> Note this is really a worst-case benchmark: lots of classes, no methods, no user data beside the classes.

Since benchgcclasses.py doesn't creates dunder methods,
cache doesn't have GC-tracked tuples, and ref cycles.

benchgcclasses2.py adds three dunder methods:

$ ./python benchgcclasses.py
GC time: 13.0 ms
gc: collecting generation 2...
gc: objects in each generation: 1 0 94942
gc: objects in permanent generation: 0
gc: done, 0.0131s elapsed
RSS:
USER       PID %CPU %MEM    VSZ   RSS TTY      STAT START   TIME COMMAND
inada-n  29604  0.0  0.1  47052 30692 pts/2    S+   20:42   0:00 ./python benchgcclasses.py

$ ./python-patched benchgcclasses.py
GC time: 17.2 ms
gc: collecting generation 2...
gc: objects in each generation: 1 0 135220
gc: objects in permanent generation: 0
gc: done, 0.0173s elapsed
RSS:
USER       PID %CPU %MEM    VSZ   RSS TTY      STAT START   TIME COMMAND
inada-n  29626  0.0  0.2  49880 33464 pts/2    S+   20:43   0:00 ./python-patched benchgcclasses.py