mirror of https://gitlab.com/QEF/q-e.git
301 lines
16 KiB
Plaintext
301 lines
16 KiB
Plaintext
TODO LIST - September 2012
|
|
|
|
0) Suspected bugs/problems:
|
|
0.1 something happened to elf calculation between 4.1 to 4.2
|
|
0.2 nstep ignored in vc-relax and other cases
|
|
0.3 wf_collect and phonons on the grid + unneeded scf calculations
|
|
0.4 phonon restart should not recalculate the entire initialization
|
|
0.5 Small energy differences between PW and CP - maybe PP-related,
|
|
but might also be a problem of Ewald calculation in CP ("raggio"?)
|
|
0.6 Numerical instabilities with: BLYP; TPSS; noncollinear GGA; LDA+U
|
|
|
|
1) Organization
|
|
1.0 Binary packages
|
|
1.0.0 At least Debian/RPM for Linux
|
|
1.3 svn/packaging:
|
|
1.3.2 run source-normalizer script dev-tools/src-normal
|
|
1.3.3 Files in flib/ use modules in Modules/, while files in Modules/
|
|
in turn call files in flib/. This kind of circular dependency
|
|
should be avoided. There isn't any good reason for the distinction
|
|
between 'modules' and 'libraries' anyway
|
|
1.4 mailing lists:
|
|
1.4.1 Send monthly reminder? possibly with netiquette
|
|
1.4.2 move to forums?
|
|
1.7 Testing
|
|
1.7.1 The automated tests should be moved to a machine that can
|
|
send e-mail to q-e-commits
|
|
1.7.2 The automated tests should be performed only when something has
|
|
actually changed, for executables that have actually changed.
|
|
The current mechanism is way too rough.
|
|
1.7.3 Many functionalities are still untested and there is not even an
|
|
example. PP/ is a notable offender: there are a lot of possible
|
|
calculations, most of which are never tested. Other untested
|
|
features: contraints, thermostats, ...
|
|
|
|
2) Documentation
|
|
|
|
2.1 Better (and shorter) FAQ list. Notably missing:
|
|
- hardware for QE
|
|
- segmentation fault (present but not visible enough)
|
|
2.2 QE-Tutorial: should be moved to the svn of qe-forge.
|
|
2.3 Documentation on how to generate PP for GIPAW calculation is missing
|
|
|
|
3) Pseudopotentials
|
|
|
|
3.1 Set up standard tests for all PPs
|
|
3.3 Implement OEP PPs
|
|
3.4 Standard recipe for generating reasonable PPs:
|
|
- Set of prescriptions for generation
|
|
- Criteria for portability
|
|
- Update PP table with PPs that follow the recipe
|
|
3.5 Extend documentation in pseudo-gen.tex to the case of USPP
|
|
(eventually to PAW)
|
|
3.6 Add default cutoff values as well as radii of augmented charge
|
|
in pseudopotential files (they are present in recent PP only)
|
|
3.7 The intermediate hard NCPP is useless for PAW and should not be
|
|
done when generating a PAW set
|
|
3.8 Finish merge of Meta-GGA (TPSS) code into atomic code
|
|
|
|
4) Development
|
|
|
|
4.0 Major highly desirable restructuring:
|
|
4.0.1 Add the possibility to run NEB/PH in a dynamical way, with a
|
|
home-made scheduler that executes tasks as soon as resources are
|
|
free; might be used for image parallelization as well.
|
|
4.0.2 Better-structured relaxation and molecular dynamics (including the
|
|
variable-cell case), with more extensive integration of PW and CP
|
|
(langevin code)
|
|
4.0.4 Accurate cleaning of exx.f90. "If" loop in Vexx is confusing.
|
|
Exx_divergence is confusing and contains some suspect values.
|
|
4.0.5 Martyna-Tuckerman has to be extended to any system dimensionality and
|
|
to exx, using coulomb_vcut trick.
|
|
|
|
4.1 New developments to be added (sooner or later):
|
|
4.1.1 new xml input/output
|
|
4.1.2 XMCD in XSpectra
|
|
4.1.3 Genetic Algorithm
|
|
4.1.4 LDA+U with Wannier functions
|
|
4.1.5 EPW (electron-phonon with Wannier stuff): released,
|
|
has to be aligned with current version
|
|
4.1.6 Converse NMR
|
|
4.1.7 QM-MM with MS2 plugin
|
|
4.1.8 Wannier-based exact-exchange in CP (available)
|
|
|
|
4.2 Small new developments, desirable or to be added:
|
|
4.2.0 Fermi energy in insulators (tetrahedron method) should either be
|
|
at the top of the valence band or in the middle of the gap; now
|
|
it is sometimes at the bottom of the conduction band (Eduardo)
|
|
4.2.1 configure issues:
|
|
- ifort+acml
|
|
- ifort+mkl+scalapack troubles
|
|
- add search for -lfftw3xf_intel
|
|
4.2.2 constraints should be implemented in all cases;
|
|
a check should be added if constraints break the symmetry
|
|
4.2.3 inversion symmetry should allow real hamiltonian and wavefunctions
|
|
4.2.4 nscf calculations are slow. There must be a way to make a better
|
|
usage of the available information from the scf calculation:
|
|
wavefunctions are just discarded. Same for phonon calculation:
|
|
it shouldn't be needed to recalculate everything almost from
|
|
scratch at each different q-point
|
|
4.2.5 Fermi-Dirac: pass T instead of "broadening", make it possible
|
|
to use it on top of smearing for free-energy calculations
|
|
4.2.7 matdyn should write frequencies in a format that is suitable
|
|
for direct plotting by gnuplot/xmgrace - see also Eyvaz' script
|
|
for phonon plotting. Also about phonons:
|
|
- perform calculations now performed by dynmat.x at the end of a
|
|
phonon run at gamma?
|
|
- perform calculations now performed by q2r.x: C(q)=>C(R), at the
|
|
end of a dispersion run
|
|
- projected phonon DOS with tetrahedra
|
|
- adapt plotband.x to phonon case; in general, simplify phonon plotting
|
|
- Bring QHA calculations inside matdyn? see fqha.f90
|
|
4.2.8 Interface to RESP calculation - requires adding radii to xml file
|
|
4.2.9 elf for USPP/PAW ; delocalization indices
|
|
4.1.10 Collection of tools and utilities for data analysis
|
|
(things like g(r) from MD simulations). Also: for ev.x,
|
|
write output file with E(V) for direct plotting of EOS
|
|
4.2.11 Gamma: same input as for PH
|
|
4.2.12 Cleanup of the 'prefix.EXIT' stuff. Stop with signals?
|
|
Also: stop with 'prefix.EXIT' and restart in D3 and Gamma (KK)
|
|
4.2.13 Various defaults for CP (proposed from Princeton):
|
|
- emass(emass=300), dt (dt=7), for preconditioning cutoff (3)
|
|
- automatic box grid for USPP from radii of augmented charge
|
|
- Electronic minimization: damp as default, sd discouraged
|
|
introduce an automatic default schedule, something as:
|
|
1st step sd followed by 5 steps with with damp= 0.8,
|
|
followed by 5 steps with damp=0.5,
|
|
followed by 10 steps with damp=0.3,
|
|
followed by 10 steps with damp=0.2,
|
|
followed by as many steps as necessary
|
|
to achieve the required convergence with damp=0.1
|
|
A max number of steps should be included to ensure program termination.
|
|
The other option allowed should be conjugate gradients:
|
|
see NM - it could one day become the default
|
|
- Ionic minimization: again SD should be discouraged
|
|
A default scheme for simultaneous damped dynamics should be given
|
|
(to be tested) for example: zero damp on ions and start with damp=0.5
|
|
on electrons to become then 0.1 or perhaps the values should be set
|
|
given the forces on the ions
|
|
When moving ions and electrons simultaneously an important parameter
|
|
is the ratio between electron and ion masses - For minimization it
|
|
is better to set up all the ion masses equal - A default value for
|
|
the ion masses (considering the defaults for emass and dt) is perhaps
|
|
10 AMU (we should do some test to see if 20 AMU is s a safer value)
|
|
- Default values for randomization should be given
|
|
amprp=0.1 is a decent value - amprp=0.01 is too small
|
|
- Car-Parrinello dynamics: the proper masses for the ions, an optimal
|
|
value for emass and dt should be set up by the code, based on the
|
|
smaller atomic mass and the default value used in the minimization
|
|
e.g. Amass_default=10 AMU. If the minimum physical AMASS is 20 then
|
|
dt=sqrt(2) dt_default and emass should be increased so to keep
|
|
emass^2/dt constant
|
|
- defaults for the Nose thermostat
|
|
|
|
4.3 Performance enhancements/Parallelization:
|
|
4.3.1 make hpsi/spsi/CG faster
|
|
- remove complex factor i**l from beta fct and q(r)
|
|
- shift structure factor from beta to psi when computing becp
|
|
(reduce memory)
|
|
- use real BLAS routine instead of COMPLEX one in hpsi/spsi
|
|
(at least 2 times faster).
|
|
- use only half of the G's when computing real integrals
|
|
(2 times faster)
|
|
- seek for CG and DIIS algoritms that only use (H-eS)|psi>
|
|
and not the two vector separately ... compute it in one single
|
|
call. In this way S|psi> is inexpensive
|
|
4.3.2 Try the new "Density-Matrix-Based" diagonalization algorithm
|
|
4.3.3 image parallelization of the phonon code: irreps and q-vectors
|
|
should be distributed across processors/grid computers.
|
|
Could be done in the same way as for NEB?
|
|
4.3.4 PH: use charge mixing instead of potential mixing
|
|
4.3.5 D3: verify status of parallelization, clean it up if needed
|
|
|
|
4.4 Cleanup
|
|
4.4.1 Increase modularization by
|
|
- collecting variables and routines acting on those variables
|
|
into modules
|
|
- classifying modules in a hierarchical way
|
|
- avoiding as much as possible that modules depend on many
|
|
other modules
|
|
4.4.2 Avoid monster routines that do too many things at the same time
|
|
depending on the value of too many variables. An example:
|
|
read_file
|
|
4.4.3 There is some confusion in the various initialization steps:
|
|
- default values at startup
|
|
- reading of the input data and copy into internal variables
|
|
- reading from data file
|
|
- initialization of general variables (that presumably will
|
|
be written to or read from file)
|
|
- initialization of variables used in a specific calculation
|
|
(that may not be written to or read from the data file)
|
|
All these steps are intermixed and/or replicated and it is
|
|
never clear what is initialized where. Same for variable
|
|
allocation: see recent GIPAW workaround for an example of
|
|
allocation confusion (qnorm, cell_factor in allocate_nlpot)
|
|
4.4.4 More PW/CP merge:
|
|
- f_inp, fixed occupancies
|
|
- lda+U modules
|
|
- makov-payne
|
|
- "cellmd" module of PW and "cell_base" of CP
|
|
- PW "real-space" approach / CP "small boxes"
|
|
- there should be a single function or routine for periodic boundary
|
|
conditions (i.e. bringing all atoms inside the unit cell)
|
|
- spherical harmonics and integration routines
|
|
- merge of atomic positions! currently CP uses a complex logic
|
|
that is very hard to follow
|
|
4.4.6 adding/removing/modifying input variables is too complex
|
|
Why are some checks on input variables performed in read_namelist,
|
|
while others apparently similar are in */input.f90?
|
|
4.4.7 Units: all units should be clearly documented and printed
|
|
on output (and also it should be clearly stated what the
|
|
printed quantites are)
|
|
4.4.8 There should be a function calculating dj_l/dx;
|
|
j_l with l=-1 should not be needed
|
|
4.4.9 too many confusing error messages are still around
|
|
4.4.10 Output should be more informative and less confused, better
|
|
structured, and ready for automatic reading (.e.g by xcrysden)
|
|
4.4.11 any possibility to merge the various solve_* in PH ?
|
|
4.4.12 Replace "use pwcom" with more "use" statements
|
|
4.4.13 Move all plots requiring Fourier (or real-space) interpolation
|
|
into pawplot.x, leaving in pp.x only gaussian cube and 3d xsf
|
|
files. Plots of sums and differences should be performed using
|
|
data files ready for plotting (gnuplot, xsf, cube; may require
|
|
some tools). pp.x should be simplified a lot and intermediate
|
|
format should disappear. Also: there is no reason to have dos.x
|
|
together with projwfc.x
|
|
4.4.14 All allocated variables should be deallocated at the end:
|
|
it makes easier to find memory leaks. Currently most variables
|
|
are deallocated, but a few (mostly in ffts and in pseudopotential
|
|
reading) aren't
|
|
4.4.16 add_efield must be rewritten from scratch: it is a mess beyond control
|
|
4.4.17 Some postprocessing codes could be used with command-line
|
|
options instead of fortran input
|
|
4.4.18 What about transforming 'bands'/'nscf' into a postprocessing
|
|
code?
|
|
|
|
4.5 Trouble-makers. inconsistencies, etc
|
|
4.5.1 Negative Charge problems (see qe-forge, H on graphene)
|
|
4.5.3 G-vector shells, especially in the variable-cell case, and the
|
|
various tricks to reduce cpu by not re-calculating things that
|
|
depend on |G| only (see e.g. qvan2). Maybe we should move to
|
|
interpolation of all quantities and get rid of shells and tricks
|
|
4.5.4 PP: complete postprocessing in Gamma case (only average missing),
|
|
and with CP data (in the latter case: when the data file does
|
|
not contain the charge/potential, issue an error message saying
|
|
what is missing and why instead of just crashing in iotk)
|
|
4.5.5 CP: add error check if dt^2/emass too large does not allow ortho
|
|
to converge or cause energy to increase as time step evolve
|
|
4.5.6 epsilon.x should be extended at least to have the nonlocal
|
|
contribution included; there should be a pointer in the
|
|
documentation explaining how to make a better calculation.
|
|
4.5.7 There should be a check on the FFT grid preventing a bad
|
|
choice of Nr1,Nr2,Nr3 (e.g. different Nr for axis of the
|
|
same length or even worse related by symmetry): this is a
|
|
frequent source of trouble with electron-phonon calculations
|
|
4.5.8 Still a few quirks with the atomic coordinate parser, if
|
|
DOS characters or tabulators are present (Lorenzo)
|
|
4.5.9 Spin-polarized cases: input is clumsy, confusing, error-prone;
|
|
the case occupations='from_input' is implemented as a special
|
|
case, for no apparent good reason.
|
|
4.5.10 k-points in crystal IBZ should be correctly calculated also
|
|
if input k-points are not in the IBZ of the lattice. It should
|
|
sufficient to use brute forces instead of group theory:
|
|
be expand to full BZ, remove symmetry-equivalent k.
|
|
Or maybe there is a problem with grids not having the
|
|
symmetry of the crystal?
|
|
|
|
5) Files and I/O
|
|
|
|
5.1 The "buffer" trick to keep wfc in memory wasn't such a great idea
|
|
after all: a better approach would be to have a k-point index
|
|
5.2 Scratch files are a big mess. It should be possible to open files
|
|
in places other than tmp_dir without resorting to obscure coding.
|
|
This is especially serious for PH, D3 etc
|
|
5.3 There should be a clearer distinction, both in the code and
|
|
in the input data, between directories to be read (and left
|
|
unchanged), directories to be (over-)written, temporary files
|
|
or directories
|
|
5.4 Some inconsistencies between PW and CP in the xml file format
|
|
(and inconsistencies with the documentation). Also: CP should
|
|
behave like PW and create a directory if not existent
|
|
5.5 Use qexml for xml file processing so that a single, easily
|
|
exportable routine, is used everywhere. By reading the xml
|
|
file in parallel, this should be simple (if the xml file is not
|
|
accessible by all processors, we can just make a local copy
|
|
with a small C routine and read it).
|
|
This is also an opportunity to update the file format with
|
|
1) removal of obsolete variables,
|
|
2) addition of variables that should be present,
|
|
3) possibility to reduce the number of files and directories
|
|
(hardcoded limit, more kpoints per directory, would replace
|
|
and extend lkpoint_dir)
|
|
5.6 There should be a lock mechanism that prevents people from
|
|
overwriting files of running processes. Should be done with
|
|
care, or else every time a code crashes will make the following
|
|
one crash as well!
|
|
5.7 Add rotation of restart files, similar to what is done in CPMD:
|
|
"The number of distinct RESTART files generated during CPMD runs
|
|
is read from the next line. The restart files are written in turn.
|
|
Default is 1. If you specify e.g. 3, then the files RESTART.1,
|
|
RESTART.2, RESTART.3 are used in rotation."
|