One thing that I remember from a performance test we once did is, that the results
are a lot more realistic, better and more stable, if you try to decouple the startup of
the different clients a little bit, so they are not all in the same section of code at the same time.
We inserted random usleeps, I forgot what range, but 10 ms seem reasonable to me.
This was another database, but it might also apply here.
Andreas