Regarding the benchmark: You can instrument a 2.x version of the
interpreter to build the data set and then have the test load this data
set in Py3k and have it replay the allocation/deallocation in the same
way it was done on the 2.x system.

I also expect that patch #2321 will have an effect on the performance
since that patch changed the way memory allocation of the buffer was
done (from using the system malloc() to using pymalloc's allocator for
the buffer as well).