Dear Marcos, Do you have an example of an arch.make file using Intel compilers (any version) and -ipo? I haven't had much luck getting that to work, even with older versions of ifort. Specifically, if I add -ipo to FCFLAGS for ifort 10.0.x or 10.1.x, configure crashes with the following error:
checking whether ifort accepts -D... no checking whether ifort accepts -I... no checking whether ifort fulfils requested features... no checking preprocessing mode we may therefore use... indirect checking how to run the C preprocessor... icc -E checking how to preprocess Fortran files... fpp checking how to redirect fpp output... > conftest.f checking whether fpp needs the -P option... unknown configure: error: fpp cannot produce code that ifort compiles Without -ipo among the other optimization settings, configure decides that ifort can do direct preprocessing. If I try to use -ipo with the latest ifort 11.1, configure also recognizes that ifort can do direct preprocessing, although make ends up failing to build the executable. Many thanks, Mike On Sep 8, 2010, at 7:40 AM, Marcos Veríssimo Alves wrote: > Leila, > > The reason I wanted to see an output was because I suspect that the slowness > of your calculations could be linked to your compilation, actually. One thing > I have noticed is that you use the -g flag for the compilation, which > produces debugging info. This is only good in case you will debug the code in > case of an error; otherwise it is worth having one version compiled with -g > and one without; use the version without -g for production and, if you get an > error which you don't know what subroutine caused it, THEN use the one with > -g to determine it. The lack of -g could, by itself, speed your calculation > up. > > Next: optimizations. Use the flags -pad and -opt-prefetch. In my own > compilations they make quite a difference. If you manage to do it > successfully, do -ipo, which also speeds your code up significantly - up to > 30% in some cases! > > Finally, it is not clear from your output: since you compile with ifort, are > you using the intel mkl? If not, do compile with these libraries immediately > - Siesta's blas and lapack are not optimized, and the mkl also speed your > code up enormously. > > Good luck, > > Marcos > > 2010/9/7 leila <[email protected]> > Dear Marcos, > > Thank you very much for your reply! > > I used to do first-principles calculations on bcc Fe with SIESTA. The PP and > basis set are almost the same with Fu et al. (C.C. Fu, E. Willaime, and P. > Ordejón, Phys. Rev. Lett. 92 (2004) p.175503.), and also thoughtfully tested > in my work (L. Zhang, Journal of Physics: Condensed Matter 22 (2010) > p.375401.) I used these PP and basis for the MD test calculations. Please > see the input and output in the attached files. > > Because I am new to MD, I will be really grateful if you can point me > some difference in DFT and MD calculations with SIESTA, for example, the > basis, K points, or any other aspects. > > Thank you in advance! > > Yours sincerely, > > Leila > > > > 发件人: Marcos Veríssimo Alves [mailto:[email protected]] > 发送时间: 2010年9月7日 17:27 > 收件人: [email protected] > 主题: Re: [SIESTA-L] how to accerlarate MD calculations? > > > Leila, > > There could be many reasons for your md calculation taking so long to > complete. With two atoms and 5x5x5 grid, reasons could range from your using > an executable created with an inefficient compiler and/or without optimized > libraries, to a wrong geometric setup, bad calculation settings, heavy basis > set (f.e. very long) and many others that do not come to my mind right now. > If any of these is the reason, you should correct for them before resorting > to more incomplete basis sets, which will compromise the quality of your > calculation. Please provide some details on what you are calculating, and an > input and output, to check for possible problems. > > Marcos > > El 7 de sep de 2010, 10:25 a.m., "leila" <[email protected]> escribió: > > Hi, everyone! > > > I did an ab-initio molecular dynamics calculations (MD) by SIESTA with a > unit cell containing 2 atoms, using 5 × 5 × 5 k-point grid. The basis and PP > are the same as those for the first-principles calculations. > > I set 100 MD steps with 1.0 fs as the time length. However, the simple test > took more than 3 hours by serial processing. > > This test almost demonstrates the impossibility for ~100 atoms > calculations. > > > I wonder if there are any methods, for example using SZ, less K-points, or > larger smearing to make ab initio MD calculations feasible. > > > Any advice is welcome! Thanks in advance! > > > Yours sincerely, > > > Leila > > >
