The session dir is also used by the shared memory system for its backing file, so you may need it if you plan to run more than one proc in a VM. This has been one of the sticking points for VM/container-based operations.
As for the orted: your description is pretty close. The socket you mention no longer exists - instead, there is one used by the PMIx client-server. It’s a Unix-domain socket, and the rendezvous file is in the session dir tree. You might look in the orte/orted/pmix directory to see what the daemon is doing to support the individual client. Specifically, the daemon needs to pass down the location of every proc, so you’d need to figure out what that means when you are using container and/or VMs. > On Oct 22, 2015, at 11:45 AM, Justin Cinkelj <justin.cink...@xlab.si> wrote: > > Normally, mpi_run starts via ssh on remote node orted process, and orted > start mpi_program via fork+exec. > orted and mpi_program communicate via: > - environment variables (ok, that's on-time setup only, but still) > - pipes (only one, right? - it is close-on-exec by child). > - file descriptors, mpi_program stdin/out/err are redirected in orted > between fork and exec. > - socket (only one?), mpi_program connects to OMPI_MCA_orte_local_daemon_uri > - session dir (OMPI_FILE_LOCATION) > (did I miss anything - or, how much?) > > How is session dir used? I saw check for aborted file (so that orted can > figure out if child died?). Is there any other use of that dir? > > Can I just ignore it, if I try to run orted on host, and mpi_program in > container/virtual machine? > > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2015/10/18245.php
