Re: [Pw_forum] Using -nimage with phonon at q=0
On Thu, 2015-08-06 at 14:00 +0200, Merlin Meheut wrote: > On 05/08/2015 18:50, Andrea Dal Corso wrote: > > On Wed, 2015-08-05 at 17:18 +0200, Merlin Meheut wrote: > >> Dear PWSCF users, > >> > >> I recently discovered with great interest the possibilities to > >> parallelize phonon calculations using the -nimage option of ph.x. > >> (example given in espresso-4.3.2/examples/GRID_examples). > >> > > GRID examples refer to grid splitting, meaning that you split the > > calculation using start_irr, last_irr etc. and in principle the > > calculation can run in different machines. Then you have to collect the > > files yourself to finally obtain the results. > > Sorry I mixed up both. Actually, I used -nimage (as explained in > Doc/INPUT_PH.txt of version 4.3.2), but I would like to restart with > grid splitting. > Thank you for the reference to the example. > > >> However, I had a problem when performing calculations at gamma-point: > >> for other q-points (therefore with epsil=.false. and zue=.false.) > >> everything went as planned, but with q=0 (and epsil=.true. and > >> zue=.true.), this just did not work. I took 80 processors divided into 4 > >> images, and instead of dividing the different representations into 4 > >> pools, the four groups of processors realized the same calculation, > >> computing the same representations. I killed the calculation at some > >> point (I have computed the electric fields, effective charges and 218 > >> representations out of 564). I would like now to finish the computation > >> without redoing it, and I have several questions to achieve this goal: > >> > > Are you sure that all the images made the same phonon calculations, or > > all images made the electric field calculation but the phonon modes were > > different? > > The names of the partial dynamical matrices (dynmat.1.$irr.xml) are the > same, the files (e.g. _ph0/LiClMag2-1.phsave/dynmat.1.100.xml and > _ph1/LiClMag2-1.phsave/dynmat.1.100.xml) are almost identical, and the > calculation lasted more than it should have (each of the 4 separate > images has computed more than 141 representations, which is one quarter > of the total). So as far as I understand it, all the images made > electric field calculation and started making the same phonon mode > calculations in parallel. > Thank you for reporting this. What is not supported is the option -nimage together with gamma-gamma tricks that you are using in the gamma calculation. One easy solution is to set nogg=.TRUE. in the input of the phonon if you do not need the gamma-gamma tricks to speed-up the calculation. Andrea > > In this case, since you stopped the calculation, you can only > > collect the .xml files in a single _ph0 directory and restart without > > images. The restart with images is still poorly supported. > > I would like to restart without -nimage (since it seems to fail in > this particular case), but with grid splitting. Do you think this is > presomptuous? > By the way, I had a positive experience restarting with -nimage on > another run (at q neq 0) . > > > In the last version of QE, the dielectric constant and effective charges > > are saved in the tensors.xml file in the _ph0 directory. > > Thank you very much for your help! > > Best regards, > > > -- > Merlin Méheut, Géosciences et Environnement Toulouse, > OMP, 14 avenue Edouard Belin, 31400 Toulouse, France > > phone +33 (0)5 61 33 26 17, fax +33 (0)5 61 33 25 60 > > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum -- Andrea Dal CorsoTel. 0039-040-3787428 SISSA, Via Bonomea 265 Fax. 0039-040-3787249 I-34136 Trieste (Italy) e-mail: dalco...@sissa.it ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] Using -nimage with phonon at q=0
If the input file is correct? On Thu, Aug 6, 2015 at 4:30 PM, Merlin Meheut wrote: > On 05/08/2015 18:50, Andrea Dal Corso wrote: > > On Wed, 2015-08-05 at 17:18 +0200, Merlin Meheut wrote: > >> Dear PWSCF users, > >> > >> I recently discovered with great interest the possibilities to > >> parallelize phonon calculations using the -nimage option of ph.x. > >> (example given in espresso-4.3.2/examples/GRID_examples). > >> > > GRID examples refer to grid splitting, meaning that you split the > > calculation using start_irr, last_irr etc. and in principle the > > calculation can run in different machines. Then you have to collect the > > files yourself to finally obtain the results. > > Sorry I mixed up both. Actually, I used -nimage (as explained in > Doc/INPUT_PH.txt of version 4.3.2), but I would like to restart with > grid splitting. > Thank you for the reference to the example. > > >> However, I had a problem when performing calculations at gamma-point: > >> for other q-points (therefore with epsil=.false. and zue=.false.) > >> everything went as planned, but with q=0 (and epsil=.true. and > >> zue=.true.), this just did not work. I took 80 processors divided into 4 > >> images, and instead of dividing the different representations into 4 > >> pools, the four groups of processors realized the same calculation, > >> computing the same representations. I killed the calculation at some > >> point (I have computed the electric fields, effective charges and 218 > >> representations out of 564). I would like now to finish the computation > >> without redoing it, and I have several questions to achieve this goal: > >> > > Are you sure that all the images made the same phonon calculations, or > > all images made the electric field calculation but the phonon modes were > > different? > > The names of the partial dynamical matrices (dynmat.1.$irr.xml) are the > same, the files (e.g. _ph0/LiClMag2-1.phsave/dynmat.1.100.xml and > _ph1/LiClMag2-1.phsave/dynmat.1.100.xml) are almost identical, and the > calculation lasted more than it should have (each of the 4 separate > images has computed more than 141 representations, which is one quarter > of the total). So as far as I understand it, all the images made > electric field calculation and started making the same phonon mode > calculations in parallel. > > > In this case, since you stopped the calculation, you can only > > collect the .xml files in a single _ph0 directory and restart without > > images. The restart with images is still poorly supported. > > I would like to restart without -nimage (since it seems to fail in > this particular case), but with grid splitting. Do you think this is > presomptuous? > By the way, I had a positive experience restarting with -nimage on > another run (at q neq 0) . > > > In the last version of QE, the dielectric constant and effective charges > > are saved in the tensors.xml file in the _ph0 directory. > > Thank you very much for your help! > > Best regards, > > > -- > Merlin Méheut, Géosciences et Environnement Toulouse, > OMP, 14 avenue Edouard Belin, 31400 Toulouse, France > > phone +33 (0)5 61 33 26 17, fax +33 (0)5 61 33 25 60 > > ___ > Pw_forum mailing list > Pw_forum@pwscf.org > http://pwscf.org/mailman/listinfo/pw_forum relax.in Description: Binary data ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] Using -nimage with phonon at q=0
On 05/08/2015 18:50, Andrea Dal Corso wrote: > On Wed, 2015-08-05 at 17:18 +0200, Merlin Meheut wrote: >> Dear PWSCF users, >> >> I recently discovered with great interest the possibilities to >> parallelize phonon calculations using the -nimage option of ph.x. >> (example given in espresso-4.3.2/examples/GRID_examples). >> > GRID examples refer to grid splitting, meaning that you split the > calculation using start_irr, last_irr etc. and in principle the > calculation can run in different machines. Then you have to collect the > files yourself to finally obtain the results. Sorry I mixed up both. Actually, I used -nimage (as explained in Doc/INPUT_PH.txt of version 4.3.2), but I would like to restart with grid splitting. Thank you for the reference to the example. >> However, I had a problem when performing calculations at gamma-point: >> for other q-points (therefore with epsil=.false. and zue=.false.) >> everything went as planned, but with q=0 (and epsil=.true. and >> zue=.true.), this just did not work. I took 80 processors divided into 4 >> images, and instead of dividing the different representations into 4 >> pools, the four groups of processors realized the same calculation, >> computing the same representations. I killed the calculation at some >> point (I have computed the electric fields, effective charges and 218 >> representations out of 564). I would like now to finish the computation >> without redoing it, and I have several questions to achieve this goal: >> > Are you sure that all the images made the same phonon calculations, or > all images made the electric field calculation but the phonon modes were > different? The names of the partial dynamical matrices (dynmat.1.$irr.xml) are the same, the files (e.g. _ph0/LiClMag2-1.phsave/dynmat.1.100.xml and _ph1/LiClMag2-1.phsave/dynmat.1.100.xml) are almost identical, and the calculation lasted more than it should have (each of the 4 separate images has computed more than 141 representations, which is one quarter of the total). So as far as I understand it, all the images made electric field calculation and started making the same phonon mode calculations in parallel. > In this case, since you stopped the calculation, you can only > collect the .xml files in a single _ph0 directory and restart without > images. The restart with images is still poorly supported. I would like to restart without -nimage (since it seems to fail in this particular case), but with grid splitting. Do you think this is presomptuous? By the way, I had a positive experience restarting with -nimage on another run (at q neq 0) . > In the last version of QE, the dielectric constant and effective charges > are saved in the tensors.xml file in the _ph0 directory. Thank you very much for your help! Best regards, -- Merlin Méheut, Géosciences et Environnement Toulouse, OMP, 14 avenue Edouard Belin, 31400 Toulouse, France phone +33 (0)5 61 33 26 17, fax +33 (0)5 61 33 25 60 ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum
Re: [Pw_forum] Using -nimage with phonon at q=0
On Wed, 2015-08-05 at 17:18 +0200, Merlin Meheut wrote: > Dear PWSCF users, > > I recently discovered with great interest the possibilities to > parallelize phonon calculations using the -nimage option of ph.x. > (example given in espresso-4.3.2/examples/GRID_examples). > GRID examples refer to grid splitting, meaning that you split the calculation using start_irr, last_irr etc. and in principle the calculation can run in different machines. Then you have to collect the files yourself to finally obtain the results. Image splitting with the -nimage option is in Image_example. In this case the calculation is split automatically, but at the end you need to do another calculation without images to collect the results. Please use QE 5.2 if you want to use these features. > However, I had a problem when performing calculations at gamma-point: > for other q-points (therefore with epsil=.false. and zue=.false.) > everything went as planned, but with q=0 (and epsil=.true. and > zue=.true.), this just did not work. I took 80 processors divided into 4 > images, and instead of dividing the different representations into 4 > pools, the four groups of processors realized the same calculation, > computing the same representations. I killed the calculation at some > point (I have computed the electric fields, effective charges and 218 > representations out of 564). I would like now to finish the computation > without redoing it, and I have several questions to achieve this goal: > Are you sure that all the images made the same phonon calculations, or all images made the electric field calculation but the phonon modes were different? In this case, since you stopped the calculation, you can only collect the .xml files in a single _ph0 directory and restart without images. The restart with images is still poorly supported. > - is there a particular procedure for using -nimage with epsil=.true. > and zue=.true., or is that just not foreseen? In other word, did I miss > something? > - following the same idea, if I want to build my dynamical matrix, with > effective charges and dielectric tensor, by a ph.x run with > "recover=.true.", can I do it and if I can, what files do I need in > _ph0? In particular, what are the files that contain the information on > dielectric tensor and effective charges? In other words, are there > special guidelines in supplement to the ones given in > espresso-4.3.2/Doc/INPUT_PH.txt to separate the phonon calculation in > several jobs, when we consider a calculation with epsil=.true. and zue > =.true. ? In the last version of QE, the dielectric constant and effective charges are saved in the tensors.xml file in the _ph0 directory. HTH, Andrea > ( ADDITIONAL INFORMATION lines 562-end) > > Thank you in advance for any help, > > the version of qe is 5.1 > > I did first a scf calculation on 20 processors: > > -scf input file -- > &control > calculation = 'scf', >restart_mode = 'from_scratch' , > prefix = 'LiClMag2-1', > disk_io = 'default' , > pseudo_dir = '$WORKDIR', > outdir = '${WORKDIR}', > tprnfor= .true., > tstress= .true., > /&end > &system > ibrav = 0, celldm(1)=23.3535, > nat = 188, ntyp = 4, ecutwfc = ${a}.0, ecutrho=${b}.0 > /&end > &electrons > electron_maxstep = 100, >conv_thr = 1.d-8, > mixing_mode = 'plain', > startingwfc = 'atomic', > mixing_beta = 0.5, > /&end > ATOMIC_SPECIES >Li7.0160 Li.blyp-sl-rrkjus_psl.1.0.0.UPF >O15.9949 O.blyp.UPF >H 1.0079 H.blyp2.UPF >Cl 34.9689 Cl.blyp-nl-rrkjus_psl.1.0.0.UPF > > ATOMIC_POSITIONS (angstrom) > (...) > K_POINTS {crystal} > 1 > 0.0 0.0 0.0 1 > > CELL_PARAMETERS { cubic } >1.0 0.00. >0.0 1.00. >0.0 0.01. > > > the scf was run on 20 processors > > srun pw.x -npool 1 < scf.${PREFIX}.inp > scf.${PREFIX}.out > > the ph input is : > -- > &inputph > amass(1)= 7.0160, > amass(2)=15.9949, > amass(3)= 1.0079, > amass(4)=34.9689, > ! ldisp=.true., nq1=2, nq2=2, nq3=2, > alpha_mix(1) = 0.7, > tr2_ph = 1.0D-18, > prefix='LiClMag2-1', > fildyn='mat.$PREFIX', > epsil =.true., > trans =.true., > zue = .true., > lraman=.false., > outdir = '$WORKDIR', > ! max_seconds= 18, > /&end > 0.000 0.000 0.000 > --- > > It was run on 20 processors using: > > srun ph.x -npool 1 -nimage 4 < ph.${PREFIX}.inp > ph.${PREFIX}.out > > > -- > Merlin Méheut, Géo
[Pw_forum] Using -nimage with phonon at q=0
Dear PWSCF users, I recently discovered with great interest the possibilities to parallelize phonon calculations using the -nimage option of ph.x. (example given in espresso-4.3.2/examples/GRID_examples). However, I had a problem when performing calculations at gamma-point: for other q-points (therefore with epsil=.false. and zue=.false.) everything went as planned, but with q=0 (and epsil=.true. and zue=.true.), this just did not work. I took 80 processors divided into 4 images, and instead of dividing the different representations into 4 pools, the four groups of processors realized the same calculation, computing the same representations. I killed the calculation at some point (I have computed the electric fields, effective charges and 218 representations out of 564). I would like now to finish the computation without redoing it, and I have several questions to achieve this goal: - is there a particular procedure for using -nimage with epsil=.true. and zue=.true., or is that just not foreseen? In other word, did I miss something? - following the same idea, if I want to build my dynamical matrix, with effective charges and dielectric tensor, by a ph.x run with "recover=.true.", can I do it and if I can, what files do I need in _ph0? In particular, what are the files that contain the information on dielectric tensor and effective charges? In other words, are there special guidelines in supplement to the ones given in espresso-4.3.2/Doc/INPUT_PH.txt to separate the phonon calculation in several jobs, when we consider a calculation with epsil=.true. and zue =.true. ? ( ADDITIONAL INFORMATION lines 562-end) Thank you in advance for any help, the version of qe is 5.1 I did first a scf calculation on 20 processors: -scf input file -- &control calculation = 'scf', restart_mode = 'from_scratch' , prefix = 'LiClMag2-1', disk_io = 'default' , pseudo_dir = '$WORKDIR', outdir = '${WORKDIR}', tprnfor= .true., tstress= .true., /&end &system ibrav = 0, celldm(1)=23.3535, nat = 188, ntyp = 4, ecutwfc = ${a}.0, ecutrho=${b}.0 /&end &electrons electron_maxstep = 100, conv_thr = 1.d-8, mixing_mode = 'plain', startingwfc = 'atomic', mixing_beta = 0.5, /&end ATOMIC_SPECIES Li7.0160 Li.blyp-sl-rrkjus_psl.1.0.0.UPF O15.9949 O.blyp.UPF H 1.0079 H.blyp2.UPF Cl 34.9689 Cl.blyp-nl-rrkjus_psl.1.0.0.UPF ATOMIC_POSITIONS (angstrom) (...) K_POINTS {crystal} 1 0.0 0.0 0.0 1 CELL_PARAMETERS { cubic } 1.0 0.00. 0.0 1.00. 0.0 0.01. the scf was run on 20 processors srun pw.x -npool 1 < scf.${PREFIX}.inp > scf.${PREFIX}.out the ph input is : -- &inputph amass(1)= 7.0160, amass(2)=15.9949, amass(3)= 1.0079, amass(4)=34.9689, ! ldisp=.true., nq1=2, nq2=2, nq3=2, alpha_mix(1) = 0.7, tr2_ph = 1.0D-18, prefix='LiClMag2-1', fildyn='mat.$PREFIX', epsil =.true., trans =.true., zue = .true., lraman=.false., outdir = '$WORKDIR', ! max_seconds= 18, /&end 0.000 0.000 0.000 --- It was run on 20 processors using: srun ph.x -npool 1 -nimage 4 < ph.${PREFIX}.inp > ph.${PREFIX}.out -- Merlin Méheut, Géosciences et Environnement Toulouse, OMP, 14 avenue Edouard Belin, 31400 Toulouse, France Université Paul Sabatier - Toulouse 3 phone +33 (0)5 61 33 26 17, fax +33 (0)5 61 33 25 60 ___ Pw_forum mailing list Pw_forum@pwscf.org http://pwscf.org/mailman/listinfo/pw_forum