Parallelisation: NPAR-tag, NCORE-tag, and LPLANE-tag

VASP currently offers parallelisation (and data distribution) over bands and parallelization (and data distribution) over plane wave coefficients (see also Section 4). To obtain high efficiency on massively parallel systems or modern multi-core machines, it is strongly recommended to use both at the same time. Most algorithms work with any data distribution (except for the single band conjugated gradient, which is considered to be obsolete).

NCORE is available from VASP.5.2.13 on, and is more handy than the previous parameter NPAR. The user should either specify NCORE or NPAR, where NPAR takes a higher preference. The relation between both parameters is

$${\tt NCORE} = \mbox{total number cores}/ {\tt NPAR}.$$

NCORE determines how many cores work on one orbital. The value is also printed at the beginning of the OUTCAR file. The current default is NCORE=1, implying that one orbital is treated by one core. NPAR is then set to the total number of cores. If NCORE equals the total number of cores, NPAR is set to 1. This implies distribution over plane wave coefficients only: all cores will work on every individual band, by distributing the plane wave coefficients over all cores. This is usually very slow and should be avoided.

NCORE=1 is the optimal setting for platforms with a small communication bandwidth and is a good choice for up to cores, as well as, machines with a single core per node and a Gigabit network. However, this mode substantially increases the memory requirements, because the non-local projector functions must be stored entirely on each core. In addition, substantial all-to-all communications are required to orthogonalize the bands. On massively parallel systems and modern multi-core machines we strongly urge to set

$${\tt NPAR=} \approx \sqrt{ \mbox{number of cores}}$$

or

$${\tt NPAR=} \mbox{number of cores per compute node}$$

In selected cases, we found that this improves the performance by a factor of up to four compared to the default, and it also significantly improves the stability of the code due to reduced memory requirements.

The second switch influences the data distribution is LPLANE. If LPLANE is set to .TRUE. in the INCAR file, the data distribution in real space is done plane wise. Any combination of NPAR and LPLANE can be used. Generally, LPLANE=.TRUE. reduces the communication band width during the FFT's, but at the same time it unfortunately worsens the load balancing on massively parallel machines. LPLANE=.TRUE. should only be used if NGZ is at least 3*(number of nodes)/NPAR, and optimal load balancing is achieved if NGZ=n*NPAR, where n is an arbitrary integer. If LPLANE=.TRUE. and if the real space projector functions (LREAL=.TRUE. or ON or AUTO) are used, it might be necessary to check the lines following

 real space projector functions
  total allocation   :
  max/ min on nodes  :

The max/ min values should not differ too much, otherwise the load balancing might worsen as well.

The optimum settings for NPAR and LPLANE depend very much on the type of machine you are using. Results for some selected machines can be found in Sec. 3.10. Recommended setups:

• LINUX cluster linked by Infiniband, modern multicore machines:
  

  On a LINUX cluster with multicore machines linked by a fast network we recommend to set

   LPLANE = .TRUE.
   NCORE   = number of cores per nodes (e.g. 4 or 8)
   LSCALU = .FALSE.
   NSIM   = 4
  

  If very many nodes are used, it might be necessary to set LPLANE = .FALSE., but usually this offers very little advantage. For long (e.g. molecular dynamics runs), we recommend to optimize NPAR by trying short runs for different settings.

• LINUX cluster linked by 1 Gbit Ethernet, and LINUX clusters with single cores:
  On a LINUX cluster linked by a relatively slow network, LPLANE must be set to .TRUE., and the NPAR flag should be equal to the number of cores:

   LPLANE = .TRUE.
   NCORE   = 1
   LSCALU = .FALSE.
   NSIM   = 4
  

  Mind that you need at least a 100 Mbit full duplex network, with a fast switch offering at least 2 Gbit switch capacity to find usefull speedups. Multi-core machines should be always linked by an Infiniband, since Gbit is too slow for multi-core machines.

• Massively parallel machines (Cray, Blue Gene):

  On many massively parallel machines one is forced to use a huge number of nodes. In this case load balancing problems and problems with the communication bandwidth are likely to be experienced. In addition the local memory is fairly small on some massively parallel machines; too small keep the real space projectors in the cache with any setting. Therefore, we recommend to set NPAR on these machines to $\sqrt{\mbox{ number of nodes}}$ (explicit timing can be helpful to find the optimum value). The use of LPLANE=.TRUE. is only recommend if the number of nodes is significantly smaller than NGX, NGY and NGZ.

  In summary, the following setting is recommended

   LPLANE = .FALSE.
   NPAR   = sqrt(number of nodes)
   NSIM   = 1