Where I at, the benchmark testing definition is to have a set of data and input rules that provide a known result, then measure other data and input rules against that result.
One example might be 100 records take 3 minutes to process every time, so this is the benchmark. You would measure your expected results against that.
You might then expect that 200 records takes 6 minutes...
Very simplistic, but the idea is to have a starting point to measure against.
Although it seems that getting involved in any definitions is starting to become dangerous, the truth is that a benchmark is any arbitrary goal that you would like to match or improve upon. It literally is derived from a workbench where your goal is to make spindled the same size by marking a line that reflects a test spindle, so you make it easy on yourself to cut the others to the same length. In comparing computers many agencies use a typical configuration as 100% performance, and compare other configurations as better or worse, just to get a relative feedback. If you deal with a formal benchmark you define a standard that you measure performance against. Jean's example fits.