Category:Parallelization: Difference between revisions

VASP makes use of parallel machines splitting the calculation into many tasks, that communicate with each other using MPI. Since a single core cannot perform enough operations, for many complex problems, this parallelization is necessary to finish the calculation in a reasonable time.

Theory

Basic parallelization

By default, VASP distributes the number of bands (NBANDS) over the available MPI ranks. But it is often beneficial to add parallelization of the FFTs (NCORE), parallelization over k points (KPAR), and parallelization over separate calculations (IMAGES). All these tags default to 1 and divide the number of MPI ranks among the parallelization options. Additionally, there are some parallelization options for specific algorithms in VASP, e.g., NOMEGAPAR. In summary, VASP parallelizes with

${\displaystyle \text{total ranks} = \text{ranks parallelizing bands} \times \text{NCORE} \times \text{KPAR} \times \text{IMAGES} \times \text{other algorithm-dependent tags}. }$

In addition to the parallelization using MPI, VASP can make use of OpenMP threading and/or OpenACC (for the GPU port). Note that running on multiple OpenMP threads and/or GPUs switches off the NCORE parallelization.

MPI setup

The MPI setup determines the placement of the ranks onto the nodes. VASP assumes the ranks first fill up a node before the next node is occupied. As an example when running with 8 ranks on two nodes, VASP expects rank 1–4 on node 1 and rank 5–8 on node 2. If the ranks are placed differently, communication between the nodes occurs for every parallel FFT. Because FFTs are essential to VASP's speed, this deteriorates the performance of the calculation. A manifestation is an increase in computing time when the number of nodes is increased from 1 to 2. If NCORE is not used this issue is less severe but will still reduce the performance.

To address this issue, please check the setup of the MPI library and the submitted job script. It is usually possible to overwrite the placement by setting environment variables or command-line arguments. When in doubt, contact the HPC administration of your machine to investigate the behavior.

Terminology in high-performance computing (HPC)

CPU

The central processing unit of a computer. A CPU may consist of multiple cores. One or more CPUs can be combined with accelerators like GPUs to form a node. Desktop computers typically contain a single CPU.

GPU

The graphical processing unit. A GPU is very efficient at matrix and vector operations and may accelerate a program by transferring particularly suitable tasks from the CPU to the GPU.

Core

When a CPU has the option to execute multiple tasks in parallel, we refer to this as a multi-core CPU. Because these computational cores are physically close, they typically exhibit a fast communication between them.

Node

The node constitutes a physical entity consisting of one or more CPUs potentially accelerated by GPUs. The communication between nodes is much slower compared to the communication within a single node.

Socket

Processes communicate via sockets. Each socket corresponds to an endpoint in this communication.

Process

A process is a program executing on one or more cores. Multiple processes may distribute work via communication. Each process may spawn multiple threads to execute their task.

OpenMP thread

These threads live on a single node. Processes can instantiate these threads for loops or other parallel tasks.

Message Passing Interface (MPI)

A communication protocol to facilitate parallel execution of multiple processes. The processes send messages among them to synchronize parallel tasks when necessary.

Rank

Each rank corresponds to one process participating in the MPI communication. These ranks determine which particular task a process works on and identify senders and receivers of messages.

Memory

Processes store the data they work on in the dynamic random access memory (DRAM) or random access memory (RAM). From there it propagates to the execution cores via the cache.

Cache

The cache is physically much closer to the CPU than the memory. Therefore, data is moved from the memory to the cache before processing.

How to

Optimizing the parallelization

The performance of a specific parallelization depends on the system, i.e., the number of ions, the elements, the size of the cell, etc. Different algorithms (density-functional theory, many-body perturbation theory, or molecular dynamics) require a separate optimization of the parallel setup. To obtain publishable results, many projects require performing many similar calculations, i.e., calculations with similar input and using the same algorithms. Therefore, we recommend optimizing the parallelization to make the most of the available compute time.

For more detailed advice, check the following:

• How to optimize the parallelization in a nutshell

OpenMP/OpenACC

Both OpenMP and OpenACC parallelize the FFTs and therefore disregard any conflicting specification of NCORE. When combining these methods OpenACC takes precedence but any code not ported to OpenACC benefits from the additional OpenMP threads. This approach is relevant because the recommended NVIDIA Collective Communications Library requires a single MPI rank per GPU. Learn more about the OpenMP and OpenACC parallelization in these sections

• How to parallelize with multiple OpenMP threads per MPI rank
• How to run on GPUs

Additional parallelization options

KPAR

For Laplace transformed MP2 this tag has a different meaning.

NCORE_IN_IMAGE1

Defines how many ranks work on the first image in the thermodynamic coupling constant integration (VCAIMAGES).

NOMEGAPAR

Parallelize over imaginary frequency points in ${\displaystyle GW}$ and RPA calculations.

NTAUPAR

Parallelize over imaginary time points in ${\displaystyle GW}$ and RPA calculations.