Closed jcphill closed 8 years ago
Original date: 2015-12-03 23:14:50
Added environment variables in https://charm.cs.illinois.edu/gerrit/#/c/935/ https://github.com/UIUC-PPL/charm/commit/a830a94f69ff6712daffd16e11fd3f41ac27b839
Original date: 2015-12-04 03:46:04
Jim, could you please open a separate issue for the suggested ++mpiexec-no-n
? I think that may require a bit more discussion. At any rate, it's a distinct change from this one.
Original date: 2016-05-12 19:28:25
https://charm.cs.illinois.edu/gerrit/#/c/942/ https://github.com/UIUC-PPL/charm/commit/3b59f9fa8306dd103b75a0403cbd20d5ff741f9d
Original issue: https://charm.cs.illinois.edu/redmine/issues/906
IBM's poe launcher uses MP_CHILD and MP_PROCS to specify the rank and size. It can currently be used with ++mpiexec via the following ++remote-shell script:
!/bin/csh -f
shift shift sed -e s/PMI_ID/MP_CHILD/ -e s/PMI_SIZE/MP_PROCS/ -i.orig $1 exec poe $*
MP_CHILD and MP_PROCS should be added to the end of the set of environment variables that the generated charmrun script checks.
It would also be good to have a charmrun command line option to not pass -n to mpiexec, since most sites seem to use an mpiexec substitute that gets the process count directly from the environment. With these changes NAMD could be launched as (e.g.):
charmrun +pXXX ++mpiexec-no-n ++remote-shell poe namd2 ...
or (on Stampede):
charmrun +pXXX ++mpiexec-no-n ++remote-shell ibrun namd2 ...