Hi everyone!
We are trying to solve an annoying problem, using the Intel Fortran 8.1
compiler, OpenPBS, and LAM 7.1.1, running in bash shells under Fedora
Core 2. As every user of Intel 8.1 compilers know, they use vaste
amounts of stack space. In order to alleviate this, one has to increase
the allowed stacksize by the stmt: ulimit -s unlimited.
Now, the problem is, that this limit or any other related limit isn't an
environment variable. When the bash shells on the execution nodes are
started by lamboot, they do not inherit any limits set in the submit
shell.
The only (not very elegant) workaround, we have come up with, so far, is
to execute a script (instead of just the program we want to run) in the
mpirun command. The script first performs the ulimit command, whereupon
it runs our program (A jobscript is shown below).
Has anyone a better (more elegant, and preferrably systemwide) way to go
around this nuisance?
Best regards, Jess Michelsen
#!/bin/sh
#PBS -N MPItest
#PBS -r n
#PBS -q workq
#PBS -l nodes=2:ppn=1
#PBS -j oe
NPROCS=`wc -l < $PBS_NODEFILE`
cat $PBS_NODEFILE
lamboot -v -b -ssi boot tm $PBS_NODEFILE
mpirun -O -ssi rpi tcp -np $NPROCS -wd $PBS_O_WORKDIR runscript
lamhalt
|
