NCORE: Difference between revisions
No edit summary |
No edit summary |
||
| Line 6: | Line 6: | ||
VASP currently offers parallelization and data distribution over bands and/or over plane wave coefficients, and as of VASP.5.3.2, parallelization over '''k'''-points (no data distribution, see {{TAG|KPAR}}). | VASP currently offers parallelization and data distribution over bands and/or over plane wave coefficients, and as of VASP.5.3.2, parallelization over '''k'''-points (no data distribution, see {{TAG|KPAR}}). | ||
To obtain high efficiency on massively parallel systems or modern multi-core machines, it is strongly recommended to use all at the same time. Most algorithms work with any data distribution (except for the single band conjugated gradient, which is considered to be obsolete). | To obtain high efficiency on massively parallel systems or modern multi-core machines, it is strongly recommended to use all at the same time. Most algorithms work with any data distribution (except for the single band conjugated gradient, which is considered to be obsolete). | ||
{{TAG|NCORE}} is available from VASP.5.2.13 on, and is more handy than the previous parameter {{TAG|NPAR}}. The user should either specify {{TAG|NCORE}} or {{TAG|NPAR}}, where {{TAG|NPAR}} takes a higher preference. The relation between both parameters is | |||
NCORE =number-of-cores / NPAR | |||
{{TAG|NCORE}} determines how many cores share the work on an individual orbital. The current default is {{TAG|NCORE}}=1, meaning that one orbital is treated by one core. {{TAG|NPAR}} is then set to the total number of cores. If {{TAG|NCORE}} equals the total number of cores, {{TAG|NPAR}} is set to 1. This implies data distribution over plane wave coefficients only: all cores will work together on every individual band, i.e., the plane wave coefficients of each band are distributed over all cores. This is usually very slow and should be avoided. | {{TAG|NCORE}} determines how many cores share the work on an individual orbital. The current default is {{TAG|NCORE}}=1, meaning that one orbital is treated by one core. {{TAG|NPAR}} is then set to the total number of cores. If {{TAG|NCORE}} equals the total number of cores, {{TAG|NPAR}} is set to 1. This implies data distribution over plane wave coefficients only: all cores will work together on every individual band, i.e., the plane wave coefficients of each band are distributed over all cores. This is usually very slow and should be avoided. | ||
{{TAG|NCORE}}=1 is the optimal setting for small unit cells, and platforms with a small communication bandwidth. It is also a good choice for up to 8 cores. 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. | {{TAG|NCORE}}=1 is the optimal setting for small unit cells, and platforms with a small communication bandwidth. It is also a good choice for up to 8 cores. 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 | On massively parallel systems and modern multi-core machines we strongly urge to set | ||
| Line 24: | Line 21: | ||
For large unit cells, we found that this can improve the performance by up to a factor four compared to the default, and it also significantly improves the stability of the code due to reduced memory requirements. Ideally, {{TAG|NCORE}} should be a factor of the number-of-cores-per-socket (or number-of-cores-per-node), since this reduces communication between the sockets or nodes. The best value {{TAG|NCORE}} depends somewhat on the number of atoms | For large unit cells, we found that this can improve the performance by up to a factor four compared to the default, and it also significantly improves the stability of the code due to reduced memory requirements. Ideally, {{TAG|NCORE}} should be a factor of the number-of-cores-per-socket (or number-of-cores-per-node), since this reduces communication between the sockets or nodes. The best value {{TAG|NCORE}} depends somewhat on the number of atoms | ||
in the unit cell. Values around 4 | in the unit cell. Values around 4 are usually ideal for 100 atoms in the unit cell. For very large unit cells (more than 400) values around 12-16 are often optimal. If you run extensive simulations for similar systems, make your own tests. | ||
Finally, the optimal settings for {{TAG|NCORE}} and {{TAG|LPLANE}} depend strongly on the type of machine you are using. | |||
Some recommended setups: | Some recommended setups: | ||
| Line 34: | Line 31: | ||
<pre> | <pre> | ||
LPLANE = .TRUE. | LPLANE = .TRUE. | ||
NCORE = number of cores per node | NCORE = 2 - number-of-cores-per-node | ||
LSCALU = .FALSE. | LSCALU = .FALSE. | ||
NSIM = 4 | NSIM = 4 | ||
</pre> | </pre> | ||
:If very many nodes are used, it might be necessary to set LPLANE = .FALSE., but usually this offers very little | :If very many nodes are used, it might be necessary to set LPLANE = .FALSE., but usually this offers very little performance boost. | ||
*LINUX cluster linked by 1 Gbit Ethernet, and LINUX clusters with single cores: | *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 | :On a LINUX cluster linked by a relatively slow network, LPLANE must be set to .TRUE., and the NCORE should be equal to 1: | ||
<pre> | <pre> | ||
LPLANE = .TRUE. | LPLANE = .TRUE. | ||
| Line 48: | Line 45: | ||
NSIM = 4 | NSIM = 4 | ||
</pre> | </pre> | ||
: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 | :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 useful 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): | *Massively parallel machines (Cray, Blue Gene): | ||
:On many massively parallel machines one is forced to use a | :On many massively parallel machines one is sometimes forced to use a large number of cores. 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 using any reasonable VASP setting). Therefore, we recommend to set {{TAG|NCORE}} on these machines to √''# of cores'' (explicit timing can be helpful to find the optimum value). The use of {{TAG|LPLANE}}=.TRUE. is only recommended if the number of nodes is significantly smaller than {{TAG|NGX}}, {{TAG|NGY}} and {{TAG|NGZ}}. | ||
:In summary, the following setting is recommended | :In summary, the following setting is recommended | ||
Revision as of 12:17, 14 October 2020
NCORE = [integer]
Default: NCORE = 1
Description: NCORE determines the number of compute cores that work on an individual orbital (available as of VASP.5.2.13).
VASP currently offers parallelization and data distribution over bands and/or over plane wave coefficients, and as of VASP.5.3.2, parallelization over k-points (no data distribution, see KPAR). To obtain high efficiency on massively parallel systems or modern multi-core machines, it is strongly recommended to use all 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
NCORE =number-of-cores / NPAR
NCORE determines how many cores share the work on an individual orbital. The current default is NCORE=1, meaning 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 data distribution over plane wave coefficients only: all cores will work together on every individual band, i.e., the plane wave coefficients of each band are distributed over all cores. This is usually very slow and should be avoided.
NCORE=1 is the optimal setting for small unit cells, and platforms with a small communication bandwidth. It is also a good choice for up to 8 cores. 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
NCORE = 2 - number-of-cores-per-socket (or number-of-cores-per-node)
For large unit cells, we found that this can improve the performance by up to a factor four compared to the default, and it also significantly improves the stability of the code due to reduced memory requirements. Ideally, NCORE should be a factor of the number-of-cores-per-socket (or number-of-cores-per-node), since this reduces communication between the sockets or nodes. The best value NCORE depends somewhat on the number of atoms
in the unit cell. Values around 4 are usually ideal for 100 atoms in the unit cell. For very large unit cells (more than 400) values around 12-16 are often optimal. If you run extensive simulations for similar systems, make your own tests.
Finally, the optimal settings for NCORE and LPLANE depend strongly on the type of machine you are using.
Some 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 = 2 - number-of-cores-per-node LSCALU = .FALSE. NSIM = 4
- If very many nodes are used, it might be necessary to set LPLANE = .FALSE., but usually this offers very little performance boost.
- 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 NCORE should be equal to 1:
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 useful 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 sometimes forced to use a large number of cores. 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 using any reasonable VASP setting). Therefore, we recommend to set NCORE on these machines to √# of cores (explicit timing can be helpful to find the optimum value). The use of LPLANE=.TRUE. is only recommended 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 cores) NSIM = 1
Related Tags and Sections
NPAR, LPLANE, LSCALU, NSIM, KPAR, LSCALAPACK, LSCAAWARE