Registers are touched on basically every operation the CPU does :) But it didn't say "registers", but "registered memory". http://blogs.cisco.com/performance/registered-memory-rma-rdma-and-mpi-implementations/
Martin On 17 June 2013 08:36, Johan Hake <[email protected]> wrote: > On 06/16/2013 11:40 PM, Jan Blechta wrote: > > On Sun, 16 Jun 2013 22:40:43 +0200 > > Johan Hake <[email protected]> wrote: > >> There are still fork calls when swig version is checked. Would it be > >> ok to check it only when dolfin is imported? That would be an easy > >> fix. > > > > I've no idea. There are two aspects of the issue: > > > > 1. forks may not be supported. > > Following [1] below it looks like they should be supported by the more > recent openmpi and it also says that: > > In general, if your application calls system() or popen(), > it will likely be safe. > > > 2. even if forks are supported by given installation, it may not be > > secure. Citing from [1]: > > > > "If you use fork() in your application, you must not touch any > > registered memory before calling some form of exec() to launch > > another process. Doing so will cause an immediate seg fault / > > program crash." > > > > Is this condition met with present state and would be met after > > suggested change? > > I have no clue if we do touch any register before we call the fork and I > have no clue whether the suggested fix would do that. Aren't registers > touched on a low level basis quite often? > > Do you experience occasional segfaults? > > Also [2] suggest that the warning might be the problem. Have you tried > running with: > > mpirun --mca mpi_warn_on_fork 0 ... > > Johan > > > > > Jan > > > >> > >> Johan > >> On Jun 16, 2013 12:47 AM, "Jan Blechta" <[email protected]> > >> wrote: > >> > >>> What is the current status of a presence of fork() calls in FEniCS > >>> codebase? These calls are not friendly with openib infiniband > >>> clusters [1, 2]. > >>> > >>> Issue with popen() calls for searching swig library was disscused in > >>> the end of [3]. I'm still experiencing these sort of troubles when > >>> running on infiniband nodes (even when using only one node) so was > >>> cleaning of popen() finished or are there any other harmful fork() > >>> calls in FEniCS codebase? > >>> > >>> [1] http://www.open-mpi.org/faq/?category=openfabrics#ofa-fork > >>> [2] http://www.open-mpi.org/faq/?category=tuning#fork-warning > >>> [3] https://answers.launchpad.net/dolfin/+question/219270 > >>> > >>> Jan > >>> _______________________________________________ > >>> fenics mailing list > >>> [email protected] > >>> http://fenicsproject.org/mailman/listinfo/fenics > >>> > >> > > _______________________________________________ > fenics mailing list > [email protected] > http://fenicsproject.org/mailman/listinfo/fenics >
_______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
