Alan Fung <alanfung <at> ust.hk> writes: > > > I experience the same performance drop in my numerical simulations. With > GSL-1.15, the simulation takes 7.7s. With GSL-1.16, it takes 2m12s. My > simulation is to integrate a group of ODEs. GCC = 4.8.1. OS = Kubuntu 13.10. > > some more figures on other computers: > using the same simulation source code: > GSL-1.15 <at> Pentium 4 with GCC3.2: ~30s > GSL-1.16 <at> E7-4850 with GCC4.4: 5m > > I am quite sure that the problem is not due to different versions of GCC. I > believe that it should be due to differences between 1.15 and 1.16. > >
I have attached a version of my source code (sorry, this is not well-written :)). Actually, I have just found out why there is a regression. I found that, if i undo the revision 4771 (http://bzr.savannah.gnu.org/lh/gsl/trunk/revision/4771), the performance of 1.16 will be at par with 1.15. For the source code I have attached, I have done some tests on different versions. 1.15: Running time: 0m0.759s Output result: -0.378978 (center of mass of dynamical variables) 1.16 Running time: 0m10.670s Output result: -0.378996 1.16 (without revision 4771): Running time: 0m0.754s Output result: -0.378978 In revision 4771, the programmer intended to use e->dydt_out to update e->dydt_in, rather than calculating a new e->dydt_in. Based on my understanding on the source code, he thought that memcpy is better than a new calculation. But, in my case, it makes a big performance regression. Also, in my experiment, without 4771, results by 1.15 and 1.16 agree with each other. So, I would like to question which version is not reliable. For the reliability concern, I have sent an email to bug-gsl to question about this. ============================== The source code: http://www.filedropper.com/cannstdsubdelay
