Benchmark different kind of SQL database :What's the best test to do?

The best test, and the only one that is at all meaningful, is to use it exactly as you would in a production environment. Testing it any other way may give completely different and irrelevant results, so write your test script accordingly.

I agree with John Gardeniers - but in order to actually measure the performance effectively you need to look at the transaction times for a simulated workload at the HTTP level - note that the "standard" log formats only provide integer response times - so you'll need to do some tweaking. Also you need to test out how the system behaves with varying levels of load (running single operations on isolation does not reflect what happens when you've got lots of other stuff happenning - speed <> scalability).

C.