tis 31 mars 2015 kl 18:32 skrev Jan Blechta <[email protected]>:
> On Mon, 30 Mar 2015 14:13:30 +0000 > Anders Logg <[email protected]> wrote: > > > The only thing that matters is that it's small enough. It overhead > > should be so small that it does not discourage users (developers) > > from putting it into the code (except for at the very low level). > > > > If there are technical reasons for using boost::cpu_timer, then I > > guess it's ok. > > The reason for boost is that it provides wall, user and system time by, > say > > list_timings(TimingClear_keep, > [TimingType_wall, TimingType_user, TimingType_system]) > > see timing demo. Moreover it should work on any supported platform. > > Jan > ok, good then. -- Anders > > > > -- > > Anders > > > > > > > > mån 30 mars 2015 kl 15:13 skrev Jan Blechta > > <[email protected]>: > > > > > On Mon, 30 Mar 2015 13:55:55 +0200 > > > Jan Blechta <[email protected]> wrote: > > > > > > > On Mon, 30 Mar 2015 11:34:19 +0000 > > > > Anders Logg <[email protected]> wrote: > > > > > > > > > See this question on the QA forum: > > > > > > > > > > http://fenicsproject.org/qa/6875/ubuntu-compile-from- > > > source-which-provide-better-performance > > > > > > > > > > The Cahn-Hilliard demo takes 40 seconds with 1.3 Ubuntu > > > > > packages and 52 seconds with 1.5+ built from source. Are these > > > > > regressions in performance or is Johannes that much better at > > > > > building Debian packages than I am building FEniCS (with > > > > > HashDist). > > > > > > > > > > PS: Looking at the benchbot, there seem to have been some > > > > > regressions in the timing facilities with the recent changes: > > > > > > > > Ok, I will look if something can be done with common-timing-cpp > > > > regression. I have a guess what causes that. > > > > > > I haven't profiled it rigorously yet but it seems that most of the > > > regression is due to switching from gettimeofday() from sys/time.h > > > to boost::cpu_timer. > > > > > > Anders, is there any strong reason for improving the current > > > timings? Isn't it fast enough? I don't think that we use Timer > > > class in any loop such tight that this slowddown would have effect > > > and I would hardly expect users doing it. > > > > > > Jan > > > > > > > > > > > Nevertheless, please note slowdown of la-vector-access-cpp. > > > > http://fenicsproject.org/benchbot/la-vector-access-cpp_ > > > last_five_years.png > > > > > > > > It could be a cause of the regression of Cahn-Hilliard demo. You > > > > could even try running it with uBLAS/UmfpackLUSolver on both > > > > versions to see whether the problem is in PETSc/wrappers. > > > > > > > > Jan > > > > > > > > > > > > > > http://fenicsproject.org/benchbot/ > > > > > > > > > > -- > > > > > Anders > > > > > > > > _______________________________________________ > > > > fenics mailing list > > > > [email protected] > > > > http://fenicsproject.org/mailman/listinfo/fenics > > > > > > > >
_______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
