The issue of the generational hypothesis is interesting. Of course if a benchmark doesn't exhibit "generational" behaviour then it won't be a good test of a gen

The issue of the generational hypothesis is interesting. Of course if a benchmark doesn't exhibit "generational" behaviour then it won't be a good test of a generational collector. But taken too far, that creates a bias: you are selecting benchmarks that fit what you know a generational collector is good at. The question then is: are real world programs as generational as we think? This might depend on whether we assume are are using a generational collector. If you assume shortlived object allocation is ~free then you will produce software conforming to the generational hypothesis but that is not good evidence that the generational hypothesis is true. It is a bit like saying "caches are important because of locality of reference" and exhibiting as evidence code optimised to make good use of caches. We have to be careful that we are not just measuring things designed around themselves.