Re: [QE-users] seeking hints for nscf convergence in supercell
Dear Patrizio Are you sure you need ecutwfc=100 Ry and ecutrho=12*ecutwfc to converge US pseudopotentials? Have you performed convergence tests on simple systems or flicked through sssp tables (I see 30/240 for pslibrary US Mg and 40/320 for gbrv US Sb...)? However, for your nscf step you require a 45 45 31 k-points mesh. It's a lot of k-points. I don't know what is the purpose of your calculations, so I can't say if you are uselessly oversampling the Brillouin zone. But there is no free lunch... HTH Giuseppe PS maybe we are nearest neighbours, you may phone... :-D Quoting patrizio.grazi...@cnr.it: Hello everybody, I'm seeking advice to fasten nscf convergence (or make it possible). I'm using qe 6.8 to study Mg3Sb2 / Mg3Bi2 solid solutions, as first step I'm starting with a Mg3Sb2 2x2x2 supercell (40 atoms) to compare with simple Mg3Sb2 cell (5 atoms). The calculation track is scf --> nscf --> fs. For the simple cell it runs smoothly (and I'm confident it is okay as I compared with a colleague of mine relatively expert in DFT, not qe). For the supercell the scf converges in around 1 day of CPU time, but the nscf does not converge in ~ 12 days of CPU time - say 23 hours with 4 nodes, 6 tasks, 6 cpu-per-task, npool=24, are not sufficient to reach convergence. I attach the .in files for both scf and nscf, as well as the nscf out file. There are several eigenvalues not converged but I am not sure if this is the cause of the long time required for convergence (I've read in the forum this can be often regarded as a warning). So, since I have limited computational resources, before making many trial-and-error attempts I would seek hints and advices to understand if it is only a matter of make a more intensive calculation (I've read the restart mode is not recommended for nscf so I should ask for more nodes), or to improve the input file in an aware way. Sorry if this is a time waste for you, and thanks in advance for any support, Patrizio -- Patrizio Graziosi, PhD Research Scientist CNR - ISMN Institute for the Study of Nanostructured Materials GIUSEPPE MATTIOLI CNR - ISTITUTO DI STRUTTURA DELLA MATERIA Via Salaria Km 29,300 - C.P. 10 I-00015 - Monterotondo Scalo (RM) Mob (*preferred*) +39 373 7305625 Tel + 39 06 90672342 - Fax +39 06 90672316 E-mail: ___ Quantum ESPRESSO is supported by MaX (www.max-centre.eu) users mailing list users@lists.quantum-espresso.org https://lists.quantum-espresso.org/mailman/listinfo/users
Re: [QE-users] seeking hints for nscf convergence in supercell
Hi Giuseppe, thanks a lot for your time and support! The answer to the first question is honestly no, I took ecutwfc to a colleague of mine (for simple cell, not supercell) and ecutrho for USPP from the qe indications, I'll work on this. For the sampling, the final purpose is transport calculations and I'm afraid this sampling can be already too rough. If it's too much, I'll go for Wannier, for which unfortunately I've no experience (not yet). I'll follow your indications and, if the case, I'll feel free to phone you :D Thanks!!! Patrizio Giuseppe Mattioli ha scritto: Dear Patrizio Are you sure you need ecutwfc=100 Ry and ecutrho=12*ecutwfc to converge US pseudopotentials? Have you performed convergence tests on simple systems or flicked through sssp tables (I see 30/240 for pslibrary US Mg and 40/320 for gbrv US Sb...)? However, for your nscf step you require a 45 45 31 k-points mesh. It's a lot of k-points. I don't know what is the purpose of your calculations, so I can't say if you are uselessly oversampling the Brillouin zone. But there is no free lunch... HTH Giuseppe PS maybe we are nearest neighbours, you may phone... :-D Quoting patrizio.grazi...@cnr.it: Hello everybody, I'm seeking advice to fasten nscf convergence (or make it possible). I'm using qe 6.8 to study Mg3Sb2 / Mg3Bi2 solid solutions, as first step I'm starting with a Mg3Sb2 2x2x2 supercell (40 atoms) to compare with simple Mg3Sb2 cell (5 atoms). The calculation track is scf --> nscf --> fs. For the simple cell it runs smoothly (and I'm confident it is okay as I compared with a colleague of mine relatively expert in DFT, not qe). For the supercell the scf converges in around 1 day of CPU time, but the nscf does not converge in ~ 12 days of CPU time - say 23 hours with 4 nodes, 6 tasks, 6 cpu-per-task, npool=24, are not sufficient to reach convergence. I attach the .in files for both scf and nscf, as well as the nscf out file. There are several eigenvalues not converged but I am not sure if this is the cause of the long time required for convergence (I've read in the forum this can be often regarded as a warning). So, since I have limited computational resources, before making many trial-and-error attempts I would seek hints and advices to understand if it is only a matter of make a more intensive calculation (I've read the restart mode is not recommended for nscf so I should ask for more nodes), or to improve the input file in an aware way. Sorry if this is a time waste for you, and thanks in advance for any support, Patrizio -- Patrizio Graziosi, PhD Research Scientist CNR - ISMN Institute for the Study of Nanostructured Materials GIUSEPPE MATTIOLI CNR - ISTITUTO DI STRUTTURA DELLA MATERIA Via Salaria Km 29,300 - C.P. 10 I-00015 - Monterotondo Scalo (RM) Mob (*preferred*) +39 373 7305625 Tel + 39 06 90672342 - Fax +39 06 90672316 E-mail: ___ Quantum ESPRESSO is supported by MaX (www.max-centre.eu) users mailing list users@lists.quantum-espresso.org https://lists.quantum-espresso.org/mailman/listinfo/users -- Patrizio Graziosi, PhD Research Scientist CNR - ISMN Institute for the Study of Nanostructured Materials ___ Quantum ESPRESSO is supported by MaX (www.max-centre.eu) users mailing list users@lists.quantum-espresso.org https://lists.quantum-espresso.org/mailman/listinfo/users
Re: [QE-users] seeking hints for nscf convergence in supercell
I'll follow your indications and, if the case, I'll feel free to phone you :D Feel free indeed! We must think and act "corporate" :-o :-D The answer to the first question is honestly no, I took ecutwfc to a colleague of mine (for simple cell, not supercell) This depends only on the pseudopotential (and on your purposes), not on the size of the system. Best G. Quoting patrizio.grazi...@cnr.it: Hi Giuseppe, thanks a lot for your time and support! The answer to the first question is honestly no, I took ecutwfc to a colleague of mine (for simple cell, not supercell) and ecutrho for USPP from the qe indications, I'll work on this. For the sampling, the final purpose is transport calculations and I'm afraid this sampling can be already too rough. If it's too much, I'll go for Wannier, for which unfortunately I've no experience (not yet). I'll follow your indications and, if the case, I'll feel free to phone you :D Thanks!!! Patrizio Giuseppe Mattioli ha scritto: Dear Patrizio Are you sure you need ecutwfc=100 Ry and ecutrho=12*ecutwfc to converge US pseudopotentials? Have you performed convergence tests on simple systems or flicked through sssp tables (I see 30/240 for pslibrary US Mg and 40/320 for gbrv US Sb...)? However, for your nscf step you require a 45 45 31 k-points mesh. It's a lot of k-points. I don't know what is the purpose of your calculations, so I can't say if you are uselessly oversampling the Brillouin zone. But there is no free lunch... HTH Giuseppe PS maybe we are nearest neighbours, you may phone... :-D Quoting patrizio.grazi...@cnr.it: Hello everybody, I'm seeking advice to fasten nscf convergence (or make it possible). I'm using qe 6.8 to study Mg3Sb2 / Mg3Bi2 solid solutions, as first step I'm starting with a Mg3Sb2 2x2x2 supercell (40 atoms) to compare with simple Mg3Sb2 cell (5 atoms). The calculation track is scf --> nscf --> fs. For the simple cell it runs smoothly (and I'm confident it is okay as I compared with a colleague of mine relatively expert in DFT, not qe). For the supercell the scf converges in around 1 day of CPU time, but the nscf does not converge in ~ 12 days of CPU time - say 23 hours with 4 nodes, 6 tasks, 6 cpu-per-task, npool=24, are not sufficient to reach convergence. I attach the .in files for both scf and nscf, as well as the nscf out file. There are several eigenvalues not converged but I am not sure if this is the cause of the long time required for convergence (I've read in the forum this can be often regarded as a warning). So, since I have limited computational resources, before making many trial-and-error attempts I would seek hints and advices to understand if it is only a matter of make a more intensive calculation (I've read the restart mode is not recommended for nscf so I should ask for more nodes), or to improve the input file in an aware way. Sorry if this is a time waste for you, and thanks in advance for any support, Patrizio -- Patrizio Graziosi, PhD Research Scientist CNR - ISMN Institute for the Study of Nanostructured Materials GIUSEPPE MATTIOLI CNR - ISTITUTO DI STRUTTURA DELLA MATERIA Via Salaria Km 29,300 - C.P. 10 I-00015 - Monterotondo Scalo (RM) Mob (*preferred*) +39 373 7305625 Tel + 39 06 90672342 - Fax +39 06 90672316 E-mail: ___ Quantum ESPRESSO is supported by MaX (www.max-centre.eu) users mailing list users@lists.quantum-espresso.org https://lists.quantum-espresso.org/mailman/listinfo/users -- Patrizio Graziosi, PhD Research Scientist CNR - ISMN Institute for the Study of Nanostructured Materials ___ Quantum ESPRESSO is supported by MaX (www.max-centre.eu) users mailing list users@lists.quantum-espresso.org https://lists.quantum-espresso.org/mailman/listinfo/users GIUSEPPE MATTIOLI CNR - ISTITUTO DI STRUTTURA DELLA MATERIA Via Salaria Km 29,300 - C.P. 10 I-00015 - Monterotondo Scalo (RM) Mob (*preferred*) +39 373 7305625 Tel + 39 06 90672342 - Fax +39 06 90672316 E-mail: ___ Quantum ESPRESSO is supported by MaX (www.max-centre.eu) users mailing list users@lists.quantum-espresso.org https://lists.quantum-espresso.org/mailman/listinfo/users
