Which machine should I use? Well, it depends on what you are running. Here we collect results from specific structures to determine the best way to use our computational resources. The structures are:
- Calcite + water, 337 atoms, self-consistent calculation (qe, pw)
- Gold surface, 112 atoms, self-consistent calculation (qe, pw)
- CO2 liquid-phase, 81 atoms, ab-initio molecular dynamics (VASP, IBRION=0, MDALGO=2)
How many seconds/minutes/hours are needed to solve these calculations? Please include basic analysis of the data and a suggestion in case someone is running something similar.
Benchmark input files and results by Prof. James Moraes de Almeida (Jul 2020).
| Machine | Cores/Boards | PWSCF (CPU, s) | Cost Efficiency |
|---|---|---|---|
| Tesla V100 SXM2 32GB GPU | 008 | 165 | 65.91 % |
| --- | 004 | 255 | 76.35 % |
| AMD Epyc 7532 (cpu) | 128 | 348 | 100.00 % |
| --- | 064 | 623 | 111.71 % |
| nanopetro-intel (cpu) | 028 | 1799 | 38.69 % |
| nanopetro-amd (cpu) | 064 | 3137 | 19.86 % |
Observations
- It was considered a price ratio (GPU/CPU) = 3.2 to estimate the cost efficiency;
- GPU/CPU price ratio has to go down to lower than 2.5 to favor GPUs over CPUs
- CPUs still present an advantage over GPUs considering price;
- For DFT, the best strategy is still scaling down CPU cores until max. performance;
- nanopetro nodes included as a reference.
This SCF calculation test was forked from here. Calculations were performed in May 2020 by @camilofs.
| Machine | Cores | PWSCF (CPU, h) | Efficiency |
|---|---|---|---|
| nanopetro-intel (cpu) | 28 | 0.14244 | 100.00 % |
| --- | 14 | 0.27366 | 104.11 % |
| --- | 12 | 0.29512 | 112.62 % |
| --- | 08 | 0.37400 | 133.00 % |
| nanopetro-amd (cpu) | 64 | 0.36804 | 100.00 % |
| --- | 32 | 0.65282 | 112.75 % |
| --- | 16 | 1.26667 | 116.75 % |
| --- | 08 | 2.01667 | 146.00 % |
Observations
- CPU time only; running quantum-espresso 6.4.0 & openmpi-3.1.4, Jul 2020. Efficiency estimated based on a full node allocation;
- nanopetro-amd --> AMD Opteron 6376;
- nanopetro-intel --> Intel Xeon Gold 5120;
- Best options are highlighted. In nanopetro-intel, use a combination of 12+8+8 cores (three jobs).
Benchmark from the leda/dia project (RCGI1, 2021). It comprises 10.000 ab-initio molecular dynamics steps using VASP. The system has 81 atoms (27 CO_2 molecules). Integration step (POTIM) is 1 fs. Please, refer to the folder to acess all the input files. Calculations were performed in Feb 2021 by @camilofs.
| Machine | Cores | Time for 10k steps (min) | Efficiency |
|---|---|---|---|
| nanopetro-intel (cpu) | 28 | 19040 | 100.00 % |
| --- | 14 | 24110 | 157.94 % |
| nanopetro-amd (cpu) | 64 | 7488 | 100.00 % |
| --- | 32 | 9060 | 165.29 % |
Observations
- CPU time only; running VASP 5.4.4 with a gcc 9.2.0 compilation, 2020. Efficiency estimated based on a full node allocation;
- nanopetro-amd --> AMD Opteron 6376;
- nanopetro-intel --> Intel Xeon Gold 5120;
- The best option is highlighted; nanopetro-amd cores perform way better in AIMD. They are the only ones that run each MD step in less than 1 min.