There is another factor at play here that I haven't seen mentioned yet, and that is the inherent characteristic of Ethernet networks to experience rapidly degraded performance/throughput once they get past about 30% busy. The reason for this is inherent in the basic architecture of Ethernet - on a single Ethernet cable (like Cat5) only one device can be transmitting at a time.
Before a device transmits it tests the cable to see if it is busy transmitting someone else's signal. If it (the cable) is busy, the Ethernet device (actually it's the network card in the device) waits a random amount of time and then tries to transmit again. This process continues until the the device finds that the cable is free and it then transmits the current data packet. Then the process starts all over again.
This architecture is called "indeterminant" because it is impossible to predict how long it will take any specific amount of data to get transmitted - again becuase there is no telling how busy the cable will be or how many times a particular Ethernet adapter will have to wait/retry it's transmissions. In general, if the cable is more than abotu 30% busy there are a lot of waits/retries that happen, and this significantly degrades overall network performance.
Interestingly enough, there is (or was) an alternate network architecture that does not have this problem. It is called a "deterministic" network becuase it is always possible to predict exactly how long it will take to transmit a known amount of data. This kind of network was invented by IBM and is called Token RIng. It has several major advantages over Ethernet, but was never widely adopted because IBM wanted to charge money for every Token RIng device, whereas Ethernet was given away for free.