So we should see which of the parameters are really useful and which ones are neutral or dangerous... By now tcp_timestamps and page-cluster are considered guilty.
Can you suggest some kind of benchmark to evaluate if pulling them out of the script can make a difference.

Or, what about a double-blind cross-over scheme for an in-vivo experiment? Where a number of volunteers load in blind 3 or 4 different parameter schemes, in a latin square order, and evaluates the effect on an analog visual scale, and then we get the data and make the stats?