Next: 3.4 Tricks and problems Up: 3 Parallelism Previous: 3.2 Running on parallel   Contents

Subsections

## 3.3 Parallelization levels

In QUANTUM ESPRESSO several MPI parallelization levels are implemented, in which both calculations and data structures are distributed across processors. Processors are organized in a hierarchy of groups, which are identified by different MPI communicators level. The groups hierarchy is as follow:

• world: is the group of all processors (MPI_COMM_WORLD).
• images: Processors can then be divided into different "images", each corresponding to a different self-consistent or linear-response calculation, loosely coupled to others.
• pools: each image can be subpartitioned into "pools", each taking care of a group of k-points.
• bands: each pool is subpartitioned into "band groups", each taking care of a group of Kohn-Sham orbitals (also called bands, or wavefunctions). Especially useful for calculations with hybrid functionals.
• PW: orbitals in the PW basis set, as well as charges and density in either reciprocal or real space, are distributed across processors. This is usually referred to as "PW parallelization". All linear-algebra operations on array of PW / real-space grids are automatically and effectively parallelized. 3D FFT is used to transform electronic wave functions from reciprocal to real space and vice versa. The 3D FFT is parallelized by distributing planes of the 3D grid in real space to processors (in reciprocal space, it is columns of G-vectors that are distributed to processors).
• tasks: In order to allow good parallelization of the 3D FFT when the number of processors exceeds the number of FFT planes, FFTs on Kohn-Sham states are redistributed to ``task'' groups so that each group can process several wavefunctions at the same time. Alternatively, when this is not possible, a further subdivision of FFT planes is performed.
• linear-algebra group: A further level of parallelization, independent on PW or k-point parallelization, is the parallelization of subspace diagonalization / iterative orthonormalization. Both operations required the diagonalization of arrays whose dimension is the number of Kohn-Sham states (or a small multiple of it). All such arrays are distributed block-like across the ``linear-algebra group'', a subgroup of the pool of processors, organized in a square 2D grid. As a consequence the number of processors in the linear-algebra group is given by n2, where n is an integer; n2 must be smaller than the number of processors in the PW group. The diagonalization is then performed in parallel using standard linear algebra operations. (This diagonalization is used by, but should not be confused with, the iterative Davidson algorithm). The preferred option is to use ELPA and ScaLAPACK; alternative built-in algorithms are anyway available.
Note however that not all parallelization levels are implemented in all codes.

Images and pools are loosely coupled: inter-processors communication between different images and pools is modest. Processors within each pool are instead tightly coupled and communications are significant. This means that fast communication hardware is needed if your pool extends over more than a few processors on different nodes.

#### 3.3.0.2 Choosing parameters

: To control the number of processors in each group, command line switches: -nimage, -npools, -nband, -ntg, -ndiag or -northo (shorthands, respectively: -ni, -nk, -nb, -nt, -nd) are used. As an example consider the following command line:
```mpirun -np 4096 ./neb.x -ni 8 -nk 2 -nt 4 -nd 144 -i my.input
```
This executes a NEB calculation on 4096 processors, 8 images (points in the configuration space in this case) at the same time, each of which is distributed across 512 processors. k-points are distributed across 2 pools of 256 processors each, 3D FFT is performed using 4 task groups (64 processors each, so the 3D real-space grid is cut into 64 slices), and the diagonalization of the subspace Hamiltonian is distributed to a square grid of 144 processors (12x12).

Default values are: -ni 1 -nk 1 -nt 1 ; nd is set to 1 if ScaLAPACK is not compiled, it is set to the square integer smaller than or equal to half the number of processors of each pool.

#### 3.3.0.3 Massively parallel calculations

For very large jobs (i.e. O(1000) atoms or more) or for very long jobs, to be run on massively parallel machines (e.g. IBM BlueGene) it is crucial to use in an effective way all available parallelization levels: on linear algebra (requires compilation with ELPA and/or ScaLAPACK), on "task groups" (requires run-time option "-nt N"), and mixed MPI-OpenMP (requires OpenMP compilation: configure-enable-openmp). Without a judicious choice of parameters, large jobs will find a stumbling block in either memory or CPU requirements. Note that I/O may also become a limiting factor.

### 3.3.1 Understanding parallel I/O

In parallel execution, each processor has its own slice of data (Kohn-Sham orbitals, charge density, etc), that have to be written to temporary files during the calculation, or to data files at the end of the calculation. This can be done in two different ways:
• ``collected'': all slices are collected by the code to a single processor that writes them to disk, in a single file, using a format that doesn't depend upon the number of processors or their distribution. This is the default since v.6.2 for final data.
• ``portable'': as above, but data can be copied to and read from a different machines (this is not guaranteed with Fortran binary files). Requires compilation with `-D__HDF5` preprocessing option and HDF
There is a third format, no longer used for final data but used for scratch and restart files:
• ``distributed'': each processor writes its own slice to disk in its internal format to a different file. The ``distributed'' format is fast and simple, but the data so produced is readable only by a job running on the same number of processors, with the same type of parallelization, as the job who wrote the data, and if all files are on a file system that is visible to all processors (i.e., you cannot use local scratch directories: there is presently no way to ensure that the distribution of processes across processors will follow the same pattern for different jobs).

The directory for data is specified in input variables outdir and prefix (the former can be specified as well in environment variable ESPRESSO_TMPDIR): outdir/prefix.save. A copy of pseudopotential files is also written there. If some processor cannot access the data directory, the pseudopotential files are read instead from the pseudopotential directory specified in input data. Unpredictable results may follow if those files are not the same as those in the data directory!

IMPORTANT: Avoid I/O to network-mounted disks (via NFS) as much as you can! Ideally the scratch directory outdir should be a modern Parallel File System. If you do not have any, you can use local scratch disks (i.e. each node is physically connected to a disk and writes to it) but you may run into trouble anyway if you need to access your files that are scattered in an unpredictable way across disks residing on different nodes.

You can use input variable disk_io to vary the amount of I/O done by pw.x. Since v.5.1, the dafault value is disk_io='low', so the code will store wavefunctions into RAM and not on disk during the calculation. Specify disk_io='medium' only if you have too many k-points and you run into trouble with memory; choose disk_io='none' if you do not need to keep final data files.

Next: 3.4 Tricks and problems Up: 3 Parallelism Previous: 3.2 Running on parallel   Contents
Paolo Giannozzi 2019-04-18