Parallelism for the ground state using wavelets¶
Boron cluster, alkane molecule…¶
This tutorial explains how to run the calculation of an isolated system using a wavelet basis-set on a parallel computer using MPI. You will learn the different characteristics of a parallel run using the wavelet basis-set and test the speed-up on a small boron cluster of 14 atoms followed by a test on a bigger alkane molecule.
This tutorial should take about 90 minutes and requires you have several CPU cores (up to 64).
You are supposed to know already some basics of parallelism in ABINIT, explained in the tutorial A first introduction to ABINIT in parallel.
Note
Supposing you made your own install of ABINIT, the input files to run the examples are in the ~abinit/tests/ directory where ~abinit is the absolute path of the abinit top-level directory. If you have NOT made your own install, ask your system administrator where to find the package, especially the executable and test files.
To execute the tutorials, create a working directory (Work*) and
copy there the input files and the files file of the lesson. This will be explicitly mentioned in the first lessons,
that will tell you more about the files file (see also section 1.1).
The files file ending with _x (e.g. tbase1_x.files) must be edited every time you start to use a new input file.
Most of the tutorials do not rely on parallelism (except specific tutorials on parallelism). However you can run most of the tutorial examples in parallel, see the topic on parallelism.
In case you work on your own PC or workstation, to make things easier, we suggest you define some handy environment variables by executing the following lines in the terminal:
export ABI_HOME=Replace_with_the_absolute_path_to_the_abinit_top_level_dir export PATH=$ABI_HOME/src/98_main/:$PATH export ABI_TESTS=$ABI_HOME/tests/ export ABI_PSPDIR=$ABI_TESTS/Psps_for_tests/ # Pseudopotentials used in examples.
Examples in this tutorial use these shell variables: copy and paste the code snippets into the terminal (remember to set ABI_HOME first!). The ‘export PATH’ line adds the directory containing the executables to your PATH so that you can invoke the code by simply typing abinit in the terminal instead of providing the absolute path.
1 Wavelets variables and parallelism¶
The parallelism with the wavelet formalism can be used for two purposes: to reduce the memory load per node, or to reduce the overall computation time.
The MPI parallelization in the wavelet mode relies on the orbital distribution scheme, in which the orbitals of the system under investigation are distributed over the assigned MPI processes. This scheme reaches its limit when the number of MPI processes is equal to the number of orbitals in the simulation. To distribute the orbitals uniformly, the number of processors must be a factor (divisor) of the number of orbitals. If this is not the case, the distribution is not optimal, but the code tries to balance the load over the processors. For example, if we have 5 orbitals and 4 processors, the orbitals will have the distribution: 2/1/1/1.
There are no specific input variables to use the parallelism in the wavelet
mode as the only parallelisation level is on orbitals. So running ABINIT with
an mpirun command is enough (this command differs according to the local MPI
implementation) such as:
mpirun -np Nproc abinit < infile > logfile
For further understanding of the wavelet mode, or for citation purposes, one may read [Genovese2008]
2 Speed-up calculation for a boron cluster¶
We propose here to determine the speed-up in the calculation of the total
energy of a cluster made of 14 boron atoms. Open the file tgswvl_01.in. It
contains first the definition of the wavelet basis-set. One may want to test
the precision of the calculation by varying the wvl_hgrid and
wvl_crmult variables. This is not the purpose of this tutorial, so we will
use the given values (0.45Bohr and 5).
usewvl 1 wvl_hgrid 0.45 wvl_crmult 5 wvl_frmult 8 nwfshist 6 ixc 11 iscf 0 icoulomb 1 optstress 0 prtwf 0 prtden 0 prteig 0 tolwfr 1e-4 ntypat 1 znucl 5 typat 14*1 natom 14 kpt 3*0. kptopt 0 istwfk *1 acell 20 20 20 xcart 6.23174441376337462E+00 1.10650949517181125E+01 8.80369514094791228E+00 6.46975493123427281E+00 1.16466203707466729E+01 7.38942614477337933E+00 6.06260066201525571E+00 9.92651732450364754E+00 7.48967117179688202E+00 7.80101706261892769E+00 1.05781643210426584E+01 7.83061522324544157E+00 6.31690262930086011E+00 8.31702761150885550E+00 7.01500573994981380E+00 7.75164346206343247E+00 7.71450195176475972E+00 8.07490331742471490E+00 7.03603095641418808E+00 9.37653827161064335E+00 6.05099299166473248E+00 9.17770300902474467E+00 7.98319528851733384E+00 8.99010444257574015E+00 7.16405045957283093E+00 1.08856491054555171E+01 6.23700501567613319E+00 7.62758557507984225E+00 5.97498074889332820E+00 7.97176704264589375E+00 7.14341418637742454E+00 1.00163217181891877E+01 9.42744042423526629E+00 8.44317230240882566E+00 9.32780744990434307E+00 8.78315808382683194E+00 6.51462241221419625E+00 6.81056643770915038E+00 7.30106028716855171E+00 8.83833328151791164E+00 6.50337987446461785E+00 8.70844982986595006E+00 Angstrom #%%<BEGIN TEST_INFO> #%% [setup] #%% executable = abinit #%% [files] #%% files_to_test = tgswvl_01_MPI2.out, tolnlines = 0, tolabs = 0.0, tolrel= 0.0 #%% psp_files= B-q3 #%% [paral_info] #%% max_nprocs = 2 #%% nprocs_to_test = 2 #%% [extra_info] #%% authors = D. Caliste #%% keywords = NC,WVL #%% description = Boron cluster, parallel calculation #%%<END TEST_INFO>
Run ABINIT with 3 processors. The overall time is printed at the end of the output file (and of the log):
Proc. 0 individual time (sec): cpu= 172.6 wall= 172.6
Read the output file to find the number of orbitals in the calculation (given by the keyword nband). With the distribution scheme of the wavelet mode, the best distribution over processors will be obtained for, 1, 3, 7 and 21 processors. Create four different directories (with the number of processors for instance) and run four times ABINIT with the same input file, varying the number of processors in {1, 3, 7, 21}. The speed-up is the ratio between the time with one processor and the time of a run with N processors.
Assuming that the directories are called {01, 03, 07, 21}, one can grep the over-all time of a run and plot it in gnuplot with:
plot "< grep 'individual time' */log | tr '/' ' '" u 1:(ttt/$11) w lp t "Boron cluster", x t "Ideal speed-up"
where ttt represents the time on one processor. The efficiency (in percent)
of the parallelization process is the ratio between the speed-up and the
number of processors. One can plot it with:
plot "< grep 'individual time' */log | tr '/' ' '" u 1:(ttt/$11/$1*100) w lp t "Boron cluster"
The first conclusion is that the efficiency is not so good when one use one orbital per processor. This is a general rule with the wavelet mode: due to the implementation, a good balance between speed and efficiency is obtained for two orbitals per processor. One can also see that the efficiency generally decreases with the number of processors.
This system is rather small and the amount of time spent in the overhead (read the input file, initialise arrays…) is impacting the performance. Let’s see how to focus on the calculation parts.
3 Time partition¶
The wavelet mode is generating a time.prc file at each run (warning: it will
erase any existing copy). This is a text file and can be read directly. There
are three sections, giving the time of the initialisation process (before
entering the SCF loop), the time of the SCF loop itself, and the time for the
post-processing. Let’s have a closer look to the SCF section (the actual
figures will vary between runs and number of processors):
CATEGORY mean TIME(sec) PERCENT CrtLocPot 1.70E-01 0.018 ApplyLocPotKin 1.83E+02 19.865 ApplyProj 1.44E+00 0.156 Precondition 3.42E+02 37.055 Rho_comput 1.10E+02 11.926 Rho_commun 5.45E+00 0.591 Un-TransSwitch 5.37E+00 0.582 Un-TransComm 5.95E+00 0.645 GramS_comput 6.84E+01 7.417 GramS_commun 8.91E-02 0.010 LagrM_comput 1.36E+02 14.784 LagrM_commun 1.44E-01 0.016 Diis 1.41E+01 1.527 PSolv_comput 2.31E+01 2.508 PSolv_commun 3.31E+00 0.358 Exchangecorr 5.01E+00 0.543 ---------------------------------------------------------------------- Total CPU time for category: WFN_OPT = 9.22E+02 Total categorized percent 98.0
With the total time of this SCF section, one can compute the speed-up and the efficiency of the wavelet mode more accurately:
N processors Speed-up Efficiency (%)
3 2.3 75.3
7 3.7 52.8
21 7.5 35.6
With the percentages of the time.prc file, one can see that, for this
example, the time is mostly spent in the precondionner and the application of
the local part of the Hamiltonian on the wavefunctions. Let’s categorise the
time information:
- The communication time is the sum of all the
_communentries and theUn-TransCommone. - The time spent doing convolutions is the sum of
Precondition, ApplyLocPotKin, Rho_comput. - The linear algebra part is the sum of
GramS_comput, LagrM_comput. - The other entries are in a miscellaneous category.
By doing the summations, one can give the percentage per category during the SCF loop:
CATEGORY mean TIME(sec) PERCENT Communication 14.9 1.62 Convolutions 635.0 68.87 Linear algebra 204.4 21.17 Other 67.7 7.34
You can analyse all the time.prc that have been generated for the different number of processors and see the evolution of the different categories.
4 Orbital parallelism and array parallelism¶
If the number of processors is not a divisor of the number of orbitals, there will be some processors with fewer orbitals than others. This is not the best distribution from an orbital point of view. But, the wavelet mode also distributes the scalar arrays like density and potentials by z-planes in real space. So some parts of the code may become more efficient when used with a bigger number of processors, like the Poisson Solver part for instance.
Run the boron example with {2, 4, 14, 15} processors and plot the speed-up. One can also look at the standard output to the load balancing of the Poisson Solver and the load balancing of orbitals (with 15 processors):
[...] Processes from 0 to 9 treat 2 orbitals Processes from 10 to 10 treat 1 orbitals Processes from 11 to 14 treat 0 orbitals [...] Load Balancing for Poisson Solver related operations: LB_density : processors 0 - 13 work at 100% processor 14 works at 40% LB_kernel : processors 0 - 12 work at 100% processor 13 works at 92% processors 14 - 14 work at 0%
One can see that, as expected, the load balancing per orbital is bad (4 processors are doing nothing), but one can see also that the load balancing of the scalar arrays is not so good since the last processor will have a reduced array. It is thus useless to run this job at 15 processors, 14 will give the same run time (since the load balancing will be better).
5 Speed-up calculation on a 65-atom alkane¶
Let’s do the same with a bigger molecule and a finer grid. Open the file
tgswvl_02.in. It contains the definition of an alkane chain of 65 atoms,
providing 64 orbitals.
usewvl 1 wvl_hgrid 0.30 wvl_crmult 7 wvl_frmult 8 nwfshist 6 nsym 1 ixc 1 iscf 2 icoulomb 1 optstress 0 prtwf 0 prtden 0 prteig 0 tolwfr 1e-4 acell 100 100 100 natom 65 ntypat 2 znucl 6 1 typat 21*1 44*2 xcart 0.00000 0.00000 0.00000 0.00000 0.88250 1.24804 0.00000 0.00000 2.49609 0.00000 0.88250 3.74413 0.00000 0.00000 4.99217 0.00000 0.88250 6.24022 0.00000 0.00000 7.48826 0.00000 0.88250 8.73630 0.00000 0.00000 9.98435 0.00000 0.88250 11.23239 0.00000 0.00000 12.48043 0.00000 0.88250 13.72848 0.00000 0.00000 14.97652 0.00000 0.88250 16.22457 0.00000 0.00000 17.47261 0.00000 0.88250 18.72065 0.00000 0.00000 19.96870 0.00000 0.88250 21.21674 0.00000 0.00000 22.46478 0.00000 0.88250 23.71283 0.00000 0.00000 24.96087 0.00000 0.61775 -0.87363 0.87363 -0.61775 0.00000 -0.87363 -0.61775 0.00000 0.87363 1.50025 1.24804 -0.87363 1.50025 1.24804 0.87363 -0.61775 2.49609 -0.87363 -0.61775 2.49609 0.87363 1.50025 3.74413 -0.87363 1.50025 3.74413 0.87363 -0.61775 4.99217 -0.87363 -0.61775 4.99217 0.87363 1.50025 6.24022 -0.87363 1.50025 6.24022 0.87363 -0.61775 7.48826 -0.87363 -0.61775 7.48826 0.87363 1.50025 8.73630 -0.87363 1.50025 8.73630 0.87363 -0.61775 9.98435 -0.87363 -0.61775 9.98435 0.87363 1.50025 11.23239 -0.87363 1.50025 11.23239 0.87363 -0.61775 12.48043 -0.87363 -0.61775 12.48043 0.87363 1.50025 13.72848 -0.87363 1.50025 13.72848 0.87363 -0.61775 14.97652 -0.87363 -0.61775 14.97652 0.87363 1.50025 16.22457 -0.87363 1.50025 16.22457 0.87363 -0.61775 17.47261 -0.87363 -0.61775 17.47261 0.87363 1.50025 18.72065 -0.87363 1.50025 18.72065 0.87363 -0.61775 19.96870 -0.87363 -0.61775 19.96870 0.87363 1.50025 21.21674 -0.87363 1.50025 21.21674 0.87363 -0.61775 22.46478 -0.87363 -0.61775 22.46478 0.87363 1.50025 23.71283 -0.87363 1.50025 23.71283 0.87363 -0.61775 24.96087 -0.87363 -0.61775 24.96087 0.00000 0.61775 25.83450 Angstrom #%%<BEGIN TEST_INFO> #%% [setup] #%% executable = abinit #%% [files] #%% files_to_test = tgswvl_02_MPI24.out, tolnlines = 0, tolabs = 0.0, tolrel= 0.0 #%% psp_files= C-q4, H-q1 #%% [paral_info] #%% max_nprocs = 24 #%% nprocs_to_test = 24 #%% [extra_info] #%% authors = D. Caliste #%% keywords = NC,WVL #%% description = chain of 65 atoms, 64 orbitals, parallel calculation #%%<END TEST_INFO>
Run this input file with {1, 2, 4, 8, 16, 24, 32, 48, 64} processors. The run with one processor should take less than one hour. If the time is short, one can reduce wvl_hgrid in the input file to 0.45.
Time measurements for a run over several processors of a C 21H44 alkane chain
As we obtained previously, the efficiency is generally lowered when the number of processors is not a divisor of the number of orbitals (namely here 24 and 48).
6 Conclusion¶
With the wavelet mode, it is possible to efficiently decrease the run time by increasing the number of processors. The efficiency is limited by the increase of amount of time spent in the communications. The efficiency increases with the quality of the calculation: the more accurate the calculations are (finer hgrid…), the more efficient the code parallelization will be.