Apart from hard disk throughput the usual culprit is interrupt performance. Most cheaper computers have interrupt response times that saturate the CPU at an interrupt rate that allows only a few MB/s on a 100 MBit/s network. Good computers, on the other hand, will achieve a total throughput of 10 to 11 MB/s.

Of course, most single hard disks will not yield that much, so you may just be measuring your hard disk performance. You want a RAID 5 array on the server to be able to feed a 100 MBit/s network at full speed.

To find out about your interrupt performance, fire up Performance Monitor and watch the CPU interrupt rate and time on both your workstation and your server. If it comes close to 100% on either the workstation or the server, you've maxed out your computers. This also explains why it is direction dependent, because the sending interrupt load may be different from the receiving interrupt load, and one computer may be better at this than the other.

The network adapter driver plays a crucial role also. Good network adapters will issue fewer interrupts than cheap ones. What the good ones do is chain several blocks after one interrupt. The Compaq Netflex-3 is an example for this, but some others also do this.

On a LAN usually only the server performance matters, because usually several workstations will be competing for the server at any one time. Thus you may be happy if you can get the server up to 10 MB/s or more, even if the workstations only manage one half or even one quarter of that throughput. Only if you have applications in which one workstation would benefit from utilizing the full network capacity, then workstation performance matters also. Some backup scenarios come to mind.

I cannot repeat one point often enough though: A good 100 MBit/s Ethernet will have a 10 to 11 MB/s throughput, if not even more. Don't let anybody tell you otherwise. There are some people around who have never seen this and have come to believe that it's impossible.