Hello Open MPI developers,
I've been away from the Open MPI code for a long time now, but I just
ran across this article that should give developers second thoughts
about using mmap'ed files in performance critical situations.
"Deferring mtime and ctime updates"
I might suggest an audit of the Open MPI code base for mmap calls
to make sure they are all needed, or if they can be done
a different way, especially in the coll and sm btl.
Yes, the article is about a patchset to lessen the performance problem,
but until that gets into the kernel, and into the distributions, and into
the parallel machines, this could be a source of lost performance in Open MPI.
I seem to recall some work was done after I left development to
make use of POSIX or SYSV shared memory system calls that avoided
having an associated backing-store file. But I don't remember if
those became the default, or if mmap'ed files was still the default.
Anyway, back to lurking on the mailing lists for me.
P.S. - Darn, my timattox_at_[hidden] e-mail alias is gone... c'est la vie!
Tim Mattox, Ph.D. - tmattox_at_[hidden]