GW parallel strategies

This modules contains very general discussions of the parallel environment of Yambo. In this tutorial we will see how to setup the variables governing the parallel execution of yambo in order to perform efficient calculations in terms of both cpu time and memory to solution. As a test case we will consider the hBN 2D material. Because of its reduced dimensionality, GW calculations turns out to be very delicate. Beside the usual convergence studies with respect to k-points and sums-over-bands, in low dimensional systems a sensible amount of vacuum is required in order to treat the system as isolated, translating into a large number of plane-waves. As for other tutorials, it is important to stress that this tutorial it is meant to illustrate the functionality of the key variables and to run in reasonable time, so it has not the purpose to reach the desired accuracy to reproduce experimental results. Moreover please also note that scaling performance illustrated below may be significantly dependent on the underlying parallel architecture. Nevertheless, general considerations are tentatively drawn in discussing the results.

Files and Tools

Database and tools can be downloaded here:

• hBN-2D-para.tar.gz

Getting familiar with yambo in parallel

Let's start by copying the tutorial files in the cluster and unzip them in the folder you will run the tutorial.

$ cp $path/hBN-2D-para.tar.gz ./
$ tar -zxvf hBN-2D-para.tar.gz
$ ls
YAMBO_TUTORIALS
$ cd YAMBO_TUTORIALS/hBN-2D-para/YAMBO

Under the YAMBO folder, together with the SAVE folder, you will see the run.sh script

$ ls
parse_qp.py parse_ytiming.py  SAVE

First, run the initialization as usual. Then you need to generate the input file for a GW run.

$ yambo -g n -p p -F yambo_gw.in 

After setting the variables in red, the new input file should look like the following:

$ cat yambo_gw.in

ppa                          # [R Xp] Plasmon Pole Approximation
gw0                          # [R GW] GoWo Quasiparticle energy levels
HF_and_locXC                 # [R XX] Hartree-Fock Self-energy and Vxc
em1d                         # [R Xd] Dynamical Inverse Dielectric Matrix
X_Threads=  0               # [OPENMP/X] Number of threads for response functions 
DIP_Threads=  0             # [OPENMP/X] Number of threads for dipoles
SE_Threads=  0              # [OPENMP/GW] Number of threads for self-energy
EXXRLvcs= 21817        RL    # [XX] Exchange RL components
VXCRLvcs= 21817        RL      # [XC] XCpotential RL components
Chimod= ""                   # [X] IP/Hartree/ALDA/LRC/BSfxc
% BndsRnXp
    1 |  200 |               # [Xp] Polarization function bands
%
NGsBlkXp= 4            Ry    # [Xp] Response block size
% LongDrXp
 1.000000 | 0.000000 | 0.000000 |        # [Xp] [cc] Electric Field
%
PPAPntXp= 27.21138     eV    # [Xp] PPA imaginary energy
% GbndRnge
    1 |  200 |               # [GW] G[W] bands range
%
GDamping=  0.10000     eV    # [GW] G[W] damping
dScStep=  0.10000      eV    # [GW] Energy step to evaluate Z factors
DysSolver= "n"               # [GW] Dyson Equation solver ("n","s","g")
%QPkrange                    # [GW] QP generalized Kpoint/Band indices
  1| 1| 3| 6|
%

Now you need to create a submission script. here below an example (run.sh) based on the SLURM scheduler. In the case of other schedulers, the header should be updated accordingly.

$ cat run.sh
 
#!/bin/bash
#SBATCH -N 1
#SBATCH -t 06:00:00 
#SBATCH -J test
#SBATCH --partition=<queue name>
#SBATCH --tasks-per-node=1

nodes=1
nthreads=1
ncpu=`echo $nodes $nthreads 1 | awk '{print $1*$3/$2}'`

module purge
module load <needed modules> 
module load <more modules> 
bindir=<path to yambo bindir> 

export OMP_NUM_THREADS=$nthreads

label=MPI${ncpu}_OMP${nthreads}
jdir=run_${label}
cdir=run_${label}.out

filein0=yambo_gw.in
filein=yambo_gw_${label}.in

cp -f $filein0 $filein
cat >> $filein << EOF
X_all_q_CPU= "1 1 $ncpu 1"  # [PARALLEL] CPUs for each role
X_all_q_ROLEs= "q k c v"    # [PARALLEL] CPUs roles (q,k,c,v)
X_all_q_nCPU_LinAlg_INV= $ncpu   # [PARALLEL] CPUs for Linear Algebra
X_Threads=  0               # [OPENMP/X] Number of threads for response functions
DIP_Threads=  0             # [OPENMP/X] Number of threads for dipoles
SE_CPU= " 1 1 $ncpu"        # [PARALLEL] CPUs for each role
SE_ROLEs= "q qp b"          # [PARALLEL] CPUs roles (q,qp,b)
SE_Threads=  0    

EOF

echo "Running on $ncpu MPI, $nthreads OpenMP threads"
srun -n $ncpu -c $nthreads $bindir/yambo -F $filein -J $jdir -C $cdir

As soon as you are ready to submit the job.

$ sbatch run.sh

Yambo calculates the GW-qp corrections running on 1 MPI process with a single thread. As you can see, monitoring the log file produced by yambo, the run takes some time, although we are using minimal parameters.

The status of the jobs can be controlled via:

$ squeue  -u $USER        # to inspect the status of jobs 
                       #(hint: make a unix alias, if you like)
$ scancel  <jobid>           # to delete jobs in the queue

Pure MPI scaling with default parallelization scheme

Meanwhile we can run the code in parallel. Let's use 16 MPI process, still with a single thread. To this end modify the job.sh script changing

#SBATCH --tasks-per-node=16

ncpu=`echo $nodes $nthreads 16 | awk '{print $1*$3/$2}'`

or set ncpu by hand:

 ncpu=XX   (e.g. 1 to 16)

This time the code should be much faster. Once the run is over try to run the simulation also on 2, 4, 8. Each time please remember to change both the number of tasks per node and the number of CPUs. At the end you can try to produce a scaling plot.

To analyze the data you can use the phyton parse_yambo.py script which is also provided.

You can use it running

$ ./parse_yambo.py run*/r-*

where $jobstring is the string you used for $jobname without the explicit number of MPI tasks.

You should obtain something like that (but with more columns)

# ncores         MPI     threads     Dipoles          Xo           X       Sgm_x       Sgm_c   WALL_TIME
      1           1           1       1.00s    7m39.00s       0.00s       3.00s    1m46.00s      09m38s
      2           2           1       2.00s    3m48.00s       0.00s       2.00s      53.00s      04m55s
      4           4           1       0.00s    1m56.00s       0.00s       1.00s      27.00s      02m33s
      8           8           1       0.00s     1m2.00s       0.00s       0.00s      14.00s      01m26s
     16          16           1       0.00s      35.00s       0.00s       0.00s       7.00s         54s

Plot the execution time vs the number of MPI tasks and check (do a log plot) how far you are from the ideal linear scaling. Below a similar plot performed on Fermi

tips:
- not all runlevels scale in in the same way
- you should never overload the available number of cores

Pure OpenMP scaling

Next step is instead to check the OpenMP scaling. Set back

#SBATCH --tasks-per-node=16

and now use

nthreads=16
ncpu=`echo $nodes $nthreads 16 | awk '{print $1*$3/$2}'`

In order to do pure OpenMP scaling the three numbers must be the same. Notice that here the last 16 entering the definition of "ncpu" defines the total number of cores (or better threads, see tips below). Since we are already using 16 threads, we cannot also split among MPI tasks, i.e. ncpu will result equal to 1. Try setting nthreads to 16, 8, 4 and 2 and again to plot the execution time vs the number of threads using the pyton script. Again you should be able to produce a plot similar to the following

tips:
- OpenMP usually shares the memory among threads, but not always
- you should never overload the available number of cores
- in principle we could overload the cores setting more threads than the available total number of cores, since a single core allows multi-thread operations

MPI vs OpenMP scaling

Which is scaling better ? MPI or OpenMP ? How is the memory distributed ?

Now you can try running simualations with hybrid strategies. Try for example setting:

#SBATCH --tasks-per-node=16

nthreads= 4
ncpu=`echo $nodes $nthreads 16 | awk '{print $1*$3/$2}'`

We can try to do scaling keeping the total number of tasks per node at 16. Parsing the data we will obtain something similar to

 # ncores         MPI     threads     Dipoles       Xo           X       Sgm_x       Sgm_c   WALL_TIME
     16           1          16       1.00s      38.00s       0.00s       2.00s      33.00s      01m25s
     16           2           8       2.00s      34.00s       0.00s       1.00s      16.00s      01m06s
     16           4           4       0.00s      34.00s       0.00s       1.00s      11.00s         56s
     16           8           2       0.00s      33.00s       0.00s       0.00s       9.00s         52s
     16          16           1       0.00s      35.00s       0.00s       0.00s       7.00s         54s

As you can see here the total CPU time decreases more and more moving the parallelization from the the OpenMP to the MPI level. Sigma_c in particular scales better. However the fastest run is not the one with 16 MPI task but the 8 2 configuration since Xo is faster using an hybrid MPI-OpenMP scheme.

However CPU time is not the only parameter you need to check. Also the total memory usage is important If you compare for example the two extreme cases (you can use)

$ grep Gb run_MPI1_OMP*/l* | grep Alloc      (use this for the case with only one MPI proc)
$ grep Gb run_MPI*_OMP*/LOG/l*_1 | grep Alloc  (use this for the case with more than one MPI proc)

For the case

 # ncores         MPI     threads
     16           1          16  
<01s> [M  0.119 Gb] Alloc WF ( 0.112)
<03s> [M  0.314 Gb] Alloc WF ( 0.306)
<46s> [M  0.074 Gb] Alloc WF ( 0.056)
<50s> [M  0.321 Gb] Alloc WF ( 0.306)

the numbers reported above refer to the total amount of memory use in the run.

For the case

 # ncores         MPI     threads
     16           16          1
<02s> P0001: [M  0.034 Gb] Alloc WF ( 0.026)
<43s> P0001: [M  0.037 Gb] Alloc WF ( 0.019)
<45s> P0001: [M  0.091 Gb] Alloc WF ( 0.076)

the numbers reported above refer to the total amount of memory per MPI task. Multiplying by 16 you obtain an estimate of the total memory: 0.091*16=1.456 (0.076*16=1.216) These last two numbers have to be compared with 0.321 (0.306) As you can see yambo is distributing memory, since the single MPI task uses less memory than the total one needed (you can even compare with the serial case). However it is not as efficient as OpenMP in doing so.

Using an hybrid scheme you may also consider running yambo on mode than one node. To run on two nodes for example you need to set

#SBATCH -N 2

nodes=2

accordingly you can now set

nthreads= 4
ncpu=`echo $nodes $nthreads 16 | awk '{print $1*$3/$2}'`

This time you will use 32 cores with (16 per node) 4 OpenMP threads and 2*16/4=8 MPI tasks.

tips:
- in real life calculations running on n_cores > 100 it is a good idea to adopt an hybrid approach
- with OpenMP you cannot exit the single node, with MPI you can

Advanced: Comparing different parallelization schemes (optional)

Up to now we used the default parallelization scheme. Yambo also allows you to tune the parameters which controls the parallelization scheme. To this end you can open again the job.sh script and modify the section where the yambo input variables are set

X_all_q_CPU= "1 1 $ncpu 1"   # [PARALLEL] CPUs for each role
X_all_q_ROLEs= "q k c v"     # [PARALLEL] CPUs roles (q,k,c,v)
#X_all_q_nCPU_LinAlg_INV= $ncpu   # [PARALLEL] CPUs for Linear Algebra
X_Threads=  0               # [OPENMP/X] Number of threads for response functions
DIP_Threads=  0             # [OPENMP/X] Number of threads for dipoles
SE_CPU= "1 1 $ncpu"             # [PARALLEL] CPUs for each role
SE_ROLEs= "q qp b"               # [PARALLEL] CPUs roles (q,qp,b)
SE_Threads=  0    

In particular "X_all_q_CPU" sets how the MPI Tasks are distributed in the calculation of the response function. The possibilities are shown in the "X_all_q_ROLEs". The same holds for "SE_CPU" and "SE_ROLEs" which control how MPI Tasks are distributed in the calculation of the response function.

Please try different parallelization schemes and check the performances of Yambo. In doing so you should also change the jobname in the run.sh script

label=MPI${ncpu}_OMP${nthreads}_scheme1

Using the python script, you can then chenck how speed, memory and load balance between the CPUs are affected. For more details see also the Parallel module

tips:
- the product of the numbers entering each variable (i.e. X_all_q_CPU and SE_CPU) times the number of threads should always match the total number of cores (unless you want to overload the cores taking advantage of multi-threads)
- using the X_Threads and SE_Threads variables you can think about setting different Hybrid schemes in between the screening and the self-energy runlevel. - memory better scales if you parallelize on bands (c v b)
- parallelization on k-points performs similarly to parallelization on bands, but memory requires more memory
- parallelization on q-points requires much less communication in between the MPI tasks. It maybe useful if you run on more than one node and the inter-node connection is slow

Prev: Tutorials Home

Now: Tutorials Home --> GW Parallel

Next: GW Convergence