On Wed, 4 Mar 2009, Carlo Nervi wrote: CN> Dear all,
dear carlo, CN> finally I succesfully compiled the code using ifort, mkl CN> and mpich2. CN> However, I'm getting odd results, very probably due to CN> wrong settings. CN> CN> I tried to run the example01 with the following settings: CN> CN> OMP_NUM_THREADS=1 CN> mpiexec -n 8 CN> CN> The example is executed, but with a very long time. All 8 CN> CPUs are running CN> between 5 and 20 % of the cpu time. For some reason we are CN> not using all the power.. please note, example01 is _very_ small. as a result there is little data to parallelize over and quite a bit of i/o. CN> I tried to run also "mpiexec -envall -n 8" (maybe the CN> OMP_NUM_THREADS is not passed) CN> but i got identical results. CN> Here is the results of the first example in the example01 CN> directory: CN> CN> results/si.scf.david.out CN> CN> Writing output data file silicon.save CN> CN> PWSCF : 0.50s CPU time, 7.27s wall time see. these are _seconds_. unless you run a test that runs _minutes_, there is little use to these numbers. as suggested before, a good test to see if your MPI is working well, are the cp.x inputs in example21. have a look at the readme first. running them with ./run_example 32 should be sufficient. CN> On the other hand, if I set CN> CN> OMP_NUM_THREADS=8 CN> and do not use mpiexec I got: CN> CN> Writing output data file silicon.save CN> CN> PWSCF : 7.10s CPU time, 1.12s wall time CN> CN> Therefore results/si.scf.david.out takes 1.12s instead of CN> 7.27s, a much CN> more reasonable timing. CN> Anyone could give me an advice? this is due to having to create 8x the number of files. see my comments in the (longish) e-mail from last night with benchmarks wrt. to the number of pools to use. it also looks briefly into the impact of file io. with this small a system, the impact is correspondingly larger. cheers, axel. CN> Thank you, CN> Carlo CN> CN> -- CN> ------------------------------------------------------ CN> Carlo Nervi carlo.nervi at unito.it Tel:+39 011 6707507/8 CN> Fax: +39 011 6707855 - Dipartimento di Chimica IFM CN> via P. Giuria 7, 10125 Torino, Italy CN> http://lem.ch.unito.it/ CN> CN> CN> _______________________________________________ CN> Pw_forum mailing list CN> Pw_forum at pwscf.org CN> http://www.democritos.it/mailman/listinfo/pw_forum CN> -- ======================================================================= 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.
