On Thu, 5 Mar 2009, Carlo Nervi wrote: CN> time is almost 1/3 of the wall time. I suspect that Axel CN> is right pointing to the disk I/O. I'm afraid that to CN> increase the performance I should set the RAID 10...
no. this cannot be I/O. this is why i suggested this example. unlike with pw.x in some configurations, there usually is no significant i/o when running cp.x. also you have almost exactly 4x the wall time relative to the cpu time. are you _sure_ you are running on a machine with 8 cores? _and_ is that machine otherwise empty? when running on a local 2x quad-core intel xeon E5430 2.66GHz node, i get 42.30s CPU time, 43,87s wall time. with serial MKL and 8 MPI tasks. CN> OMP=2, MPI=8 CN> v0: CP : 0m46.98s CPU time, 1m36.08s CN> wall time CN> little worse than OMP=1 and MPI=8 there is no point in oversubscribing a node. you rather want to test, whether less than the maximum number of total tasks per node is not giving you a better speedup. cheers, axel. [...] -- ======================================================================= 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.