Cluster Performance
The performance of the computer clusters can be an important criterium up-on deciding which one to use for large projects. Additional informations on parallelisation parameters for the from us commonly used DFT-code VASP are essential to speed up simulations and make the most of the obtained computational time.
First the performance on the various computer clusters (Vera, Beskow, Tetralith) are compared. For Tetralith the results for VASP5.4 and VASP6.2 are shown. Overall VASP6.2 is not faster than VASP5.4. The systems investigated are: A small zeolite pore with adsorbate inside, a In2O3 (111) pristine surface and a CO molecule adsorbed on top of a Pt atom for a 3x3 (111) surface. For all systems one single-point calculation was performed on an a prior relaxed structure. The different structures and VASP-input files can be downloaded in this zip-file.
As can be seen from the picture above, over the investigated range of Number of Nodes, the Nodes on Beskow are generally the slowest. For all systems the Nodes have 32 physical cores. Strangely the single Node run on Vera is much worse than paralysing over various nodes. For N(Nodes) >= 2 Tetralith and Vera show the same performance.
For VASP several parallelisation options are available. The main ones are NCORE (connected to NPAR) which is the number of cores which work on one Band. The second one is KPAR, the number of cores per kpoint. NCORE is also mainly effecting the work balance distribution for memory demanding jobs. The above picture shows the influence of NCORE on the different simulations for Beskow and Tetralith respectively. KPAR=1 in all cases. It can be seen that the performance for the various NCORE values is strongly system dependent and should usually be tested a-prior long simulations or if one has a set of similar calculations at one of them.
To run similar tests by one-self I attach here also a collection of useful set-up, submit and plot scripts:
Set-up a set of simulations for various Number of Nodes
Obtain results in a dictionary or plot directly from various OUTCAR files
Although all used clusters use SLURM as queuing system, their submission scripts are quite different, especially the run command: