Hi Masoud,
In older materials, collisions were categorized as either local, remote, or late, and this categorization is actually helpful to explain how the collisions were detected. This is probably going to be more extensive but you may find it interesting.
Local collisions referred to the type of collisions that actually occured on the same medium that interconnected two or more senders. The characteristic symptom of these collisions was the violation of link coding - invalid voltage levels, invalid signal state transitions, their frequency and ordering, etc. that was the result of two or more independent signal transmissions being superimposed onto each other. Detecting a collision could theoretically be based on simply detecting whenever these rules of link coding were violated. The circuitry on a NIC should always know it because it knows whether it was able to follow the signal and decode it bit by bit, or whether it lost its track. Alternatively, with more advanced digital signal processing technology, you could try listening to the media at the same time you are sending, and subtract your transmitted signal from the signal present on the medium. This difference should always be zero (or equal to a background noise). If, however, someone else is transmitting as well, this difference will be nonzero, and is therefore an indication of a collision.
Local collisions as described above were typical on coaxial Ethernets. However, on TP cabling where Tx and Rx pairs are independent, local collision is not detected by violated link coding (because these transmissions can never be superimposed onto each other), but rather simply by sensing an incoming transmission on the Rx pair whenever another transmission was being sent over the Tx pair. This assumes, of course, that half duplex is in use. If full duplex was used, seeing transmissions both on Rx and Tx pairs would be entirely normal.
If two stations connected to a hub via a TP cable start transmitting data simultaneously, this will be registered as a local collision inside the hub (a hub has yet another means of detecting a local collision apart from the means mentioned above - simply by seeing that more than one of its ports is receiving a transmission in the same moment). The hub only needs to transmit whatever it could read from the stations back to these stations over their Rx pairs to make them see that while they're transmitting, someone else is transmitting, too. This will be registered as a collision by the stations attached to the hub.
Note that with a hub, the collision does not occur in the cable itself, as opposed to coax cabling. Receiving and sending signals over the same TP cable in the same time does not alone result into collision. Rather, what is considered a collision is the fact that during half duplex operation, a station started receiving a transmission on the Rx pair while it was sending data on the Tx pair.
I have mentioned remote collisions. Remote collisions are local collisions as seen from another port on the same repeater (or a hub, for that matter). In other words, if two stations on a common coaxial cable segment jump into each other's transmission, that is a local collision with all its link coding violations etc. However, if this coax cable is connected to a repeater, this repeater will attempt to read and make sense of the incoming mess - and while it won't be intelligible, it will nonetheless attempt to interpret it bit by bit and regenerate these bits on its other port. On the other port of this repeater, we won't see any link coding violations because a repeater (or a hub) does not simply amplify signals, rather, it reinterprets the incoming signal as bits and then sends these bits anew from scratch. We will only see the frames that are shorter than the minimal length, that are misaligned (they do not end on byte boundaries), and obviously, whose checksum is invalid. These frames, sometimes also called collision fragments, do not register as collisions on the receiving NICs, only as invalid frames - too short, misaligned, wrong CRC. On hubs, remote collision is what all other listening (not transmitting) stations will see while some senders end up in colliding with each other.
Lastly, a late collision is a collision, local or remote, that occurs after the first 64 bytes of a frame were already sent. This collision is an indication of a faulty network because if the network was built according to standards, this would never have happened. Late collisions were usually caused by excessively long cabling and/or chaining too many repeaters or hubs together. They are notoriously difficult to detect because some NICs do not check for collision after the first 64 bytes of a frame have been sent. Thus, a frame can end up destroyed in a collision, while the NIC thinks the frame has been sent properly.
I think there is mixed concept of hub and switch in the market when third parties sell their products.
I hope not :) My sincere belief is that you cannot even purchase a hub nowadays. Whatever device to build an Ethernet LAN you get, it is a switch. Do you have a different experience?
Best regards,
Peter
