Thanks for catching this. I agree with you, the first option seems
like it is closer to what we want to have happen in Open MPI. I'm
less concerned with dropping a couple of pages of memory as I am with
searching our registration cache more often. It shall be committed
to SVN today.
B
Hello,
I found one more problem with ptmalloc and registration cache.
In arena.c:grow_heap() when heap is shrinking ptmalloc tries to be smart
and is using mmap() to change pages protection instead of mprotect() because
as a side effect mmap() drops underlying pages. In the case the area is
regis