One pair is used to transmit, the other - to receive.
If you connect two computers without a switch or a hub, in the past you had to use a cross-over cable, where one pair is connected to pins 1,2 on one end and pins 3,6 on the other end. Hubs had reversed pinout of the socket, so you can use a straight cable to connect a PC to a hub, but would need a cross-over cable to connect two hubs (unless one hub had an "uplink" port).
Modern Ethernet devices can be connected with any cable - they will figure out whether the cable is straight or crossed and will reconfigure themselves accordingly. Gigabit Ethernet works a bit differently - it uses all pairs (instead of just two) and can reconfigure each pair as "transmit" or "receive" as needed.
Now, as to why pairs are used instead of single wires:
To transfer data, you need to be able to get some current to the receiving device. As we know current only flows when there is a closed circuit, so you need at least two wires connecting the devices. Your "scheme 2" will not work as you drew the "batteries" not connected.
This can be done in one of two ways - easier is to have one or more data wires and one ground wire (called Single Ended system). Here ground is shared among all the signals and you need less wires. However, this system does not work well for long distances - noise can get in the cable quite easily and the receiving device may not be able to understand the transmission. One solution is to use a coaxial cable (it shields the data wire from noise), but they are expensive and you would need one cable for each data pin. Still, multiple coax cables are used, say, for connecting a VGA monitor to the computer (at least in the better monitor cables). It is also true for analog audio.
A better way to do things is to have two wires for each signal. Now you send the signal in both wires, but invert one of them, that is, if you send "1" in one wire, you send "0" on the other - so the voltage between those two wires is always non-zero. You also use a twisted pair cable. This is called differential signaling. Now, the noise affects both wires in a pair equally and the receiver can cancel it out (by measuring the voltage between the wires instead of each wire to ground). This allows the signal to be sent further using cheaper twisted pair cables. Professional analog audio also uses differential signalling for, say, microphones etc (the XLR connectors have three pins - positive signal, negative signal and ground), so that longer cables can be used without noise affecting the signal.
An example of differential signalling:
As you see, in this case what matters is the polarity of the received voltage, so if whatever noise affects both wires the same, the polarity will not change and the information will still be transmitted.
To transmit in both directions (but not at the same time, so-called "half-duplex") over the same pair of wires you can do it like this:
Now when any switch is closed, both lamps light up, so any end can transmit taking turns. This arrangement is called "open collector".
The answers to the previous question you referenced, give some strategies but do not ground and/or not light the conditions in those that strategies are useful. Therefore, the short answer to your question: it depends.
The long answer is the following:
Scenario A: you trace a two-layer PCB
On a two-layer PCB, it is too difficult (or simply impossible) to route impedance matched pairs, therefore place the phy, mag, and jack as close as possible to make that traces short as possible but also keeping wires lengths matched within pairs at least.
Scenario B: you trace a four or more layer PCB
On an at least four-layer PCB, it is simple to organize the corresponding reference planes and there are two sub-scenarios here:
1) If you (can) organize AGND reference plane only, than only phy-mag pairs can be traced impedance matched, therefore you must keep the mag-rj45 distance as short as possible. (Keeping the lengths matched is also mandatory here.)
2) If you (can) organize both AGND between phy-mag traces and (let's call it so) MGND between mag-rj45 traces, than you can trace all the pairs impedance&lengths matched. But you must be aware that each mag-rj45 path must have its separate reference plane, rather than AGND that can be shared.
Some tips on how to do MGND is shown below.
Now, on your sub-questions:
Which of the two, remote the connector or remote the PHY, is “correct”/ easier to implement for maintaining signal integrity and minimizing EMI?
IMO, Scenario B1 is preferred, because tracing many pairs (including many pairs of many phys) with respect to one reference is simpler than what is needed to do in other cases.
What is the maximum length that is possible to remote the PHY or the connector?
Without a reference, up to one inch limit is recommended. With a reference, it can be much longer.
Do I have to run the long lengths off the board?
It depends on your construction. Running on a PCB, use at least the approaches shown above.
How does a 48 or 96 port gigabit switch run their signals while maintaining signal integrity?
They use many approaches, mostly including (but not limited to) shown above.
Are there any definite specs on how to proceed?
Maybe, but i think they all are case dependent.
Good luck.
Best Answer
The method is called echo cancellation, and it requires a bit of signal processing. Basically, the idea is since you know what you're sending out, then you can separate the signal you just sent from what is coming in from the far end of the link. The way the circuitry is set up, the transmit and receive signals are superimposed on top of each other, more or less adding together.
Simple example to give you an idea of how this works: if the transmitter sends
+1, +1, -1, +1
and the local receiver gets
+2, 0, -2, +2
then you can work out that the signal from the other end must have been
+1, -1, -1, +1
That's more or less the gist of how it works, but it's significantly more complicated due to delays and reflections. The technique is called 'echo cancellation' because sending just a lone +1 down the line will not result in receiving a lone +1, rather you will get several delayed copies at various amplitudes. For example, if you send
+1, 0, 0, 0, 0, 0
you might get back
0, +0.8, 0, +0.2, -0.1, +0.1
due to discontinuities along the line. The received signal then becomes the 'convolution' of the transmitted signal with this pattern. For example, if you send
+1, +1, -1, +1, 0, 0, 0, 0
then you will get something like
0, +0.8, +0.8, -0.6, +0.9, -0.2, +0.4, -0.2, +0.1
The transceivers send training sequences to figure out what the echo looks like (e.g. send a lone +1 while the other end is sending 0 and measure what you get at the receiver). This information is used to reconstruct what the receiver would expect to see from the transmitted data echoing back. This reconstruction is subtracted from the received data, leaving behind the signal from the other end of the link.
This method cannot tolerate as much loss or noise as using separate signalling pairs for each direction, however it means that you can re-use the old 100 Mbit cabling that you already have routed to every room in your building.
Incidentally, 10 Mbit and 100 Mbit signalling is horribly inefficient: both use a single receive pair and a single transmit pair, even though the cable has four pairs. When gigabit ethernet was developed, the designers wanted to keep compatibility with 10 and 100 Mbit ethernet as much as possible. Since there was no way they were going to get 10x the bandwidth out of one single pair, the solution was to improve the single pair bandwidth by 2.5x and then use all four pairs. They now have 10G ethernet over a slightly improved version of the same cabling (mainly it requires a lot of shielding), but it is currently very uncommon (most 10G ethernet uses completely different cabling that has one pair in each direction running at 10G). I seriously doubt we will see anything faster than 10G ethernet over RJ-45 cabling.