VASP多节点并行参数设置

VASP currently offers parallelization and data distribution over bands and/or over plane wave coefficients (see NCORE and NPAR), and 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.

以下假设计算总核心数为N，单节点计算核心数为n

• NPAR与NCORE
  • 两者相关：NCORE*NPAR=N/KPAR,  取其一设置即可，vaspwiki上建议优先设置NPAR，但NCORE更为方便 (NCORE is available from VASP.5.2.13 on, and is more handy than the previous parameter NPAR)
  • NPAR决定能带并行策略（NPAR determines the number of bands that are treated in parallel）
  • NCORE决定轨道并行策略（NCORE determines the number of compute cores that work on an individual orbital）
  • 大规模并行，vaspwiki上建议NPAR≈Sqrt(N) or NCORE = n. 若N开方非整数，则取开方结果附近的整数，这里需要注意：NPAR需要适当选取，使得NCORE=N/NPAR是n的factor，以减小节点间通信的overhead，否则VASP会采用默认设置，即NPAR=N or NCORE=1，默认值适用于小核数（8核以内）和小的通信带宽
  • NCORE的最佳值取决于一个unit cell中的具体原子数，对于100个原子左右的unit cell, NCORE∼4；对于大的unit cell (more than 400 atoms), NCORE∼12-16
• KPAR
  • The set of k-points is distributed over KPAR groups 
  • KPAR决定K点并行策略（KPAR determines the number of k-points that are to be treated in parallel）
  • choose KPAR such that it is an integer divisor of N
  • the data is not distributed additionally over k-points

---

总结：

以上源于vaspwiki的一般性总结，具体请以实际测试为准！

特别在超算上，一定要花时间测试！否则可能吃（算）力不讨好！

计算测试从单节点开始，不要一上来就做多节点测试，确保每一次增加节点后要比之前快！

---

个人实践测试经验

VASP针对k点和能带做了并行计算处理，可以从vasp刚开始输出的结果看出

running on N total cores
distrk: each k-point on N/KPAR cores, KPAR groups
distr: one band NCORE cores, NPAR groups

其中NCORE需要为n的factor. 相对于NPAR，设置NCORE更为方便，因为直接设NPAR还要验证NCORE=N/NPAR是否是n的factor. vaspwiki建议对于100个原子左右的unit cell, NCORE∼4；对于大的unit cell (more than 400 atoms), NCORE∼12-16. 因此，可根据自己的体系和资源合理选取.