Roland has pushed his new Linux "ummunotify" kernel upstream (i.e.,
it's in his -next git branch):
http://git.kernel.org/?p=linux/kernel/git/roland/infiniband.git;a=commit;h=2fadea9acc19674c07ae7a9d90758f4b9b793940
It's not yet guaranteed that it will be accepted, but it looks good so
far. With some bug fixes from Pasha/Mellanox and Lenny+Mike/Voltaire,
I think it's ready for wide-spread testing (I mailed some of you
yesterday asking for specific testing). I'm asking all to give the
prototype code a whirl to shake out any remaining design bugs.
I describe the issue that we're fixing in my new MPI-themed blog:
http://blogs.cisco.com/ciscotalk/performance/comments/better_linux_memory_tracking
The HG where this OMPI work is being done is here:
http://bitbucket.org/jsquyres/ummunot/
You need to have a very recent Linux kernel (2.6.31+) and Roland's
umunotify module installed/running. Build the OMPI HG tree with the
"--enable-mca-no-build=memory-ptmalloc2" to disable ptmalloc2 and
enable the ummunotify stuff.
This hack-ish "disable ptmalloc2" step is only necessary while we're
shaking out the design issues. I'm halfway through merging the ummunot
+ptmalloc2 code into a new opal/mca/memory component named "linux".
This component will choose at run time whether to use ptmalloc2 or the
ummunotify stuff (i.e., the --enable-mca-no-build... step won't be
necessary when all is said and done; a default OMPI Linux build will
do the Right Things).
Thanks.
--
Jeff Squyres
jsquy...@cisco.com
