On Wed, 4 Mar 2009, Lex Kemper wrote: LK> Hey Axel, LK> LK> In my case, exporting OMP_NUM_THREADS from my .bashrc was sufficient. LK> I guess that's not the case everywhere.
right. this happens mostly on batch systems when launching jobs through libtorque (very convenient, but you don't get a login shell) and some ssh setups also do not read .profile. also one has to consider that debian based systems (incl ubuntu), don't use /bin/bash as /bin/sh, but have a version of /bin/ash as /bin/sh and that one only reads .profile (which many bash setups in turn don't) but, of course, not .bashrc. so it is actually a good idea to verify this, like JR did. the case of a not exported environment is nasty, because it will slow down the calculation (if you run 8 mpi tasks on a 2x quad-core node, mkl with spawn 8 threads for each of them), without the telltale sign on the rank0 node of cpu time being multiple times larger than the walltime. cheers, axel. LK> Thanks! LK> LK> Lex LK> -- ======================================================================= Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu http://www.cmm.upenn.edu Center for Molecular Modeling -- University of Pennsylvania Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323 tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425 ======================================================================= If you make something idiot-proof, the universe creates a better idiot.
