Archived from groups: comp.dcom.lans.ethernet (
More info?)
Rich Seifert <usenet@richseifert.com.invalid> said
>In article <cvtci1pats9jpsfk29l92r1ilm48qh2cbl@4ax.com>,
> DHP <me@privacy.net> wrote:
>
>> With Manchester encoding, even a stream of zeros is actually a square
>> wave.
>>
>
>It *could* be a square wave, but we intentionally slew-rate limited the
>signal impressed on the coaxial cable; it has a 25 ns nominal rise/fall
>time. This both reduces the effect of tap reflections and reduces the
>EMI generated by the signal.
Or, to be perverse, it limits the bandwidth
I agree about it being simpler in the time domain though, because we
are really talking about narrow pulses - the derivative of the edges.
>Let's see if I remember the numbers correctly:
>
>The nominal voltage resulting from a single transmitter on the coaxial
>cable is ~2V p-p. The current in the capacitive tap (which gets
>reflected into the coaxial cable) can be calculated as:
>
>I = C dv/dt
>
>dv/dt is 2V / 25 ns, or 80 MV/s (that's MegaVolts per second)
>C is 4 pf worst-case, so the current is 4 pf * 80 MV/s = 320 uA.
>
>The impedance seen by the capacitor is 25 ohms (it effectively sees two
>50 ohm cables in parallel, one in each direction away from the tap).
>Thus the "noise" voltage generated by a single tap is 25 * 320 uA = 8 mV
>
>8 milliVolts by itself would not be a problem, however, if we had the
>worst-case situation of all 100 transceivers lumped together, we would
>have 800 mV of signal, which would blow away our required 5:1 signal to
>noise ratio. So the idea is to make sure that as few of these 8 mV
>spikes (they only last for 25 nS, while the voltage on the cable is in
>transition) add up in phase. Also, by creating a minimum cable length of
>250 m for those 100 taps (i.e., by spacing them by at least 2.5 m), we
>are sure to get a fair amount of attenuation, at least for the taps that
>are farthest away.
>
>The simulations simply tried a zillion variations on numbers of
>transceivers, placement along the cable, spacing requirements, and data
>patterns to determine if there were any pathological situations where
>the noise exceeded the budget allowance.
Interesting, because with 2.5m taps, a 25ns pulse produces
(practically) non-overlapping reflections. So if you're trying to
break the system you don't gain anything by clustering them. That
means the amplitude is that of a single reflection per half-bit of
line, 11.7m. So the worst possible case is about 42 reflections or
328mV. What's that, about -10dB in a 1V system? But as the noise isn't
random it can *never* go any higher so, ignoring other noise, it
should be possible to set threshol voltages that avoid errors
altogether. But only just.
> By the way, this took a HUGE
>amount of computer power, at least by the standards of the time. I
>managed to distribute the simulation runs across dozens of VAXen (780s,
>the only ones in existence at the time) all around the world, using
>DEC's private network. I used the idle compute power of just about every
>machine in those time zones where the normal work day was over. It was a
>rather ambitious task for its day.