On Mon, Aug 6, 2012 at 12:06 PM, Cristian Degli Esposti Boschi <degliesposti at bo.imm.cnr.it> wrote: > Dear all, > > I have a (maybe basic) question on the performance of Quantum Espresso > with respect to openmp, even if I do not realise if the answer lies in > the way QE handles different levels of parallelism or lies at the system > level instead. > > However, I am running QE 4.3.2 on a quad-core machine with 2 CPUs 64bit. > SO is Debian stable, and it recognises 16 threads in the CPU table. > > I have compiled QE with pure openmp support (gcc) without any MPI > environment. > > The question is: even if I do NOT set the OMP_NUM_THREADS variable > at runtime, from the table of the "top" command my process > now uses 1000 % CPU time, as if it were "parallelised" over 10 threads. > > Is this real? Is there some kind of default when pw.x is invoked? > If so, how is openmp parallelism exploited by QE?
this has nothing to do with QE but is a typical implementation default. if OMP_NUM_THREADS is not set, most OpenMP runtime environment automatically default to using all CPU cores for threads. the fact that you see about 1000% and not 1600% is probably due to the limited threading during the course of execution for your specific input. in any case, you are more likely to get better performance with hyperthreading disabled and/or OMP_NUM_THREADS set to 8. BTW: this default behavior of OpenMP is a major nuisance for codes that can do both MPI and OpenMP, so that any reasonable system admin will set up machines, so that OMP_NUM_THREADS is explicitly set to 1 by default in the system environment. axel. > > Thanks for your time and support. > > -- > Cristian Degli Esposti Boschi > CNR-IMM, Sezione di Bologna, > via Gobetti, 101, 40129, Bologna, Italia > tel. ++39 051 6399152, fax ++39 051 6399216 > email: degliesposti -AT- bo.imm.cnr.it > web: www.bo.imm.cnr.it/site/staff/personal_pages/degliesposti/ > _______________________________________________ > Pw_forum mailing list > Pw_forum at pwscf.org > http://www.democritos.it/mailman/listinfo/pw_forum -- Dr. Axel Kohlmeyer akohlmey at gmail.com http://goo.gl/1wk0 International Centre for Theoretical Physics, Trieste. Italy.
