HowTo: Benchmark TCP

LeWiz Communications - The Wizard of Internet Communications - Press Releases

This was the best TCP benchmark I've done, and made a good comparison of TCP offload engines at the time.

A lot of TCP, or web server benchmarking gets done. Most of the results are wrong. A lot of web server traffic is measured, across an internet that never maxes out the web server. The web server is doing many other things, so you never know if it can actually run the network at full speed - because it never does.

If it's a 1G network, can your PC or whatever actually put out 1G of data? Most benchmarks don't really go there. Most setups can't measure it.

Get two routers, could be home quality as long as its' 1G.

While you could do this with a switch, you don't want your benchmark machines running other TCP things. There's not much to do, so the router should not interfere with the speed. We are trying to measure the network at maximum speed.

Make cheap transmit and receive subnets.

Measure something (target) that has two ethernet ports...

Put a server for transmits on one subnet, a server for recieving on the other subnet.

Make sure the servers with thier OS/drivers etc can run 1G...Otherwise the obvious bottleneck will occur.

Put the target on both subnets. Start with running something like ttcp with the target recieving data from the transmit server, and sending to the receive server which just throws away the data (no follow on traffic).

With this base, you are set to run some more real world benchmarks and applications to get some good numbers.