N-connector assembly were most
>common.

We also had problems with the people not cleaning the outer braid out
of the tap holes. As far as the N connectors, there were only two per
run, and we tended to take special care when putting them together
(though not the kind of care you used with RG-9 or 214 at 3+ GHz.).

>Also water incursion via unused probe holes and the usual mouse chewing.

Didn't have much of that, but then we left the transceivers in place once
they got installed.

>I built my own TDR (time domain relfectometer) which was a big help
>because I could "see" the probes and terminations.

The rise/fall spec was 20-30 ms, though the spec also required the
harmonics down (second/third down 20 dB, fourth/fifth down 30 dB,
sixth/seventh down 40 dB, higher down 50 dB) which really limits
the resolution of TDR. Funny thing is, no-where in the spec do they
give a required VSWR/return loss, whatever. I guess it's implied
with the transceiver loading.

>Early hardware PAD chips just discarded malformed packets and didn't
>report them. At best, it lumps everything into "discards" without any
>detail. That was the problem with using PC's to do the
>troubleshooting. The runts never showed up on any of the diagnostics.

Even late model hardware didn't report it. That's why we needed the
NetGen sniffer. (Well, that was one reason.)

>Everyone knew how to load paper and were told to keep the printers full.
>The reason was not very subtle. This company had previous experienced
>a computer meltdown and would print out every customer order so as not
>to lose anything. Between the dozen or so printers, that was about 2-3
>reams per day plus a huge pile at the end of each accounting period.

Wowser! We just had a pretty good tape backup (yeah, I know) program.
We knew the tapes were good, because they were verified during the day
following the backup, and we had enough user mistooks that we were
restoring files on a regular basis. There was off-site storage as well.

>If you search Google, you'll find a few articles by me on using RG-6/u
>75 ohm coax for 10base2 (Cheapernet). Works fine if you don't have
>any BNC T connector taps and only transceivers at the ends of the
>coax.

Huh? What kind of transceivers had built in terminations?

>I learned networking the hard way with Arcnet (over both coax
>and flat telco wire) and Starlan with CAT3 and 25 pair telco bundles.
>I still have a few 10baseT runs over 25 pair telco cables.

<shudder>

>So, I would show up at about 7PM putter around until about 9PM, and
>then bring everything down. As I recall, I was done with the rewiring
>exercise at about 2AM and spent the next 3 hours fixing my crimping
>and punch down mistakes. I now have various cable testers, but at the
>time, it was with an ohms-guesser and clip leads.

We had very little twisted pair, and once the orange cable was in (and
working), the rest of it was stringing four pair, and poking pins into
DB15s, which we often fobbed off on the interns.

>Somehow, I thought that I would remember where everything went and
>label things later. Later never arrived.

Naturally. We tried to pre-mark the drop cables so that there was both
a plastic marker, and a felt tip marking at the transceiver end. The
drop end had a felt tip marking and a Dymo label that went on the
face plate. Early on, we also marked the transceivers with felt tips on
masking tape, but the heat killed that. When we started installing
fiber, this was outsourced, but one of the contract requirements was
that the fibers had to be marked with serial numbers every ten feet
and within a foot of each end, and the serial numbers had to be entered
into "the book" with full location information not later than noon of
the next day. We randomly tested this as part of the acceptance
inspection. This seems to have worked, though I have no idea how much
it added to the tab.

>The segment length can be up to 500 meters for 10base5 so that's fine.
>Howeve, you're only suppose to have 100 nodes per segment, so you're
>more than slightly out of spec. I'm suprised it worked at all.

The spec was ALSO 2.5 meters between transceivers - we violated that
to heck and gone. ;-)

>Even repeaters wouldn't have made that conform as there were only
>suppose to be 3 segments max per system.

Not so. There could be a maximum of two repeaters (or four half
repeaters or what ever) between any two "stations". If you wanted
to have a 250 meter cable with repeaters every 2.5 meters feeding
100 OTHER segments, that was OK. Well, it was permitted - not
that having 100^2 hosts in one collision domain would work worth
a darn. ;-)

Old guy
.

Relevant Pages

  • Re: Why is there a minimum spacing?
    ... it creates an out-of-phase reflection of a ... one would want the transceivers to be *randomly* ... cables with randomly-distributed markings for this purpose! ... The idea is not just a *minimum* 2.5 m spacing; ...
    (comp.dcom.lans.ethernet)
  • Re: Why is there a minimum spacing?
    ... A 500m cable could have 43 taps with that spacing, ... To all other devices on the cable, this reflection appears as asynchronous "noise," i.e., a signal that interferes with the desired signal. ... Ideally, one would want the transceivers to be *randomly* spaced along the cable; this would ensure that the reflections added not algebraically, but on a root-mean-squared basis, yielding much less reflected energy. ... It was relatively easy to mark the cables with a uniform 2.5 m marking; as the cable comes flying out of the extruder, it passes across a roller with a 2.5 m circumference, which places a mark at every rotation. ...
    (comp.dcom.lans.ethernet)
  • Re: [patch libata-dev-2.6 1/1] libata: sync SMART ioctls with ATA pass thru spec (T10/04-262r7)
    ... > hate to think there would be a collision like that. ... Previous versions of the spec did not specify a value. ... John W. Linville ... send the line "unsubscribe linux-kernel" in ...
    (Linux-Kernel)