Copied from Experts Exchange;
first let's look at a comparison of bytes to bits:
Bytes/sec. bits/sec.
12.5 MBps 100 Mbps ethernet
so theoretically a 100Mb file would take 100 / 12.5 = 8 seconds
If we assume TCP, protocol overhead will consume some of that, so say 11.0 MBps is available for data with normal packet sizes
100 / 11 ~ 9.1 seconds
This is assuming that both machines have disk subsystems and ethernet adaptors and drivers that can send and accept the data at that rate. A busy cpu, badly fragmented disk, irq sharing and driver implementation can all affect performance. Operating systems come into play as well as tasks are prioritized so that something like a file transfer won't slow all other tasks to a crawl. A netware server will generally see file much higher server performance than windows98 on the same hardware because it's written from the ground up to do that and little else.
Now the intermediate networking equipment comes into play. In a full duplex crossover cable scenarios one should have no collisions and minimal latency. Using a (good) hub will add some latency and drop you down to half duplex so it will also introduce some collisions. Bandwidth drops as latency goes up. (see http://www.babinszki.com/Networking/Max-Ethernet-and-TCP-Throughput.html) The use of a switch between two nodes can actually be slower than a hub as the hub has only to repeat the packet whereas the switch has to examine the packet and it's span tree to find the destination, though a switch starts becoming an advantage very quickly when more machines are sharing the wire.
So you have to test your particular setup to see what bandwidth you're getting and work on each bottleneck individually, optimize each machine and operating system for file sharing.
Another example of factors to consider (an apple site, but same principles).
http://www.atpm.com/network/problems/network_speed.htm