GFN-FF scalability

[TinacciL]

We try to run a short MDs with GFN-FF on xTB (precompiled 6.4.0 version) on a 1214 atoms (protein + water) on the following specific node:

CPU - 2x Intel® Xeon® Processor E5-2680 v3, 12 core 2.5Ghz
RAM - 128GB/2133 (8 x 16 Gb)

using these default values to run the test:

export OMP_STACKSIZE=4G
export OMP_NUM_THREADS=n,1
export OMP_MAX_ACTIVE_LEVELS=1
export MKL_NUM_THREADS=n

where n is the number of core used in the test.

Here there is the image of the result test.

The scalability test shows that our system reach the plateau at 6 CPUs. Same results (with different offset in the graph) with SP on ~5k atoms and with different machines.

Is it expected such behavior? We was expecting a pseudo linear scalability at least until several tens of cores also considering the large size of the system. Directly compiling the xTB source code wuold change the result?

[awvwgk]

We don't have a xTB-FF method implemented here, I guess you are referring to GFN-FF?

[TinacciL]

sorry for the mistake, yes I was referring to GFN-FF

[sespic]

This behavior is expected. The energy terms in GFN-FF are OMP parallel and work best for a small number of threads. We found that more than 8 threads don't speed up the calculations anymore. This behavior also occurs with the xTB methods. The numerical Hessian matrix calculation on the other hand parallelizes very well. Here, you will find (almost) linear speedups.