Briefly, Pavel discovered a situation in which replacing the Python memory allocator with TCMalloc resulted in really bad performance. The latest is that there appears to be a bug or gotcha in TCMalloc with glibc, where TCMalloc does a poor job in cases where mremap can be used by glibc. The TCMalloc folk are going to look into it more, I gather. (See google-perftools thread here.)
Anyway, this was a situation where we just threw the task at the students to see if anything interesting would pop out -- not expecting much of anything other than a learning experience for the student -- and yet through some simple-yet-dogged testing, Pavel really contributed something.
There have been several real success stories to GHOP. I need to write them down, sigh... my kingdom for some time :)