Dear all, Many thanks for your suggestions, but there is one fundamental obstacle: there is no mpiexec available at all, everything must go through a scheduler.
I will see into what Satish proposed, i.e. to first get hold of known nodes. But I am afraid this may fail, because these run some very limited system. Regards, Dominik On Mon, Dec 12, 2011 at 2:21 PM, Satish Balay <balay at mcs.anl.gov> wrote: > On Mon, 12 Dec 2011, Matthew Knepley wrote: > >> On Mon, Dec 12, 2011 at 7:11 AM, Dominik Szczerba <dominik at >> itis.ethz.ch>wrote: >> >> > Thanks for your answers. Meanwhile I clarified the situation a bit: >> > >> > I can bring xterm up manually from the command line, but the job is >> > run using a scheduler (slurm). It then gets executed on arbitrary >> > nodes (some stripped down linux) which apparently can not make X11 >> > connections. >> > >> >> Sometimes you can set the env on the compute nodes, and get DISPLAY right. > > > With compute nodes -its not easy. Even if they have X11 installed - you > might have to do multiple things: > > 1. allocate nodes > 2. create ssh-x11 tunnels to each node thats allocated [and hope its the same > localhost:10 value] > 3. Now start up the parallel job with this display > > mpiexec -n 4 ./ex2 -start_in_debugger -display localhost:10 > > Satish >
